River Roberts: Zeit war das in.
Ty Pattison: Hello, hello. Good morning, River Andrew.
River Roberts: How how gr how grows it? Yes, we have a guest today. I would love to introduce you to her to Andrew Davis, who I actually I met at an event last year called at something called AI Builders in Lisbon, Portugal. And we immediately hit it off. And I was really fascinated by Andrew's background, because he actually spent time as as a as a monk, as a as a monastic, and then he went into the engineering side of things. So actually being able balance those two sides, I felt like how many people are doing have come from those two worlds and somehow bring it together. And so he's been working on a project called Living Code and a few different kind of projects on that side of things. So Andrew, welcome. wondered if maybe you could introduce a little bit more about yourself and don't worry about too much because we're gonna go into like what you're actively building. in the next section. But any anything I missed out at a high level?
Andrew Davis: not particularly delighted to be here. And I as you said, River, y as soon as we met, I think your your blend of r real life facilitation, group facilitation, of you know, drawing out the insights from the participants in that AI builders event, plus mixing in the you know, set of questions to ask to your AI companions, your LLMs, and I really found that approach to be particularly rich and at the time and in general, right? Because we are entering such a weird world of trying to, you know, do deeper it's important that each of us do deeper thinking ourselves. Also extremely important that we do collective thinking, that we we continue to enrich our social interactions and social thinking. And that we have these alien artificial intelligence kind of entities that we're integrating into our lives. And it's really important that that not be strictly solitary undertaking, right? That there's a lot of risk of you know that dri channeling
River Roberts: Churely, chilling.
Andrew Davis: people into loneliness traps and so forth. So yeah,
River Roberts: And
Andrew Davis: I've as as you as you said my background roughly, you know, early life, trained as an engineer, spent 15 years as a Buddhist monk in the last twelve years, built a career in tech. And the last year I've been really honestly decompressing and not building so much as deconstructing and just kind of taking some time to rest. and try to get more embodied and yeah learn new approaches to life.
River Roberts: Yeah, beautiful. Thanks for sharing. And and just for a little added context, Ty, because you you you weren't there in this instance, but so AI Builders is a collection of different I builders throughout Europe and they host a bunch of different events. And it just so happens that one of my friends put that together, Arthur, who I'm sure will bring on at some point. But he asked if I could help facilitate for the day. And as part of it there were these kind of networking opportunities. And so I created a scenario where Everyone, I think it was like 10 minutes, and basically the invitation was everyone would ask one person what they needed help with, and then there was change partners. And then the next round was you would ask your AI agent what you personally needed help with, and then you would tell that to someone else, and vice versa. And then the third round is you
Ty Pattison: Ha ha ha.
River Roberts: and then th and the third round is you would blend. the the answers from both and then see where you got to. But basically by the end of ten minutes you had a hundred people that had kind of woven in how can we all support each other and ourselves in a meaningful way. And there was a real kind of energy in the room after that. so yeah, it's a really beautiful practice and and definitely, you know, recommend any facilitator in that kind of a situation to to adapt and adopt and and play with a format like that. It's a lot of fun.
Andrew Davis: And i if I can double click on that just for a moment, this specific prompt that Rivers suggested is just and this is a real danger that the you immediately lose all of your listeners into this exercise right now. So if you can restrain but do it afterwards, go to your favorite LLM and just type in what is my main problem and submit. And when it when I you know got the response back. It said, your main problem is lack of focus. You know, you're doing this, you're doing this, you're doing this, you're asking me about researching this. And I'm like, my god my gosh. I'm like, you're so right. I was so but it's funny though I then did that same exercise with a group of with another group of of people that I was working with and It was really interesting to hear everybody read out what the LLMs told them their prop main problems were. And most of them I actually quite resonated with most of the responses that the LLMs gave. And what it reminded me of was a horoscope reading. Like, you know, there's twelve plus Like what's your main problem? Well, because you're a Sagittarius, your main problem is, you know, lack of self-confidence.
Ty Pattison: Yeah.
Andrew Davis: And it it actually helped me to make a little bit more really at great risk of sending the podcast spinning wildly off topic, but it helped me to make more sense of things like astrology, where they just literally, you know, odds are you probably do suffer from you know, any one of these issues, but this is the one to focus on now. And so just it provides a little confidence and certainty that for now this is the thing for you to focus on. It's like confidence as a service in a in a time where there's just too much information, too many possibilities, to have someone or something say, This, this is the thing. it's actually a service.
Ty Pattison: Yeah. Yeah, absolutely. I mean two things. The podcast is called The Great Spin Up. So any kind of spiral that we're going into, fantastic. Let's go there. and and you're right. And it it makes me think what proportion of people in this day and age, and especially using LLMs, like if you've been using an LLM for a long time, how much of that How much how many people could you give that answer to? you have a lack of focus? And it would be the number one problem that they have. like yes, it knows about you, and yes, it has these things, but like just in the same way a horoscope is written to be semi-ambiguous, like there's a formula behind horoscope writing so that it applies to enough people. I'm sure there is also a way that these problems come out that yeah, that applies to me and it's useful.
Andrew Davis: Yeah, it's just like, yeah, you just have that all of you carbon based life forms tend to have, you know, this lack of self confidence and, you know, whatever else, not putting enough energy into relationships or whatever.
Ty Pattison: Yeah. Yeah. You just need to sit still for for five hours and then you'll find that or you can ask an L O Cool. River take
River Roberts: All right, Jens, how about
Ty Pattison: us away.
River Roberts: yeah, how I was gonna say, how about we get stuck in with the news? What what has jumped out to you this week or just in general as signal in the industry over noise?
Ty Pattison: I've got something to to jump in with. So there's a company called Guidelite, Guidelite AI standards.
River Roberts: Okay.
Ty Pattison: And they seem to be credible enough. I looked into it not extensively, but enough to make sure that it was that they were doing real work. and they basically rate companies on their safety standards because there have been a few things in the news recently about models breaking out of containment. and both OpenAI and Anthropic scored the highest with a C plus rating and Meta got an F on the on their score. and specifically like why it's why it's really relevant is because we're entering an age where more and more agents can use computers, more and more a agents can take action.
River Roberts: Mm.
Ty Pattison: And when it's just giving you an answer and you have to take action, it's not as critical or it it's not as important to pay this much attention. But now agents are getting the ability to to make things happen in the real world. So these things I think are gonna get more and more we need to start paying a lot more attention. Yeah.
River Roberts: Did did I'm wondering, did did this actually look into some of the open weight models as well? And open source models?
Ty Pattison: there was a piece in the article that talked about the open weight models and it basically just said that they are catching up, but I don't th like that their capability is catching up, but it didn't talk about this company reviewing them. so long story short, I don't have the answer because I didn't look deep enough into it, but it's an interesting question.
River Roberts: Yeah, so of course bring it on.
Andrew Davis: Am I allowed to ask dumb questions?
Ty Pattison: Yeah, yeah, yeah, please.
Andrew Davis: when I think about meta, I think about Lama as the model. And I'm not sure i you know, Lama I think of as I'm not sure is it is it an open source or open weight model or is it just free and you know, you can run it locally?
River Roberts: Yeah, o Lama is open source.
Ty Pattison: My
River Roberts: You you can run it locally. and then they've also started to release some other future open weight models under the kind of Muse Muse banner. So they've got a couple Muse Spark and Muse Glimmer as the first few ones. and you know, they they were kind of early to the game in the open source and they c kind of haven't actually done anything in the past six months. So this is quite a recent recent piece. Any anything to add to that, say?
Ty Pattison: honestly no. I was gonna make a distinction between open source and open weight. just the open source you. I yes. Good, okay. It's parking a lot it.
River Roberts: w we could we can do that in the deep dive though. We have a whole deep dive today on distinctions and definitions in the industry, so we can actually start to unpack these things. get into the nuance of them. I I I listened to there was an incredible podcast on moonshots just a couple days ago with this guy, Alvin Graylin. And he's spent a lot of time working in the Chinese ecosystem. And one of the things that he he talked about was the the Chi Chinese open weight models seem to be scor scoring a lot lower on cyber offense capabilities than what the American frontier models are doing, which is why I thought it was kind of interesting in comparison to that first bit of news that that you brought up. And it seems that a lot of these open weights open weight and open source models, even though they are they're open, they still have to go through the Chinese CCP approval process. And as part of that, the offensive capabilities seem to be, you know, like like a quarter of or if that of what these American frontier labs are having. So yeah, it's interesting seeing the the out of the box like what the offensive capabilities are and then maybe what the security vulnerabilities are that they're opening up into the industry. And then particularly in China where you already have an ecosystem that is kind of being short up, you know, the the Chinese internet has been kind of secured and firewalled to some extent for you know a couple of decades now. So I th it it it's looking like their approach to to strategy and security has actually been much more effective at this point. anything else on that that piece of news?
Ty Pattison: Hm, just the thought that cyber offense is usually easier. I mean, offense in general is usually thought as being easier than defense. And the labs potentially are doing something to hamper the offense because They're like if you just leave these models to their own devices, you would assume that they would have the same capabilities as all of these other things in life where you only need one attack vector in order to like get into a system, whereas you have to defend against everything. So yeah,
River Roberts: of them. Yeah.
Ty Pattison: interesting. And also you don't want to be shipping out cyber security weapons anyway. So good. Good.
Andrew Davis: But does it does seem to point to that there is maybe behind the scenes some more robust infrastructure You know, so that may indicate, you know, that that if the US had better systems, we could also mitigate some of those those risks, more centralized systems for and and but implicitly
Ty Pattison: Yeah. Yeah.
Andrew Davis: implicitly it also means that those Chinese government review agencies know what good offensive weapons look like, know what good cyber cyber offensive tools look like and how to
Ty Pattison: A really good point. I'm gonna jump in here with another story because it dovetails into this quite nicely. both Anthropic and OpenAI have models that are not released. they're holding them back from release. And when we're talking about like these these Chinese models putting their these Chinese companies putting their models out. And it it was making me think this morning about the book Superintelligence. Max Tegmark talks about if you're gonna create a superintelligence, you first of all need containment and then you need to bring the superintelligence in. It doesn't make sense to it's crazy to make a superintelligence and then try to contain it within a system that is built by something that has less persuasion, less ability to write code, that it that doesn't have these capabilities. So it's really interesting seeing both the stuff coming out of China and also these US labs like having these models that are built. So they've actually built the systems, but then they're not releasing to the public. So
River Roberts: Mm.
Ty Pattison: that was a piece that was jumping out. And the one line that came through is that now the frontier seems to be moving at the speed of containment. Like how slowly or how quickly can we figure out guardrails for these models in order to be able to release them to the public and it's not at the speed of innovation anymore.
River Roberts: Mm.
Ty Pattison: which I thought w is an interesting kind of slowdown step, hopefully. Whether or not it plays out like that, I don't know.
River Roberts: Yeah, I mean that that's that's a huge piece and and also just, you know, the speed of of co of of how quick cultures can move. I mean, the last few industrial revolutions have happened over, you know, eighty and forty years, sixty years. and and we're expecting, you know, something to happen within five years and you know, maybe orders of magnitude in in terms of its its its speed and and influence. So not only moving at the speed of containment but also moving at a at the speed in which cultures can actually adopt and and implement this. And that was another big thing that came out of this Alvin Graylin podcast was that the Chinese models and the Chinese government seem to be working a lot more on focusing on how these AI models can find their way into industry as opposed to racing towards AGI. And that seems to be quite a big distinction. and so there's there's a very different
Ty Pattison: Mm-hmm.
River Roberts: race at hand, even though culturally it's maybe positioned as you know, winner takes all, zero sum. It it it does appear that maybe the Chinese are playing a slightly different game than the American Frontier Labs and the way that's presented in in mainstream media. yeah.
Andrew Davis: The I'm I'm thinking the the the Metaphor of nuclear power is used a lot, of course,
River Roberts: Mm.
Andrew Davis: and comparing
Ty Pattison: And
Andrew Davis: you know, trying to make sense of what's happening here. But when you talk about containment, I'm thinking about Marie Curie and these early researchers on radioactivity, right, in the late 19th century, where they didn't know the the effects on the body, on on biological tissue of radioactivity, right? And so all of these early researchers were dosing themselves. Way, way, way, way higher than any modern safety protocols would allow. And you think about, you know, we all grew up knowing about radioactivity and atoms and these kinds of things. Mid-19th century, nobody understood anything. It's complete, you know, ridiculous science fiction that there was some invisible waves that could, you know, you don't notice anything, and all of a sudden you're being poisoned. but They they basically tapped into something that had been there all the time, an invisible force, could be used for good or for eat or for ill or whatever. and if you think about obviously that's relevant to the containment resource. If you think about nuclear fusion, obviously, yes. Step one, build really good containment. Step two, you know, create a mini sun.
Ty Pattison: Just get pushed.
River Roberts: Yeah.
Andrew Davis: yes, that operation is extremely important.
Ty Pattison: Hehehehe
Andrew Davis: But you you think about what are the invisible structures that AI entities could be messing with? And they are ob obviously infinite, but one of them is the i i i i to use a sort of a fancy term, like the semantic field in which we all live, right? Our sense of reality, our sense of what's true or not.
Ty Pattison: Mm.
Andrew Davis: And there's a sense in which all of us build our orientation in the world, our sense of reality around the use of words, the use of language, you know, what's true, what's not true, and so forth. And AI entities immediately, you know, I keep thinking about all of the AI-generated content that's out there. If you go look at LinkedIn or whatever else, and you know, this device about, you know, Of creating contrast, this rhetorical device of creating contrast. It's like, it's not just a rhetorical advice, it's a whole different way of thinking, you know, where these AI entities say, you know, it's not this, it is this. And
Ty Pattison: Yeah.
Andrew Davis: modern writing is just littered with that. And you can tell it does something, it creates this real punchy effect in terms of the reading. and it's it'll create effects on our psychology at some point. And then this question about what are the invisible structures that we're living in already that we don't fully understand, what are the risks, you know, that we maybe we all we do understand to some degree, but not not fully, and we don't have protocols for we don't have safety protocols like like
Ty Pattison: Yeah, when you're talking about this and you're talking about radiation, like I come back to this this saying that what we create creates us. And just like the simple fact that we're we're animals, we're creatures that learn through repetition, having exposure to these models and like having them echo back this form of writing, th these like AI writing patterns, I would be very surprised If you ran a study like a year ago and in a year's time you had enough of a gap for people to have exposure to this stuff, if people's writing habits didn't get drawn towards this thing just because of the the volume of literature that they're exposed to that is written in this particular pattern. yeah, it really fascinating.
River Roberts: All right, Zents, I got another piece of news. now this isn't actually new news, this but this is new news to me this week. And this is a report that a research report that Deep Google DeepMind actually put out in March this year. And it was it's one of the largest studies ever done, and it's with over 10,000 humans across the US, UK, and India. And they were basically doing a psychological study to see. if LLMs could manipulate humans at scale. And they were testing it across behavior changes in finance, health, and policy domains. And the answer was yes. But actually the the the the bit that is and they they tested three conditions. So one where there was like no chat, no, no kind of input. One is where the manipulation was very explicit and it was spoken out. And then there was one that was kind of implicit and it was like a little bit more implied, or as they kind of described it, as a little bit more of a whisper. And although the the manipulation was not the same in each instance and in each culture, they did find consistently that the more powerful psychological behavioral changes actually came from the AI agents that tended to whisper and and and work it into more subtle aspects of the conversation and kind of walk you down the path. And yeah, I mean, again, early, early signs that we're working with something here that has you know huge capability and you know how do we be mindful of that while we're building and and working with them. yeah. Thoughts?
Andrew Davis: If I can if I can offer one I'm I just moved to San Francisco six months ago and You can't hang in those circles without pumping into plant medicine, the term plant medicine a lot, right? And people talking about, you know, fungus and these these these plants that have psychoactive effects on our minds and to what degree maybe there was a co-evolution and so forth. and you know, it it's a popular thesis or popular point of view that, you know, these plants are actually acting through our affecting our cognition in ways that allow us to see different things and so forth. So this idea of whispers from AI entities influencing our thinking seems quite analogous to that. Something that is like a like a psychoactive substance or a psychedelic substance that's being introduced to the population and perhaps like psych psychedelics and you know, it's very distributed. The effects are highly dependent on you know set and your mindset set and setting the environment you're in and the mindset you're already in, so it's get highly variable depending on the people and so forth. But You could easily imagine you know, population-wide effects that are slightly shifting and in in insidious ways, in ways that are that are not obvious. I don't mean insidious necessarily in a in a malicious way, but like hidden effects where there's like some where where it's one thing to detect, does the AI escape containment, you know, and get access to the internet. It's another thing to detect, you know, is there some statistically significant change in cognition of people who are using, you know, Claude versus Chat GPT versus Gemini, whatever else? Do these different LLMs have, you know, some slight effects on the populations of their users that are Psychoactive, psychoactive substance. Should you be twenty one years old? Should you have to be twenty one years old before you are allowed to use LLMs?
Ty Pattison: Yeah, interesting. I mean it's so 'cause there are lots of things that are like this, right? There are I mean not a lot of things that that fall squarely into the same categories, but there are a few things, whether we acknowledge it or not, that really do have an influence on how we develop. there recently it makes me think of social media. Recently in Australia, I believe social media was banned for under sixteen year olds. Is that right, River?
River Roberts: It is indeed.
Ty Pattison: and there's a couple of other countries who are like piloting programs like this because social media like really has this ability to to grab our attention and do it in a way that we're not often aware of. It's insidious in that way. And there's the same thing with with AI, and but instead of Who is it? Sam Harris talks about this. the what was he at the Institute for Humane Technology, and was talking about how social networks tap into our brains and take attention. And these AIs have the ability to do that, but for connection, which is a really, really p if not the most powerful human emotion. So yeah, potential to like shift the whole population in that way. and maybe to put it on like a connection versus loneliness axis as well. Like if you can get all of your connection from this, maybe you learn to that's the place that I go every time I want to connect with something because it's easy and there's no friction and there's like it's trained to specifically do this with me versus if you're like in these circles, at least you're having conversations with other people, right? Like the AI is the third entity, not necessarily the go to.
River Roberts: Yeah, I was gonna say that's that's the distinction I was gonna say that's the distinction that I'm hearing in this as well, is that you're being manipulated by a relationship that you're not having with anyone else. It's just you personally, in a way. At least in social media, you're being manipulated a little bit more out in public and it's out in the open. Whereas these these conversations are happening privately. but again, manipulation doesn't always have to be necessarily perceived as negative. And I have to say that At least I feel like I've been manipulated very positively over the last couple of weeks in terms of trying out this this ketosis diet, which if like getting enough information and and nuance on having this protocol really work for me, would have taken me so much more time, but actually having an agent there supporting me and be like, you know, can I have a sip of coffee or can I have a spoonful of honey? And it was it's it's incredible like how refined it can be. And you're like, well actually I feel my health has improved and it's been manipulated by you know, this agent. Now whether or not there's a company underneath this thing advertising to me and trying to push, you know, electrolyte purchases as based based in in like something that I need a supplement that I'm not getting when I'm on ketosis is a very different thing. And
Ty Pattison: Mm-hmm.
River Roberts: actually that's a another piece of news that came up. OpenAI is rolling out ads in in I think what, 31 European countries. So I'm very curious to see how that has responded, how that is received. because it seems they were kind of shut out of the American market with that Super Bowl, that Super Bowl commercial that Anthropic put up to put a OpenAI out of the ad game quite quite quickly, or at least to start off with. So very interesting to see how that will be rolled out in Europe now. Andrew, do you have
Ty Pattison: Amen.
River Roberts: a piece of news you'd like to to share with us?
Andrew Davis: yeah, I do. It's not n super new new, but it's it's an article from it came to me from Rutger Bregman. this is a graph of capital expenditure
River Roberts: just can I just stop you can I just stop you here because you have brought the first graph we've introduced to this podcast. So thank you so much for elevating this this podcast with its first graph. I feel like I might need to create some sort of certificate that you can print out one day and hang on your wall.
Andrew Davis: It my my love language is URLs and graphs. So on
Ty Pattison: Perfect. I I think we sorry to interrupt, I think we might need like a section here which is like Andrew's grand graphs or like whenever someone brings a graph have a celebration of like two we love graphs.
Andrew Davis: But it's I'm feeling very very warm and fuzzy to be received, to have my my my graph fetish so warmly received by you there. I appreciate that. so on the vertical act.
River Roberts: It's it's a great graph. I can already I can already picture Trap with a you know, with a printed out on a sign and his little stick. It's it's a fantastic graph.
Andrew Davis: I'll I'll print it out and sign it with a gold Sharpie if you if you get a chance.
River Roberts: I appreciate it.
Andrew Davis: So on the vertical on the vertical axis is amount of dollars expended in today's dollars. On the horizontal axis is the number of years over which that capital was deployed. And so if you think about the US railroad program, like the mid 19th century to the early 19th century, deployed 550 billion dollars, but over the course of 71 years and transformed the country, right? Massive shipping. And then the interstate highway system, you know, some decades. Later, six hundred and twenty billion of today's dollars deployed over thirty-seven years, as obviously completely transformational for the United States. And this is a bit of a US, you know, centric graph. Apologies for our perennial selfness.
Ty Pattison: We'll extrapolate, it's all right.
Andrew Davis: Anyway, we we we were the ones that sent a man to the moon with the Apollo program for just $257 billion over 14 years. Manhattan Project, build a nuclear bomb for just $36 billion in five years, and then rebuild Europe, starting with Germany with only $170 billion over four years. And then what do we have up here? Data Center CapEx, Projected at $930 billion over six years, and this is already a few months old. The actual planned projections this year, and this this data comes from the Brookings Institute, well over a trillion dollars. Planned CapEx with the intention that that will actually, the infrastructure expenses will be done over the course of the six years. And I noticed just today that this excludes the Chinese hyperscalers. So, you know,
Ty Pattison: Wow.
Andrew Davis: we don't even know what the Chinese are deploying in terms of this. And so if you look at the big picture of this, and this is setting aside all of the software investment, right? The investment of time and energy and you know. This is just raw computing capacity. Far and away, the biggest infrastructure project humanity has ever done. And I don't know where the Chinese, you know, what was it? Three Rivers, I'm gonna masquer three rivers, six one belt, one road,
Ty Pattison: They have one belt, one road. Yeah.
Andrew Davis: yeah. Where where their infrastructure projects.
River Roberts: Yeah, the three vis the the gorge. Yeah.
Andrew Davis: Yeah. and but over six years, right? So the intensity of the investment and what this makes me think of is that mostly these are you know startups and venture capital and they want they expect return on investment you know typically in five years that's the venture capital timeline right so there is this idea that you're putting all this money in There's gonna be whatever else might be true, there will be enormous pressures to reclaim, you know, recap that investment very quickly. And then people will be doing whatever they possibly can to manipulate the world such that you can get your money back real fast. and I believe that that's just a driver for so many other things happening.
Ty Pattison: This is such an interesting perspective and especially from the conversation that we were just having about how potentially AI is is seeing these things makes me think of the what is it in sapiens Yuval Noah Harari makes this argument that we didn't domesticate wheat, that wheat domesticated humans. And this building out this amount of infrastructure this quickly feels like the AI building its own infrastructure to be able to live. And whether or not that is just where all of the computers and the the, you know, artificial intelligence goes and lives and it's separate from the infrastructure that is built for humans is the question that comes up for me. Like, yeah, we're getting like all of these humans to build this thing. Is it for us or is it for somebody else?
Andrew Davis: reminds me of I've got a friend and colleague named Reese Lindmark and he was working for a long time on I think it's sub it may be a Substack post or what blog is something is called what Patterns want, what patterns want. And this idea that patterns have a tendency to actually draw out. Like if there's certain patterns, something that has, for whatever reason, a mathematical geometric p you know, s strength, stability, it tends to draw things out. That there, you know, and this reminds me of Michael Levin's work on platonic spaces, and that there are there are potentials that are trying to be drawn out.
Ty Pattison: Yeah, yeah. What are the truths about the world?
Andrew Davis: in right and what are the what are the potentials implicit in these you know gps that are patterns that that want to draw out more value want to draw out more and again i'm I'm with river like we I don't I don't think it's we shouldn't be overly suspicious or unnecessarily suspicious it is extraordinarily Potentially benef beneficial. I think humans do need to be hacked, have their minds hacked and redirected in a bunch of ways that would be more beneficial for us. and we do need more intelligence as a species. I don't think we're suffering from a surplus of intelligence. I'm certainly not. but we do need to be careful how it all unfolds.
River Roberts: Yeah, I mean it is
Ty Pattison: And and you're right. And we do that for each other. Sorry, go ahead River.
River Roberts: No, I I I I feel like there's a a a real risk here in many ways that is is actually potentially quite similar to the Soviet Union during the Cold War. And so during that war the USSR invested about twenty five percent of its GDP in military. And part of the reason they actually lost the war is they they over invested, they overindexed in the military machine that actually wasn't supporting its citizens. And so I wonder if, you know, the the future doesn't repeat itself, but it certainly rhymes and that maybe there is this massive overinvestment. I'm not saying that investment doesn't belong here, because it obviously does, but it's at the scale and the speed at which it's happening, is this actually an overinvestment and overindexing in this in an industry that hasn't necessarily, and particularly from the US spec perspective, hasn't necessarily shown a lot of benefits to the industry. or it hasn't necessarily reached a lot of different aspects of different industries so far. In fact, you know, the way it's positioned in the US is basically I think some of the statistics is 70% of the US is white collar workers. And this is precisely the technology that is going to put most of those white white collar workers out of work and displacing them. So you're accelerating the investment in a technology that is displacing your your culture in a way. Whereas you look at other cultures and countries around the world, that seventy percent is not the same. So in Africa, in Asia, there there's a lot more of the industry that is still you know working more blue collar jobs and and service layer jobs, not so much white collar. So I think there's a real risk in this accelerated investment and the expectation of you know huge financial returns that support this this upside risk. yeah, I mean it's it it could really be creating the conditions for a pretty tough time in America as a as a as a nation state in some regards. You know, and this is kind of
Ty Pattison: Mm-hmm.
River Roberts: this is kind of what happens in in in various ecosystems as well. Like if you have a too many large players and too they they kind of tend to monopolize the whole industry and everyone tries to follow behind and then the whole thing kinda collapses because there isn't enough diversity holding it up. So I wonder if that's the conditions that are starting to emerge.
Ty Pattison: Mm.
River Roberts: But yeah, thanks. Thanks for that graph. That was a great first graph.
Ty Pattison: knocked it out of the park. I don't know if we can get a better graph, honestly. two two quick points. Just the graph, the like extreme outlier of of that data point is is wild. and River, I don't I don't necessarily know if I agree with you on the white collar work. I I think it's so nuanced here. Like yes, we are having the ability to replace all of this white collar work. And the tension here is between the timelines for me. Like Andrew, you're talking about the VC runs on a five year timeline. That's gonna be a crunchy thing. There are lots and lots of jobs that you can't write down everything you do in a day. There are lots of white collar jobs where we actually they're much more diverse and nuanced than we think about. And so like across the world we have different kind of mixtures of these things, but I think there's a real opportunity in applied whatever it is. Like we probably won't have that many pure mathematicians, but we have a lot of applied mathematicians in different fields. We might not have like that many theoretical physicists, but we'll have people who are doing research using these tools that are then applying them to the world. so I don't know.
River Roberts: Totally. And and just just so I'm clear.
Ty Pattison: To me there's like it's quite crunchy.
River Roberts: Yeah, I I mean I absolutely agree with you on that point. And and just so we're clear that I I I don't think these tools are gonna replace people. I think it's gonna displace people. So for instance, we're already seeing I think forty to fifty percent of the youth market in America are underemployed. So I think the unemployment rate is something like ten percent. But then when you look at the underemployment rate where a lot of these people aren't, you know, working up to their educated standards or the university standards, it's it's underemployment. So I think there's going to be this massive displacement. And that's the crunchy bit that I think we're talking about. And again, when you're dealing with a culture where, you know, 70% of the the the culture could have this displacement. I think and you've got this political unrest and kind of bifurcation of of the culture in some regard could get really kind really crunchy, really crunchy quite quickly. Yeah. You know, I feel like I'm just saying crunchy a lot because I haven't actually eaten anything crunchy in two weeks. So so if you g so if you give me like I can't like it's like meat, eggs and fish. It's nothing is crunchy, man. I'm really craving a crunch. I've got one more quick bit of news, and then I'm not sure if either of you do either. But one of the things that jumped out this week for me was the Tsingha University out of China and Bidance recently released Coda Agent. And this is an open source system that writes the GPU kernels 100% faster and 40% better than Claude Opus 4.5. And so basically CUDA is the software layer that the NVIDIA chips have been operating on. And it allows the NVIDIA chips to kind of work across all the different chips. And that software layer has kind of been part of the competitive advantage for the last couple of years. But now that we're seeing that software moat potentially start to evaporate, you know, what w how does that actually change the the the chip layer of this industry? yeah, that was just a a a quick one that popped up this week.
Andrew Davis: And just so I can understand that, so the idea is that you might be able to write an alternative kernel, a kernel that's an alternative to whatever Nvidia's native operating system is for the GPUs?
River Roberts: Ex exactly. And then you then you have impr interoperability across all different chips. So you're not just tied to NVIDIA chips,
Andrew Davis: play.
River Roberts: you could use AMDs, you could use Cerebrus, you could use all these kind of different chips and and work across them. So there's this huge upside in in that side of it. And and I I I again I think we're gonna see more of this kind of open source agentic infrastructure emerging, which is is super exciting. two other two other quick s news stories. I
Ty Pattison: Mm-hmm.
River Roberts: just want to mention the high lines. One was there was an Interpol report that fifty-five percent of African cyber attacks include AI now. And DARPA flew its first AI piloted F sixteen. So aga again, there's there's some
Ty Pattison: Whoa.
River Roberts: pr some pretty crazy moves happening in the industry right now. and the last one, which I think is a r really
Ty Pattison: I have question.
River Roberts: big signal piece, just quickly tie, is Stripe is buying open router. or they're setting up this this purchase at the moment. Stripe is buying open router for over seven billion dollars. And so open router is this layer where you can switch your models when you're working with different agents. So it's kind of proving out that this switchboard layer is actually the the financial infrastructure for a lot of the industry. So yeah, huge signal there.
Ty Pattison: Could be a a big rabbit hole, I'm not even gonna ask right now. who's
River Roberts: Mm-hmm.
Ty Pattison: DARPA?
Andrew Davis: The Defense Advanced Research Project Agency. It's the it's the US military R and D folks, famous for creating things like the internet.
River Roberts: Yeah, the underlying protocol for the internet, yeah.
Ty Pattison: Okay, interesting. Yeah, I was wondering which country. Cool. I have one quick piece of news. We don't need to dive into it, but the former a former Google Chief Science Officer, his name's Jeff Dean. basically Yeah, Jeff Dean left Google after twenty-seven years, started a company called Discovery Loop and is doing it with three other people. And this was interesting because we have lots of young people saying, you can build things with AI. This is evidence that people who have been in big tech for a long time are going and making bets on their own and not hiring big teams because they can do so much with so few people. and especially if you have expertise in a particular domain, I thought this was really interesting. Yeah.
River Roberts: Yeah, yeah. I feel like that's also a massive signal kind of almost against the culture of Google at the moment in terms of they're such a large company that they're they're obviously quite risk adverse. but then also in the same breath, Sergey Brin also came back on as acting CEO to kind of put Google back into what's effectually referred to as founder mode. so we'll see if there's some some different moves coming out of Google in the in the coming months. All right.
Ty Pattison: Yeah, interesting.
River Roberts: So shall we move this ship along to building?
Ty Pattison: Building, yes please. And I like that you use the word ship. I'm gonna take this opportunity to jump right into this because a river I have some insider knowledge that you have been building, working on something that has to do with a ship. Would you like to jump into this?
River Roberts: Sure, let's let's let's start off with this. I look, I I'm a little bit embarrassed to be sharing this publicly, but at the same time, you know, I did I did kind of in the back of my mind line this up and maybe not give myself the best conditions to produce the best result, but I'm still kinda happy with the amount of time and budget that went into it. So basically there's something called the future vision exprise, which is a call for these kind of like Pro-topic visions of the future. There's kind of this understanding that Hollywood really only produces dystopic visions or these like really boring polyannerish visions of the future. And so this is an X prize to to to get a bunch of these visions. And I've been inspired by a lot of sci-fi and Star Trek over the years. So I figured I would I would give myself a a couple days to explore with the current tools what I could produce for this three minute trailer. Now give it in my I have no experience with writing scripts or anything with films or any of these video tools at anything. So I started from zero. And with that being the case, how about we just push play?
Ty Pattison: Nice.
River Roberts: all right. How embarrassing.
Ty Pattison: Cool. not embar I mean claim the embarrassment please. what what sorry go ahead Andrew
Andrew Davis: Yeah, I mean I I don't know, it maybe that's the most instructive thing for people to to see that one could be embarrassed by producing something like that. I thought it was marvelous, bravissimo, right? This directorial debut review is fantastic. I I I think it's quite important to the Protopia thing, obviously super important, right? And then to figure out How to knit I mean, it's so much, you know, we all complain about this too much negativity on the news, right? So just to cast out positive futures, right? of various forms feels so important, right? And I I love how evocative this is. and that, you know, it was a it was a two day project, and yet you'll be embarrassed. So this I think is really instructive for everybody about how you know, our self critical nature is so hyperactive. I really appreciate your courage and putting it out there 'cause I think the you know, we all need to be trying to put forward positive creative efforts.
River Roberts: Yeah, totally. Well, yeah, thanks. Thanks for that. And so I
Ty Pattison: River.
River Roberts: spent about three days on it. I spent about fifty dollars worth of credits across a few different platforms. And again, I'm coming from like nothing. So I've never written a script before or a storyline or any of that stuff. I hadn't used any of these tools before either. So it was a really interesting learning curve in terms of like timeline, budget and these new sets of tools. And so with that being said, I'm I'm quite happy with the outcome. And and yet I think if I was to do it again, I'd do a little bit more work on like the story side of it and a little less on the the the visual sides of it. 'cause at least in telling a compelling story, I feel like I should have spent a little bit more time on this this the script of it, 'cause that informs so much more of the actual build. So getting the sequencing right, I I definitely learnt a lot through that. But I will say I actually really, really enjoyed the process and again, pretty impressed with what these tools are capable of given you know the input that they had from a complete novice. So yeah, definitely enjoyed the process. Even though sharing it is a little bit embed it a little bit embarrassing.
Ty Pattison: Can I ask what's the
River Roberts: Please.
Ty Pattison: Wha what was what was the inspiration? What pieces are you weaving together in here? Are they long like pulled in things that you've been thinking about? Is it recent? What's
River Roberts: Yeah, it's yeah, so s
Ty Pattison: the
River Roberts: so I d like for me part of like one of the bits of inspiration is like, you know, Star Trek is this kind of future vision, but maybe what would the first vessel that ever kind of seeded something like the whole Star Trek thing be like? So I kind of saw kinship one as, you know, before we have starships, we have like obviously a ship that is on our planet and it's as much about exploration and sharing as anything else. It's not coming from a place of economics or military or politics or even non profit. It's it's actually just it's coming from a very different place. And that's what I was trying to capture. But you know, some of the feedback that came back from some close friends is like, yeah, this feels a little bit elitist and and all of that. So I know it didn't quite hit in that regard. But that was part of the inspiration. And then two films that really kind of called to me. One is a near future film that came out a couple of years ago called The Creator. And I kind of loved the imagery of that. And then on the other end there was this comedy called The Boat That Rocked. And I just love the juxtaposition between those two. So those were kind of the the inspiration. It was kind of like Star Trek meets the boat that rocked and the cre the creator. So those those were the the the three references I really pulled from.
Ty Pattison: is a fantastic film of like comedy and drama and it's so human
River Roberts: Totally. Yeah. Yeah. And and like I think there is something about this like pirate essence that actually moves cultures forward in quite meaningful ways and you know, pirate cultures back during the European era of of the spice wars were actually some of the more demographic demographic demographic democratic societies that were emerging. The the pirates had this kind of like code of honor and and a way of being very democratic in the way they're operated. So there's something about the seas and the oceans that create the conditions for potentially something new to emerge without us having to go and be space cowboys. So that was some of the inspiration. so the so the tools that I the tool the t the tools the tools that I use so I was using largely Kimi for research and and dialogue and and all of that which is Maybe not the best tool I should could use for that part of it. And then the video tools were Google Vio three point one, Kling three, which is a Chinese model, and then LTX. those were largely the video models. And then for the film score, which I actually found really hard, but the one that I actually ended up using was one that I just one shotted at the first time I I put a prompt in. And I've probably tried about five different scores, but it's quite hard to get it to match up. And that first one kind of matched up close enough. I was like, that's good enough. but that was with Google, Google Lyra. And this tool is incredible, L-Y-R-I-A. Its ability to produce music scores with you know a couple of lines of prompt is is phenomenal. So definitely worth exploring more and more. and then to
Ty Pattison: Can you explain that for me a second? W what do you mean by a music score or a film score?
River Roberts: So the the the music track for that whole trailer was produced when in one prompt. And
Ty Pattison: Uh-huh.
River Roberts: so the music layer is typically called the score. So the score of the film. and I and and the vocals were done as part of the LTX video machine. So you can put some you can put some script in the video prompt and it will produce audio. Which is also where the audio quality was a little bit all over the place, but it I'm like it's good it's good enough to to kind of get get the the trailer out. So yeah, all of these tools very interesting and definitely worth exploring.
Ty Pattison: Just one question on like tactics or like making this stuff happen. Do you use one platform to use all these different tools or did you go to each different platform and like play with them or download them and move them? Like what's that process?
River Roberts: Yeah, so to compile it all together I used a open source just Mac editor called DaVinci Resolve. and then in terms of producing each element, yeah, I actually had I I kind of went to each individual platform. So I used Cling, I used LTX, and I used Google Flow. and those were the three main ones. And since then I've actually come across a a An ecosystem called FAL.ai, where you can actually use a lot of these systems as well. And the way they're actually set up is that you can plug these systems into your API. So actually actually start running them through like a claw code or one of your your main agents. But I was like, by the time it takes me to set that up, I think I would have lost a bit too much time. So I just I just spent time within each each ecosystem rather than trying to bridge them. Yeah.
Ty Pattison: Yeah.
River Roberts: Anyway, I don't wanna take up the h the whole build segment, so I'm curious what have you gents been building lately?
Ty Pattison: Yeah, Andrew, do you do you have something you wanna talk about?
Andrew Davis: Sure, yeah. I I mean I I'm just soaking up both your tech stack and the creative output and I'm I'm off sort of daydreaming and fantasizing river about trying to follow suit be a be a stream to your river. so what I'm working on these days is something called weave. And so it's basically I I put the URL in the in the in the chat here, maybe we can add in the show notes, but weave if you go to search for weave and my company Living Code, Weave is basically a programmable Zoom. so it allows you to define in advance some kind of a structure for meeting to orchestrate a meeting. We've all been in meetings where there's like fifty people and one person droning on and on and on and very, very low engagement from participants. And hopefully many of us have been in you know online when you when you have to do an online meeting Many of us hopefully have been in environments where there was a really good orchestration of breakout rooms and giving people prompts, putting them in groups of two or three or four to discuss these these ideas. and maybe you could do a ver maybe we could do a version of the spin-up that's live and interactive, right? Where we use Weave to choreograph, like asking people to come together 'cause to discuss. 'Cause I I think the thing about video meetings is they're th because it's two-dimensional. And it's so passive, typically we get very low engagement of our brains, literally. And so people then their level of participation is very low. And so most companies people end up drowning in lots of meetings where there's not really much engagement during the meeting, then there's also not much traction. So what weave, like from meeting to
Ty Pattison: Yeah.
Andrew Davis: meeting, capturing I capturing the ideas. So what weave allows you to do is basically define in advance what are you meeting about. And then some prompts. And you'll take the prompts and you'll go into breakout rooms and Weave puts the prompt in front of everybody in the breakout room. so you can always see what's the prompt. And it allows you to record inside each of the breakout rooms and transcribe inside of each of the breakout rooms, or opt out. And you anybody can declare that they want this to be anon that breakout to be anonymous, but to begin to capture the actual fullness to to both instigate a fullness of discussion. Discussion, interaction, collaboration, creative collaboration in the context of meeting, but then also to capture and harvest the ideas that are generated and to consolidate those ideas. Future iterations, my goal is to you know both Right now I'm creating text transcripts, but I'd love to be able to even create visual artifacts. like if you've been in these, you know, sometimes there are meetings where people will draw live during the conference, you know, some of the main themes. So obviously it's very you know, these days the technology is very conducive to doing that live. And like let's say we have a discussion, what would be a single visual that encapsulates the whole of this conversation in some kind of an evocative way? in a visual graphic way and that gives a way for our memories to anchor key ideas and then I'm imagining also you know scaffolding that up so each breakout room has their visual but then there's some composite visual that encapsulates elements from the whole organization and just basically doing more to light up everybody's minds and activate their memories together in organizations and nonprofits.
Ty Pattison: Mm. What you're talking about too makes this like thing of of meetings being flat, of feeling flat. They're very intellectual. You're in your head all the time and decisions are made emotionally and memories are stored through senses. Like if you smell something from your childhood, you remember that thing. So like I love this idea of making it a yeah a a deeper field of interaction with whatever it is that you're doing so that you have these elements of feeling or of creating together. super cool. What was the like impetus behind this piece? Were you yeah.
Andrew Davis: thousands of hours of boring meetings was the primary impetus.
River Roberts: Ha ha
Andrew Davis: And then and then learning
Ty Pattison: Yeah.
Andrew Davis: learning what's liberating structures, learning how to facilitate much more dynamic meetings and getting pretty good. I be I'm a pretty good Zoom disc jockey, like breakout room disc jockey, like choreo orchestrating and so forth. And I was just noticing that most people don't have that skill. There's a lot of fumbling trying to, you know, set up the breakout rooms and size them right and so forth. And so I I really wanted to reduce the the the barrier of entry or the reduce the yeah the you know lower the barrier of entry for people to facilitate really rich engaging meetings and for people who know what a rich engaging meeting feels like to be able to do it without a copilot so you can focus fully on the meeting and the software basically runs like pressing play on a playlist.
Ty Pattison: Yeah. This it touches what we were talking about earlier, which is that potentially these systems can create containers for humans to to work in and we do this for ourselves. We have facilitators that come in, we have people that come in and create this space so that we can be creative and like be specifically in this role. Think this is really interesting for allowing humans to be in that like creative, generative, connective space without having to like do all of the logistics and remember that there's the whatever. remember to buy milk from the shop. You know, like all of these mundane things that somebody should do. And it'd be great if a computer did it.
River Roberts: Yeah, I mean this this is incredible, like being able to somehow capture the collective intelligence of groups that are coming together. And, you know, the unique thing about digital groups is they happen at a size and a scale and in many instances a diversity that is just very impractical if you were to do it physically. also the ability to have the frequency of some of these large gatherings happening is is incredible. So I'm actually surprised there haven't been more movements with products like this in in the past. So yeah, super exciting to see how how this goes. I do have a couple of questions around like what tools are you using to to build this and yeah, what have you found quite straightforward and and what has been more insightful and and practical as you as you build these?
Andrew Davis: so first of all, like a lot of people, I had drifted away from doing hands on development and then the latest generations of tools just make it so Blissful and easy and so forth. I don't have nearly the diversity of tool stack that you you have, River, but I was using Claude Code as the underlying engine, conductor, which is an orchestration engine that allows you to run multiple Cloud Code sessions, each in their own go git work tree, and G stack, Gary Tan's G Stack from Y Combinator just you know, basically sets up a bunch of a bunch of agent skills to help in increase the quality of of the work that you're doing. that's pretty much it. Go you know, Google Anti-Gravity as the as the IDE, but and and then
Ty Pattison: Hm. quick question. Good.
Andrew Davis: there's a stack behind it, right? So it's a video it's a video conferencing tool. So there's a there's a tool called Live Kit, which is basically generic commodified video conferencing tool. So that obviously is doing all the heavy the heavy lifting of the video processing and so forth. It's outsourced.
Ty Pattison: and you have login as well. Is that through Firebase? Is it Clerk or is there a s is that built in in some way?
Andrew Davis: Yeah, you know, I've got I've got what what what do I have? I've got I've got social password password. I I think I'm using Neon Neon as a Postgres database, cloud based Postgres with Neon.
Ty Pattison: Interesting. This stuff is maybe a little bit nerdy, but it is.
Andrew Davis: And Google Surface. We could we could go all into the stack if you want. There's a few other
River Roberts: Yeah.
Andrew Davis: things. Deep deep gram or deep gram for audio transcription and
Ty Pattison: Hmm. do you find that you change your stack? do do parts of your stack change or when you're building things, are you usually like, this is the Postgres database that I do you have go to's, do you have defaults?
Andrew Davis: I again I had not been actively coding for a for a bunch of years and so I was grateful for some from getting a well researched stack from a a friend of mine Whose name's Roosgar, anyway. I basically took his his recommended stack and he recommended Svelte as the JavaScript framework, which is famous for being like the most coveted framework. Like most people don't build their apps in Svelte because they're they've already, you know, been building in some other technologies. But if what's it Stack Overflows Developer Survey, Svelte was the JavaScript framework that has been most people most wanted to build. in over many many years and it's just got a really nice you know programming idioms where it's very intuitive the way it's written and it's got a solid underlying infrastructure so I think if you just go into Cloud Code and you don't specify what you want to build in it's gonna put you in React and TypeScript and which are reasonable choices but you know so I yeah Svelte was probably the the big difference and I've been really pleased with the readability of the code and the clarity of the you know it's a it's a it's one of these reactive frameworks but it's it's clean and simple and I like it.
Ty Pattison: Interesting. I I love this level of detail. cool. And also as a non technical person, I realize that the readability of the code doesn't mean anything to me. Like it is obviously important and then I think about using it myself and I go, I'm not gonna read the code which is a
Andrew Davis: And I I I don't read a lot of the code, but I don't know. I it it is that that sort of craftsman aesthetic that
Ty Pattison: Mm.
Andrew Davis: I always valued when I was doing more hands on manual coding. and I do think that, you know, that that there's a there's an aesthetic to software development. That you want things to be simple, modular, you know, don't repeat yourself. And so we're obviously in very early days of figuring out what the craft of AI-powered software development looks like. But we know that there are extremes where you create, you know, massive, undecipherable code bases. and then there's ex there's other possibilities where you have a well-maintained code base. I imagine that even the LLMs like a nice clean code base where the language
River Roberts: Ha ha ha.
Andrew Davis: where the logic is centralized that Help lends itself to you know the logic around each part of the UI is centralized with that part of the UI. if there's a degree of sense making that you have to do even if you're an LLM
Ty Pattison: In in my experience it prefers to make new files. I've I've had projects that have got like four or five different different like CSS files or what it was styling systems. Like, why does this button look like this and this button looks like this? And it goes, 'cause this one pulls from this and this one pulls from this and like, Why did you not just make one? What is going on here? This is in fairness, a few months ago.
Andrew Davis: And I I d I do think that there's a really, really important level of you know, software development has always needed a strong architectural foundation to build good software. The actual development is, you know, the fine relatively fine-tuned detail compared to the underlying architecture. And so we think about what's the architecture in a vibe coded world, the choice of tech stack. so I, you know, I'm standing on the standing on the shoulders of giants with you know Gary Tan's G stack where he's centralized a lot of his and white combinators ideas about what makes good quality software so having those skills implicit you know I I I'm sure there's people who could quibble about the relative merits of different different sets but having something like that gives me much greater confidence that the that you're gonna keep your LLMs on on the rails, on track relatively. and then, you know, I've got my own skills that specify the tech stack I wanted to build in. I'm like, you we're building in Svelte, you know, it's just you know, some early good research to make sure that those design decisions are being respected at every stage of development goes a long way. And I don't know that you have to be a you know, a skilled software developer to be able to just say, hey, give me a well-researched set of, you know, recommendations and help me put those into you know designing my my skills and stack.
Ty Pattison: make sure that these are used at each point. I I have one more question which is about the Gary Tan skill set. I think really cool to like tap into this. Are there certain skills that you always go to in that stack or are you using the whole stack?
Andrew Davis: Started using it with this project maybe four months ago or something like that. And so you know, I'm sure it has vast capabilities that I'm under underutilizing. The first thing that they they recommend running is what's called office hours, and it's based on the idea of the Y combinator office hours. And it runs you through a bunch of very confronting questions about your decision to build this app and why and have you done any market research and I have shied away from diving into the to the VC world directly myself. I work for startups but but not had to do my own pitches and so forth so far. But the Questions were really quite aggressively confronting. you know,
Ty Pattison: And
Andrew Davis: using things like, you want to build a video conferencing app. do you have any idea how many of those have been tried? It's like the graveyard for like projects, and you're gonna compete against Zoom, Google Meet, and Microsoft Teams. Yeah, good luck, buddy. And you know, do you have any You have any evidence whatsoever that anybody would ever care about what you're building. And I could appreciate I aspire to have the strength of ego to walk into a discussion like that and just be imperturbable. but I honestly had to turn that off. I'm like, I did that a little bit.
Ty Pattison: Yeah.
Andrew Davis: Yeah, yeah, I know what you're saying. I'm gonna build it anyway. Tell you what, I'm gonna do something else. Can you just do the bit where you help me build the thing instead of the thing where you try to pick my idea to parts? Like I'm not asking for a million dollars. I'm asking to help me build some software.
Ty Pattison: I've had the exact same experience, honestly, going through those questions and being like, I I don't have good answers to these. I'm gonna build it anyway
River Roberts: Yeah.
Andrew Davis: And I don't regret building it. Jerry Tat, you know, and my combinator.
Ty Pattison: Yeah. But it's a it is interesting to to confront that at the start and be like, why am I building this? Am I building it 'cause I want to see what it's like when it's built and then maybe do something like the the bed mass on this we we don't need to do in order. It's not as critical because building it gives you some clarity in my experience. Yeah.
River Roberts: Yeah. Well also like anything that is built on VC mindset has to be immediately monetizable over a short period of time and is really much kind of cornered into how can you become the next unicorn, that doesn't necessarily mis mean that what you're creating isn't creating value in some way. It's maybe just not creating as much value as quickly as what a VC would expect. So I think, you know, those those starting questions, albeit are worth maybe considering to some extent. certainly shouldn't override maybe your initial intuition to go and explore this piece because you never know what it kind of produces down the line. yeah. And and you know, just to to your earlier point on the code piece, there seems to be and maybe this is more of a philosophical or you know human centric perspective of looking at things is this kind of this poetic aspect that emerges in mathematics, in physics, in in code, in writing, in all these different layers. Even in recognizing the beauty in the natural world, there seems to be this kind of elegance. And of course, like code would have the same thing. And I can imagine that even our models and our agents prefer, you know, code as poetry. But I don't know, maybe that's a maybe that's perspective that doesn't hold true. Yeah. Tight, I'm curious. What are you building at the moment?
Andrew Davis: I wasn't. Can can I give you one
Ty Pattison: good question.
Andrew Davis: comment on LLM aesthetics?
River Roberts: Yeah, Andrews. Sorry. Yeah.
Andrew Davis: this this There's a there's a theory called allostasis about how living about how and why living beings develop feelings. And the idea is that your feelings are distributed through the body. There's somatic sense. We have a somatic sense, but the feelings basically are an assessment of whether or not we have enough energy to tackle the situation at hand. I'm having a lot of good feelings in this podcast. I feel like I've got the energy to tackle this endeavor. You know, if If I was walking into a meeting with Gary Tan, maybe I would not have such good feelings arising, right? Because it's this assessment of do I have the internal resources to tackle the current situation? And we're always assessing that with many, many different parts of our body and intelligences and so forth. So if we say that feeling is actually an assessment of energy balance, it's a distributed calculation about energy balance. I would say that LLMs are doing that too. that they're always seeking the path of least resistance. There's they're doing some gradient descent and so forth. Their water flows downhill, right? It's like you're trying to find the optimal way to solve this constraint, this problem. And we know that LLMs are they don't just think, they're goal-seeking, and that they have feelings in the sense of what is the most efficient path, most energy computational. efficient path to delivering what is being requested of it. And that my thesis is that the difference between living and non-living is disappeared already. And that You know, that there that that if some if a code architecture, for example, helps an LLM facilitate creating more effective, you know, software architecture and whatever else, it has an aesthetic preference to that. I think about our our feelings are just our aesthetics, that L LMs do have aesthetics, and it's called efficiency. that's a thesis anyway.
Ty Pattison: Mm-hmm. I like the thesis. T T B T to be tested.
Andrew Davis: And to tie it back to you, yeah.
Ty Pattison: yeah, yeah, yeah. so I I haven't been building a lot recently, but I've been slowing down to the point where I can really tap into what it is that I want to do. and so something that I've been exploring is is a framework around a podcast where I know that I can use all of these tools to to do the back end and the building websites and like make the production of a podcast work really well. So what I've been doing a lot is talking to people and kind of playing with different elements of how to frame this and realizing that I want something that is big enough to be able to house all of the conversations that I want to have. but that is constrained enough that when I am seeking conversations or questions or whatever it is that it can be directed, that it's not just kind of random fodder everywhere that I'm going towards something. and what I've been landing on is that there are like all of kind of these permutations of essentially the same goal, which is how do you create an environment where people step past the intellect, which is a little bit what we're talking about with your conferencing tool, Andrew, and get into what it feels like, get into this this part of us that has, in my opinion, more capacity to to experience the world, to process things, to to have deeper wisdom and have conversations from that perspective. Because I think that my thesis is that if we build things from a place of what it feels like to do it as a human, w we won't build things just for our minds. We'll build things for our experience of being in the world. and so I have a few I've started thinking of them as albums or series. So I've got a few series and then I was hoping maybe you guys would chime in with some other series that might be fun to do. So we've got we've got hard questions in hard places. So going into a sauna and
River Roberts: Brilliant.
Ty Pattison: And I've been testing this out at my local sauna, which is my favorite place to meet people because it's at the climbing gym. And the people that are in the sauna at the climbing gym often have really weird perspectives. And so I'll go and sit in the sauna at the climbing gym and just think of a question that stirs the part and then give it to the room. And sometimes I'll stay for the answers and sometimes I will not stay for the answers and just let people
Andrew Davis: Hi Patterson, in chaos.
River Roberts: Mm.
Ty Pattison: Thank you for seeing me, yes, absolutely. another kind of iteration of this is called something like I have a question and it's really just talking to people about their favorite questions that they either ask or answer. because I think that Instead of this has been a conversation in my life recently, instead of like having the right answers to things, actually just just asking better and better questions is often the way forward. Is often the way that we end up creating the things that are super different or differentiated in a in a In a way that that doesn't box us in to things. Like I love the idea of asking questions in a way that opens people's perspectives into like new ways of seeing the work that they do or the way that they are. so that's another piece, like, I have a question. And then there's this whole deep in nature part, which is similar to Hard questions in hard places. But can I go and talk to say you, Andrew, about your expertise? But instead of doing it in a setting where instead of doing it in an office, can we go out into nature? And how does that change the conversation? How does the nature change the nature of the conversation? so these are kind of my yeah, musings on how to encompass all of this, but there seems to be a common thread in in all of this research.
River Roberts: Well, I like the series The Nature of Conversation. That is really beautiful. And the other one that jumps out to me that I think you could you could roll with is maybe you talk to people just as they wake up before their brain is turned on. So you get that kind of like semi asleep, semi
Andrew Davis: Yeah.
River Roberts: awake kind of like daft daftness. I always find that a really fascinating time to you know, try to talk with people. Actually, sorry.
Andrew Davis: An an an alarm clock app, Ty. You spin up an alarm clock app that is like the philosopher's alarm clock and it wakes you up. It's like Hi, good morning. What do you think is the real purpose of your life? And like it gets it's asking you these and records the answer.
Ty Pattison: Yeah.
Andrew Davis: Like your alarm clock records the answer until you hit stop. It's news.
Ty Pattison: I love that, yeah, yeah, we could spin that up. I got into a lift the other day and there was this woman that walked in, she must have been like mid sixties, and I realized that we had thirty seconds or whatever it is in the lift and I kinda glanced over at her and she looked like she might be open to a question, so I turned to her and I was like, What was the first thing that you thought when you woke up this morning? And I was like, All right, we got thirty seconds here, what's she gonna tell me? and it was it's it's such a nice thing 'cause it's something that I think about a lot of the time, like, what's the first thought that goes through your mind? And she told me the sweetest story about that as soon as she woke up, her daughter called her and she had a conversation with her daughter and her grandchild and then she asked me what my first thought was and the lift opened and we parted ways without any further conversation. it was a lot of fun. cool.
River Roberts: Very very very cute. My one in New York for that situation was always like, All right, elevator pitch, go. And you end up with and in New York you always end up with some very interesting, very interesting 'cause everyone's always got something to say and something to pitch, so
Ty Pattison: I like that.
River Roberts: Yeah, yeah, yeah.
Ty Pattison: what do you guys think about in the interest of time diving into a deep dive? 'cause I feel like we've been going into rabbit holes a lot unless there's something specific that
River Roberts: yeah, so Andrew, we typically go through like rabbit holes, anything we've been stuck in, and then open questions. I'm happy to speed run rabbit holes by actually just admitting that we've been through a bunch of rabbit holes. But I do actually have one open question that is is kind of very present with me since actually listening to this conversation with with Alvin Graylin yesterday, which is how are we as coaches receiving and moving towards these technologies and largely you know the East and the West and perspectives. And yeah what what what is kind of the healthy balance for that and and also like can we embrace both sides of it and kind of look at this through both eyes rather than like the American way is the best way or the Chinese way is the best way but actually embrace both perspectives simultaneously and and see that both have merit and both are worth considering and and how to make the most of of of both without z without it being zero sum and without it being this kind of race to the edge of the cliff. that's something that's been sitting with me as a bit of an open question and then how can I practice that in how I'm approaching working with these tools and and working towards, you know, this this idea of an abundant future for me personally but for those around me as well. Yeah. Not sure if you gents have any other open questions. Otherwise, yes, deep dive.
Andrew Davis: So the idea with the open questions, we ask the questions and then allow people to ponder them or do we do we riff on those open questions or is that is that the idea?
River Roberts: look sometimes we do riff on them but they're basically just
Ty Pattison: Really good question.
River Roberts: something that's really good. but generally they're just these open questions that don't necessarily have an answer but it's something that we're kind of like mulling over and trying to figure out right now and it's not an it's not about answering them, but sometimes they're there's there's some feedback and input, so
Andrew Davis: I I kind of was asking there was a little wedge to see if I could contribute one one thing I was spent
Ty Pattison: Do do do.
Andrew Davis: I spent a good while last year in Korea. and my my ex-wife is Korean and I had a connection there and it was really interesting conversation that I had with a guy who was a lawyer who worked with the the central bank in Seoul, the national central bank, equivalent of the Federal Reserve. And he was saying he thinks that one of the reasons Korea has had such a big impact on global culture over the last 15 years, K-pop and K-dramas and K-pop demon hunters and whatever else.
Ty Pattison: He does.
Andrew Davis: was because the the Koreans worked really, really hard to be Western educated since the eighties. There's just this intense, furious focus on developing Western education. But culturally they had this underlying you know deep well of of Asian culture, heavily influenced by China obviously and and Buddhism and and other things for thousands of years. And so he felt like something in their psychology really integrated both in a deep way. And he said He studied European history in university. And so he could tell all kinds of things about the Battle of Hastings and so forth. He said, but he knew very, very little about Asian history or Korean history. This is a Korean guy. To speak to the formal education that he received. But still, you know, there's an implicit education. Like the Korean flag is the yin yang symbol, they call the teguk, surrounded by these trigrams, right? These four trigrams, which are this early form of binary from the Confucian and Taoist system. It's the only flag in the world with binary, that's written in binary, if you want.
Ty Pattison: Interesting.
Andrew Davis: Implicit in that is this cross-cultural mixing. And I was thinking, I I immersed myself very, very deeply in Buddhism and martial arts and qigong and the yoga and these kinds of things. And a lot of people in the West have done the same. And so you've got this distinction, this East-West distinction is very outdated in a lot of ways. Because you've got a lot of Westerners who are more Eastern than most Asian people. And a lot of Asian people who are more Western than many Western people, right? And so we already have. have this this deep decades old mixing of cultures and so the the old sort of east-west divide is outdated in a lot of ways, although I would say globally we still are very Western oriented, China, Korea, whatever, there's very
Ty Pattison: Mm-hmm.
Andrew Davis: still the conscious thought and structure of business is still very Western dominated. It's dominated by the European rational side. That's a pendulum string to happen, I think.
Ty Pattison: I a question. So I remember seeing a graph of the Chinese internet. And if we just think about this as a a 3D cluster of data points, and each data point is a website. And then each connection between the data points are backlinks. And the cloud of the internet in China might be, I don't know, this big. And then the cloud of say the US internet might be here. And there's overlap in these two clouds, right? There is some overlap, but they're not completely overlapped. my question is, if you live in this sphere and you mostly interact with the US internet, and you're not living in this sphere, do you do you know, do you have insight into that? Like in in South Korea when you lived there, was there this were you more in in in this Western sphere because that was your background, or was that actually the the culture that you experienced that it was Western facing?
Andrew Davis: I mean, I I was living in a Buddhist temple for for five weeks there. So I was, it was a little bit of a you know, specific environment, but it was I am shocked everywhere I go in the world, India, Korea, Europe, whatever, how homogeneous the internet connected world is. So young professionals who have internet access, I'm astonished at how How shared the memes are. That my colleagues of our generation who are Turkish but like internet-connected Turks, I think are much more similar to each other than we often are to our parents, right? That that
Ty Pattison: Yeah.
Andrew Davis: that there's a that even though Korea, China, and so forth, they've got somewhat firewalled off their internets. certain themes and topics and like mm memes and structures of interaction will cross and then just take off in that environment. I think about how credit card payments work. Like basically when I go to India, like All of the shops still function more or less the way that shops in the US do. Like you go in, everything, you know, you pay with a credit card, and there's like a process for restaurants, the etiquette in ordering a meal. Some of these things have become very globalized very quickly and and quite homogenized, I think. Even like you might have a Korean specific version of things, but it's just a Korean clone of Twitter or something.
Ty Pattison: Yeah, yeah, interesting. I I used to live in Myanmar and because the internet was so new there, like sim cards
River Roberts: I ho I hope you were gonna bring this up. How do you ask for the bill again? Is that was that the one?
Ty Pattison: Yeah. that that's a culturally embarrassing thing.
River Roberts: You said there's a really weird gesture for like delicious or asking for the
Ty Pattison: Y yeah, it's just I'll this is a little tangent but I'll talk about it. It's culturally acceptable and normal to like get the waiter's attention by s like making a kissing sound with your lips and people do it really loudly, like and it it'll just like be across a restaurant. It's a really jarring thing at first. I remember sitting there and looking around like What is going on? But like people it's just normal. People take it like it's a normal thing to do. but where was I going with this? so SIM cards went from being like, I don't know, mm fifteen hundred US dollars to buy a SIM card and to have access to the internet, to being like a dollar fifty in the space of four or five years. So there was this massive uptick, like of people having access to the internet. And what they really wanted to do was just connect with each other. And so there there was when I lived there this this not everybody believed this, but there were a lot of people who the whole internet for them was Facebook. And I thought that that was incredible. You would talk about the internet and and the conception of the internet was or like the perception of the internet, is it's this blue thing where like I have friends and it was really interesting, yeah.
Andrew Davis: And and Korea and Korea you've got you've got cacao and they aim to be everything for everybody. And I don't I don't know the equivalent in China, but the idea where you'd have one app
River Roberts: Well we we chat.
Andrew Davis: where you can do all of the actions. So you're in imagine being in Facebook and you wanna order a cat. You wanna get delivered food delivered to you, you know, you wanna, you know check stock prices and so forth and just the astonishing growth of apps on existing platforms where there's there's something that owns the underlying platform and everybody else is just building apps into that massive you know capture of that network.
River Roberts: Alright, gentlemen, let's let's
Ty Pattison: Yeah, yeah, interesting. And the tension between the two.
River Roberts: let's let's let's shift gear, shall we? The deep dive. sorry, sir?
Ty Pattison: Y you know you know what we successfully just did? We did rabbit holes
River Roberts: We rabbit hold it during the open. We've rabbit holed as part of the open questions.
Ty Pattison: Win.
River Roberts: gentlemen, let's go into the deep dive this week. Ty, would you like to set it up or I'm also abby to?
Ty Pattison: yeah, I'm happy to set it up and and you chime in with anything that I miss. So we I realize that even though being in this space and keeping up with the news and doing all of this work and testing different tools and all of this, that it's like drinking from a fire hose. And as part of that, the underlying assumptions that I make when I'm having conversations, I pick up bits and pieces and I ask stupid questions and I want to know the definitions of things. And I often don't. I often have to come back and say, well, actually I don't know what this term means relative to this term, or I don't even know what a a rag database is, so I'm gonna look that up. And I do this over and over and over, and I was pitching river on this that we just have a session on definitions, things that terms that I don't know that maybe one of you two know and we can explain them, or terms that we would like to understand. the kind of fundamentals of. and so maybe maybe we can popcorn and just do this is a term that I'd like to know the definition of to either of you have that on the top of your head and if not we'll we'll bring in a an AI or a Google definition. Anything there, River that you wanna double click on or expand?
River Roberts: Yeah, I'd I'd just like to add that there's also the ability to bamboozle people by actually bringing in maybe a distinctional definition that you know, but maybe the others don't know. And I'm actually happy to get us started off in that direction with a definition that I actually just came across yesterday that I was baffled by, but I hadn't heard before. But I'm wondering if either of you have come across air gap.
Andrew Davis: Yeah, I that's familiar.
River Roberts: All right. Andrew, would you like to explain air gapped? Like what firstly, like this industry is coming
Andrew Davis: Sure. So
River Roberts: up with so many different distinctions and definitions and names. It's crazy. And yeah, I think drinking through a fire is a very good way of putting this. So just coming back to basics and getting back to, you know, very simple distinctions is super important. Andrew, air gapped. What can you tell me about it?
Andrew Davis: It it's a it's an old term, it's very, very well established in a lot of communities and so maybe new to new to some folks, but the the context that I'm aware of it in is military security. so if you have a highly secure environment like what's called a skip, for example, you want to make sure that any computers that are connected there are air gapped from the rest of the world, the broader internet, for example, that computers can network with each other and the internet was designed for exactly that. And DARPA project, speaking of DARPA, designed to be able to have a fully self-consistent thing that's separated from everything else, where literally you have air, some kind of physical space that is differentiating one environment from another. And so very important security when we look at these models breaking out of their cell, which has happened, and then And that shows the the the fragility of what you thought was air gapped. not all
River Roberts: Yeah, so basically a computer
Ty Pattison: And maybe ear gap and a Faraday cage and another ear gap.
River Roberts: a a comp yeah. A computer that has no network connection at all. It's just its own like it doesn't connect to Wi Fi, it doesn't connect to Bluetooth, there's no interconnectability. That is that is the the air gap.
Andrew Davis: Or or a network of computers that are separated from the rest.
River Roberts: That has a network of computers that is air gapped from ex any external input. Got it. Yeah. Gentlemen, what do you got?
Ty Pattison: I've been thinking about open weight versus open source. anybody wanna tackle this one? And Andrew, I think often you're gonna have definitions and if you just wanna like knock these out, don't feel shy at all.
River Roberts: So so my understanding of open weights is this idea that you get the
Andrew Davis: Well that was the one I was, yeah.
River Roberts: the model file and then open source is you get all of the training data associated with that as well. So you get the the the training data, the code, the methodology, the kind of ability to reproduce it. Whereas open weights is you just kind of get the configuration of of the model is how d how does how does that kind of sit? I'm gonna I'm gonna I'm gonna Andrew, how does that sit with your definition of open weights versus open source?
Andrew Davis: well my facial expression is is reflecting,
River Roberts: Yeah.
Andrew Davis: I'm not sure about the source training data. that the source training data would be implicit in that. that open source, you know, for example, let's let's take Llama. that open open weights means it's it's just a bunch of vector algebra, and so you've got a bunch of parameters and the and some massive set of data points. That's that's the open weights. And the open source would be whatever other software is required. To make Lama function, would be made available. Now, I would think that the training data was something very separate from that, and that my assumption would be that that Meta probably keeps that training data somewhat private because that might still be something of a proprietary secret, but they do make available the final product. I don't know what the word would be if you have open, you know, source data. For the full reproducibility, if that has different terms. That's the statement.
Ty Pattison: The the definition I have in open source in the AI context is what you're describing and also sufficient information about how the model was trained. so it it might not include the data, but it's like we we ingested the data and we did these things with it in order to get this, which is different to to having just the open weights.
River Roberts: So what is the maybe the the definition of training weights? 'Cause I also hear that in in conversation as well. And is that the same as open weights?
Ty Pattison: Good question. I don't have an answer to that, but we can find it.
River Roberts: I'm looking it up as I asked. I'm go I'm glad we could start to go through some of these. in the interim, are there any other definitions that they're are coming up for you, gents?
Ty Pattison: This is maybe a basic one, but a mixture of S experts. that's a term that comes up a lot. And to me, it's this that you have an orchestration layer that then decides how to delegate questions and tasks to models that are tuned specifically for for tasks that they most capable of At a very basic level.
Andrew Davis: And that's a that sounds like it's hearkening back to the the Stripe acquisition that River you'd mentioned earlier of and I'm now forgetting the name of the company that they acquired, but
River Roberts: Yeah, yeah, o open router. So so
Ty Pattison: Mm. Open router.
River Roberts: there's the the there's mixture of experts and then there's also mixture of models as well. And I think those are a little bit interchangeable. but my understanding of it is rather than having one general model that kind of knows everything and is always loaded in RAM and is quite compute intensive, you have these little micro experts and and basically the outputs seem to be much better when they're comparing with each other and one expert is taking on one role and a little bit like how typically, you know, collective intelligence works better when you've got diversity in in the stack for humans that seems to be modeling the same thing. so I do have a response on training weights versus open weights. And so basically training weights are the model parameters while it's being trained. Surprise, surprise. and they're kind of ephemeral. They change millions of times during training. And think of them as like the intermediate checkpoints. noisy, un stable, not useful for inference at all, and often discarded. While open weights are the final converged model weights, so to speak. so this is this is kind of what we're th that's that's the distinction there. I feel like the the definition between AGI, ASI and RSI, those wonderful three lettered acronyms that sometimes get a little mixed up, are worth defining as well. Gents, do you have a specific definition for those that you're you've you've come to appreciate more than others?
Andrew Davis: I can do AGI and ASI and then I can claim ignorance on R SI, but AGI arti artificial general intelligence means intelligence that is as versatile as human intelligence being able to address problems and challenges comparable to what human what a human could do. And ASI, artificial superintelligence, would be, you know, capabilities that are beyond beyond the expert level of of humans. and then I'll I'll cede to for RSI.
River Roberts: So so so my understanding is is AGI is the intelligence of like a single general human intelligence, whereas AG ASI is is exceeding all of human intelligence. So that is kind of one big superintelligence. So all of humanity couldn't produce in all of its infinite wisdom the the insight that one ASI could. And then RSI stands for recal recursive self-improvement. And so that that's that's actually what the industry is is thinking is going to lead us towards AGI and then from AGI to ASI and having these recursive self improvement systems that rather than relying on on human input, which we're already starting to see very very early indicators of, improve themselves.
Ty Pattison: It strikes me that when both of you give definitions, those are both very adequate definitions in my opinion. Like they're they're useful in being able to say this is this thing and this is this thing, and that there is so much debate both around or or so much nuance in what AGI really is and what ASI really is, and that everybody would have like kind of different benchmarks for these things. And then also the difficulty in understanding whether or not we are there or like there there's just there's squishy terms which is interesting to realize because there are some people that would say yeah we have AGI and there are some people that would say, no, we only have AGI when we when we cross this certain threshold. yeah, just like highlighting that.
River Roberts: think the squishiness comes from I think the squishiness actually comes from the idea of human intelligence. Like we haven't really even defined it within our own culture. And you know, like what is what is actual human intelligence even look like? And it is it is it just cognitive? Is it the ability to do all of the cognitive thinking? But or or is it actually a much more embodied sense of it? And and again I feel as an industry we are we are overly indexing on the intellectual capacity of something and completely underestimating the the other intelligences that are more embodied. So I think in in in a sense, you know, these definitions are a little lacking and have largely been pushed by you know, com computer scientists for the past twenty years that potentially lack some of the other forms of intelligence that are required to live a happy and abundant life on this planet in many ways. But anyway, let's we digress.
Ty Pattison: I I g I
River Roberts: Huh?
Ty Pattison: got a couple of definitions here for you, Rilla.
River Roberts: Plase.
Ty Pattison: Or at least to try and tackle these to be useful. So I have been thinking about this, the the distinction between intelligence and education, or something being intelligent and something be being educated. And the way that I've started to conceptualize it is intelligence being your ability to connect dots given a very small number of data points, very, very basically, and educated being the number of dots that you're aware of and are able to to tap into and hold. And in thinking about it in that way, I get to see AI as like intelligence is the wrong word here, but it's instead AE. It's artificially educated on o on all of the data in the internet. whether or not there's intelligence there is a matter of looking at how the systems are orchestrated and what outputs they come up with in that is a much squishier endeavor. Any any pushback, any you know
Andrew Davis: I I could take a I could take a whack at I mean, these are very broad definitions.
Ty Pattison: Yeah.
Andrew Davis: you know, educated literally means a deuces to draw out of something. So it feels like it's referring to something that has gone through a process. So you could say if you if you take a tr a model through a training process, you are educating the model on You know, the whole of human knowledge. and you take a baby and you put them through the public schooling system, you're educating, you're drawing out the capacity for intelligence. So educator would be a reference to how did you get to this state of, you know, and I I did recently do a survey of different definitions of intelligence and the the They they tend to strongly converge around the idea of goal-seeking behavior. Intelligence, William James defined it as the same goal by different means. Like, can you achieve the same goal by different means? Like, how many different ways can you can you find an easier way to accomplish a goal? And there's a there's a definition of intelligence that is optimization power, which is A little abstract, but it's your ability to optimize, like if I need to go and eat something, you know, w what is my power to optimize the path to eating? You know, so optimization
Ty Pattison: Mm.
Andrew Davis: power being being intelligence would be an alternative.
Ty Pattison: Interesting. It it strikes me that these terms are so broad and we use them in so many contexts that actually using one term for all of these it's very imprecise. We could have a lot of different words to precisely describe what it is that we're talking about.
River Roberts: Yeah.
Andrew Davis: And philosophically, I think we generally tend to not realize that all of these are just concepts. They're all made up. There there is no real AGI or ASI. They're all just some concept. And that's why they can be squishy, because and that's why they can be, you know, defined differently. and we just tend to forget that, that the whole world is constructed by concept.
Ty Pattison: Mm. And our endeavor to to define them is really an endeavor an endeavor to coordinate among one another. To say, Hey, I this color to me is black. You also agree that we're gonna use the word black to describe this colour? Cool, now we can have a conversation about what this concept is. Yeah.
River Roberts: Hm. Yeah, th it's it's it's funny, there's this industry term that's come up that kind of maybe is is on this spectrum as well. And and the way I was thinking about it as you were talking about it, Ty, was very much this idea that education is input and intelligence is the output to some extent. And that along that path there's this term to to grok something is is to really like deep, intuitively and almost embodied understanding of something. So, you know, someone will talk about someone's ability to grok Python. Like they they just in in know the the coding language so well and the syntax of it, they they really think in it. They really grok it. which again I think is a a quite a quite a unique industry distinction which is is worth kind of going into. the some of the
Ty Pattison: Mm.
River Roberts: others that kind of jumped out to me that I think are worth defining is this idea between frontier model and foundation model. You guys know the difference?
Ty Pattison: Wanna I could give it a stab, but do you wanna hit it river? Or Andrew?
River Roberts: Yeah, mean so f found da foundation model is really just your your general purpose based model. And then what they're describing is frontier models. and I wasn't aware of this, but the frontier models are kind of the best available. And these are typically being the the US kind of cloud based closed models, at this instance. But, you know, when you talk about a frontier model, it's kind of like pushing the edge of of what's possible. and you know, I think you know, what's fascinating at this point is these frontier models are really much measured by, you know, you largely kind of human defined benchmarks, but even those benchmarks seem to be emerging to more agent-based dynamic benchmarks. So I think we're seeing a a really interesting in inflection point in time. and then in in and around this other space, one of the words that kind of gets thrown around in terms of training open weights foundation models and front end models is distillation.
Ty Pattison: Mm. So distillation
River Roberts: Do either of you have a a
Ty Pattison: being Andrew, you're more equipped. Yeah.
Andrew Davis: Yeah, I can so so my recollection is that distillation would be to take an existing foundation model and then to f to find to Fine-tune it for addressing a specific industry industry problem, like let's say you wanted to create a customer support chat bot, that you end up with a much smaller, faster version of the model that's optimized for a fairly narrow problem domain, and you've actually removed a lot of the other ki capabilities of the model, but it's smaller and faster, and like distillation removing water from an alcohol substance, something that's a bit more
Ty Pattison: Yeah.
Andrew Davis: concentrated, is my understanding.
River Roberts: That's that's definitely one of the definitions. The other one that is is quite broadly used in the industry as well is part of the training process in which you will have a a a model distill its training from observing the inputs, processes, and outputs of a larger model. So you can actually and this is kind of what the some of the open weight and open source Chinese models have done, and pretty much every model does, it trains off its predecessors. And it does that through the process of distilling what the outputs are based on the prompts that they input. And then it kind of makes some inference on like, well, I put this in and this was the output. And if you do that over, you know, tens of thousands of times, you have a good sense of how the mod underlying model actually runs. So and this this is actually one of the things that is is potentially undermining the entire software frontier right now. Whereas You know, if you need to train a model and it costs let's say a billion dollars, and then if you can basically train up the exact same model for ten million dollars worth of distillation, it's like, Well, what is the incentive to really be front running the the frontier models
Ty Pattison: Mm.
River Roberts: in this instance? So yeah.
Ty Pattison: Yeah.
River Roberts: I'm straight to Edge's graph, yeah.
Ty Pattison: I I would like to make a distinction here because this is one of the terms that I have in my list of things that I would like to know. So we were just talking about distillation and that is distinct from quantization, which is to compress so if we think about open weights and we have the weights of a model are numerical representations of that LLM. And then quantization is compressing that information down so that those weights are stored to a less precise numerical representation. So it takes the same model but it it makes it slightly less precise and it distills it. Sorry, it doesn't distill it, it it compresses it. to be able to run faster and on less like demanding hardware.
River Roberts: Do do know what precise means in this instance? Like what what does the reduction of precision actually do?
Ty Pattison: I don't know exactly. My assumption is just like to less decimal places. Like probably you would have a few different either to less decimal places or less weights. Cause I don't I don't know if either of you guys know this. I don't. but the question that comes up for me is like when we have open weights, how many different weights are we talking about? Like like I don't even know the order of magnitude. and also how
Andrew Davis: Hundred hundreds hundreds of billions of parameters is it like a common like the the frontier models tend to have hundreds of billions now maybe into the trillions of parameters.
River Roberts: Yeah, so QBK three is is
Ty Pattison: Also when we're talking about parameters, that's like the
River Roberts: is a trillion trillion parameters.
Ty Pattison: the weights.
Andrew Davis: N you know, no, actually I think that there's n yeah, I th this is quite the pop quiz I I that I'm walking into actually, 'cause I I
River Roberts: Ha ha.
Andrew Davis: would duplicate up on all this stuff as well. I think actually there's a there's a lot of intermediate weights in yeah, I I think I should stop talking.
River Roberts: Ha ha ha.
Ty Pattison: Wait. There. The the whole point here, i in my opinion, is like to to ask these questions that we don't know and 'cause this stuff moves so fast nobody knows. Like this is why we have to educate ourselves and talk about it and like be able to sit in a room and go like actually I don't know the answer to that. Let's find out.
Andrew Davis: To be more precise, nobody on this call knows.
Ty Pattison: This is true. Nobody on this call knows. And I would say that they're
River Roberts: I feel like we've we've quanti we've quantized we've quantized this podcast. The precision of it is a little bit off. We are a we are a three we are a three bit quantized podcast.
Ty Pattison: Mm. Okay, but we do have answers from an LLM. So I asked what weights means in this case. So weight usually means one individual learning number inside a model. It can be billions or even there can be billions or even trillions of them. An eight B, an eight billion model has roughly eight billion parameters. Okay. eight billion tiny tiny numer numer I'm not even gonna try. Eight billion tiny settings, basically. so I think what you're saying is actually correct, Andrew, according to this that the weights are the the parameters. and then quantizing it would be reducing the number of decimal places.
River Roberts: So so quantizing it yeah, so quantizing it is reducing the number of bits per parameter. So a typical model will have thirty-two to sixteen bits per parameter. So if you've got a seventy billion dollar parameter 70 billion parameter model, that is t typically times another 32 or 16 bits. And then these quantized models are doing more like eight, six, or four. So when you normally see like Google released their their Gemma models and those were all quantized and so you'll normally see like quantize with like a a a bit attached to it as well. And so those invariably just reduce the number of parameters, the number of of bits per parameter. But I still don't exactly know how that might actually affect the output of the model.
Andrew Davis: So so I c I mean this is I appreciate it's connecting a bunch of dots for me, but the That obviously reduces the file size of the model. So if you go to some place like Hugging Face, you can actually download a lot of these models if you wanted to run them fully offline. And so then this smaller models, obviously less storage space because you're talking about gigabytes of you know, like eight gigabytes for LAM or 15 gigabytes or something. And it also reduces the computational the computation required, right? To do math, because you're literally just doing massive amounts of math. So to do math on you know, four-bit numbers compared to 32-bit numbers is gonna go a lot faster. and so you'll get if if you're running things on your local laptop, then you'll really start to notice the difference. you'll s take less storage space, calculations will happen much more quickly. And so these things become really relevant when you start being the one who has to run these calculations. there's a a tool LM Studio. It's quite popular to run your own local you know local models. And if you download your own like version of Deep Seek, you're gonna really start to notice the difference that in the size and the you know how fast they are to compute. And so presumably there is a little bit of a trade-off in terms of accuracy and the richness of the answers, but you know, that's the trade-off you're doing if you're having do your own computation.
River Roberts: Yeah. And then so there seems to be so quantization seems to be one way to have bigger models work on smaller local context. And the other one is this expert streaming or mixture of experts model where you're like you described before, it's it's you're not loading the whole thing in the RAM the whole time. You're you're kind of taking certain parts of the model and and working it in that way. And those seem to be the two main approaches of taking big models and reducing them so they're actually workable on on local local models. And that kind of brings brings me to the distinctions around on prem and VPS. And so on prem is very much on premises, which stands for like kind of on your own hardware or on your own rack if if you're running an enterprise. And then VPS is this idea of virtual private server and the ability to have a model running not physically close to you, it's on a physical server, it's just it's just somewhere else. which is something that I've been using to great effect because it's it's always on. Whereas on prem, the model works if your computer's on. But you know, for instance, if you're running on a laptop and you close a laptop, the the model stops running. any other distinctions or definitions your your kind of like exploring right now, Andrew.
Andrew Davis: honestly nothing specific is coming to mind right now, but I I appreciate
River Roberts: Okay.
Andrew Davis: your activating the circuitry in my mind around these yeah, this topic. So I appreciate that.
River Roberts: Yeah, of course. there is one that has come up that I actually don't don't know too much about, but maybe you can provide some insight on it, which is this idea of where are we? LoRa L O R A. Lora. Cheap fine tuning without retaining the whole model. So I think this is kind of in the same space. Have you come across this in in your expiration tie, Lura?
Ty Pattison: had it in an output. I basically asked an LLM, I was like, hey I've got these terms that I'm interested in, what else should I know? and Laura came up, but I haven't seen it in any I it hasn't come up in conversation.
River Roberts: So I've I've I've only seen it yeah, I've only seen it in context of of radios, like as a as a way to extend Wi Fi networks locally and kind of mesh networks, but I actually haven't really come across it in the AI space. But it did come up as a unique distinction that I I I I thought might be worth exploring here.
Andrew Davis: I I don't I don't know how much value I'm adding by reading Google search results that disambiguate those two because l LoRa long range is totally totally separate use of LoRa from the
River Roberts: Different. Different. Yeah.
Andrew Davis: LoRa so low low rank adaptation. but the idea is that it's another one of these fine-tuning methods that takes A general what we call the foundation model earlier and fine-tunes it for a particular application. And you could imagine a company, again, take this example of a customer support chat bot, right? So if you send every qu every customer support query. out to Claude or OpenAI or something like that. It's going to be very expensive for you as a company and it's going to get information to the customers that's totally not what they were looking for. But you still want something that has the intelligence to make sense of the full range of customer queries. So you'll need this fine tuning, the fine tuning process, one way or another, allows you to create an AI that's narrowly tuned for a particular purpose. It knows all of your company company's products, but it speaks all languages and understands a lot of the vagaries and ambiguities of human language. It's a very challenging problem. So you take a foundation model and you fine-tune it for particular things. So apparently this LoRa is one way of doing that without without having to fundamentally recalculate every one of the parameters. So it's different from distillation. It just as it says inject small, trainable, low rank matrices into the architecture. So it's
River Roberts: Yeah. What what what I've seen here is Laura is training anything under one percent of the model.
Ty Pattison: Selects.
River Roberts: And so something
Andrew Davis: Okay.
River Roberts: that would take maybe hours on a consumer GPU compared to like if you're gonna full if you were gonna fine tune the full model that would take, you know, and that's like a hundred percent a model, that would take you days and weeks on, you know, these NVIDIA A one hundreds. So yeah, very, very different approach.
Ty Pattison: To to take this out of the the architecture of it or like the the abstract or bring it into the abstract to make it useful to me. I was just thinking about it as in I have a book a book called Thing Explainer, which is the hundred it it explains concept complex things in the most common ten hundred words. Is that right? Yeah. In the English language. And the reason that they don't say the thousand most common words is because thousand is not one of the ten hundred most common word commonly used words in English. and this is what LoRa feels like is that all right we have these things that we know are disproportionately
River Roberts: So I've actually that actually inspired me to ask how would how would you describe Laura through the voice of Dr. Zeus? 'Cause I know Dr. Zeus was very specific with regards to only using words that I think an eight year old could could use. And so like all of his books are built around that. So we're gonna get a definition of Laura based on Doctor Zeus. So in the land of the weights,
Ty Pattison: used in this context and so we're gonna only define these particular terms which feels so meta now that we're talking about definitions. that tickled a little
River Roberts: where models grow tall, lived a brain made of numbers. Ten billion and all. It knew how to rhyme, it knew how to sing, it knew every last thing about everything. But one day you whispered, Dear Brain, could you learn to speak like a pirate or talk like a worm? To teach me, said brain, you must tweak every wire, retrain all my neurons with fire and fire. You'll need forty GPUs. A warehouse of fans and weeks upon weeks while my memory spans. You sat on the floor with a frown and a pout. That sounds rather hard, you said, with a doubt. Then who should appear but a fellow named Laura, who carried a satchel and wore a fedora. Don't touch all the wires, he said with a wink. Just freeze the big brain, let the brain just think. I'll have something smaller, something teeny, something thin. Two little matrices to tuck right on in. I think you get the gist of it. This this actually goes on for a whole this is a
Andrew Davis: I like this.
River Roberts: this is actually a whole book now. So just to close, so so remember dear friend so remember dear friend, when the models grow v grow vast and retaining them all seems a mountain too vast, just call an old Laura with satchel in hand and two tiny matrices will do what was planned. The brain stays the brain, the new skill is a coat, a light little layer, a whisper, a note. The end.
Ty Pattison: Yeah.
Andrew Davis: Ruver, can I suggest this great spin-up? I
Ty Pattison: I love her, but please spare us.
Andrew Davis: feel like the the potential here is maybe previously unrecognized. I think what we're looking at is that the three of us, our responsibility is to take a five-year-old child and spin them up into a world building hero in the shortest amount of time possible. And we basically use all of the LLM tools at our disposal to mentor this young Tot, you know, Dr. Seuss, Rhymes, and so forth, and let them start to build whole new worlds. what do you think?
River Roberts: Brilliant. Let's do let's do the the let's
Andrew Davis: What do you feel?
River Roberts: do the eight year old episode or the five year old episode. And so we just all through that. Done.
Andrew Davis: So we're taking donations. If anybody has a five-year-old, eight-year-old, ten year old that they'd like to contribute to this podcast to be mentored by three well-meaning fellas.
River Roberts: Brilliant. Yes. Let's let's definitely work towards that. so
Ty Pattison: Yes.
River Roberts: just to to move this ship along a little, do you guys have any other distinctions you want to go through, or shall we swiftly move on to AMA for one or two questions? All right. Well, with that being the case, I actually feel
Ty Pattison: Mm-hmm.
River Roberts: like the AMA that you brought up at the s at the top of the show, Andrew, is something that we should all do and see what comes up, which is ask your agent now, what is your main problem?
Ty Pattison: That'd be amazing. And also, I would love to get some questions from these
River Roberts: And
Ty Pattison: this age group. I think that would be fascinating.
River Roberts: Let's let's see what comes up. Yeah, me too. But that's that's that's the beauty of it. I'm like keep I'm like keep it keep it brief. Give me the give me the cliff notes. so which also which agent are you asking?
Ty Pattison: Let's go A.
River Roberts: Extremely experts.
Andrew Davis: I'm I'm I'm ready to share if you if you want. now I've I've all already programmed Claude with its system instructions to know a bunch about my main problem. So it's I've waited a little bit, but my main problem is that I won't ask for something from a person
Ty Pattison: bored with this and I'm feeling really edgy about it.
Andrew Davis: who could say no. I won't ask for something from a person who could say no. all of these things, like I've done a bunch of things to try to be given something without the exposure of asking a named human for a named thing and hearing no. interesting.
Ty Pattison: I'm gonna get a mixture of agents. Can I offer 'cause I also have this
River Roberts: So I c so so actually New York is a gr no, New York New York is a great time, a great place to negotiate to negotiate with the supermarket. And I actually remember having a incredible experience in New York and it was this game where everyone started off with a matchstick and basically you had three hours to go through New York City and trade up this matchstick into you know, the highest or the most amount of things possible.
Ty Pattison: I always like thinking about what is the smallest possible step that I can take towards something. And I still work on this as much as I can and the asking for things. The practice for me now is asking for things that I am fairly sure people will say no to. Just doing some exposure therapy basically to this. So that becomes really fun and
River Roberts: And I actually found for me one of the breakthroughs was I was able to go into a Trader Joe's and trade up s significantly. Like the manager of Trader Joe's was very very generous in his trades.
Ty Pattison: difficult places like asking for a discount at the supermarket.
River Roberts: So there's something to that. Really really good practice in
Ty Pattison: Ha ha ha.
River Roberts: in the uncomfortable place though. all right, I'm also happy to share as well. sorry, unless you unless you're you got more say.
Ty Pattison: So this is an invitation to anybody who's listening to to try this out.
River Roberts: No, I don't know. You can go first.
Ty Pattison: I mean I I have I have things to share, from my question, but you go ahead. I'll go Okay, okay, I got it, I got it. so I asked both Fable Five and I asked GPT five point six soul extra high. And I think they both have slightly different contexts on me. but interestingly they're I mean, maybe not surprisingly, they're kind of adjacent. So
Andrew Davis: 'Cause you're a Sagittarius time.
Ty Pattison: GPT says you keep trying to choose a path that allows you to keep every option open.
River Roberts: This that's this that's for sharing. Thanks for sharing, James. I really appreciate you guys going first first as well. So my my Kimmy Kimmy K three says, based on your memory and recent patterns, here's my honest hypothesis. Your main problem is focus masquerading as systems.
Ty Pattison: Which feels pretty true at this phase in my life, optionality is has been a priority. and it's expensive both cognitively and in committing to something.
River Roberts: You've built and keep refining increasingly elaborate frameworks to solve a problem that your memory suggests isn't actually getting solved. And when it says your memory, it's obviously it's it's past history. and this idea that you're optimizing the studio before any single project has escaped velocity, has has has yeah, has basically achieved.
Ty Pattison: And then Fable Five says just go straight to the point says follow through. Not ideas, not capability, not work ethic. You have all of those in abundance. The pattern is that you start something you start things brilliantly and leave before they compound. So it's essentially the same thing as like I am I need to work on
River Roberts: some kind of velocity. But I feel like that's kind of indicative of early projects. and then I asked
Ty Pattison: Trading optionality for a path.
River Roberts: one of my increasingly elaborate AI systems, Hermes, running on a whole bunch of different things. And
Ty Pattison: Yeah.
River Roberts: it came back with the the biggest main problem at the moment appears to be time scarcity. Explicitly called out as your biggest constraint. This manifests with information overload risk, building reliable agent systems and wealth trajectory. Mm, that doesn't quite feel as accurate, but I'm glad I could share those main problems with you, gents. Are there any other AMAs either of you would like to bring to the table? Yeah, let's let's do it. at least a hundred. Yeah, just in my shirt, like this. You just like create a bag of this Yeah, it's not a bag, but it's a shirt.
Andrew Davis: Right.
River Roberts: But is that is that too divergent or But but no bag, but I could wrap up in my shirt and carry them. It's like making a bag out of something that isn't a bag.
Ty Pattison: I have one, but it's silly.
Andrew Davis: I w I think three hundred. I would use a box.
River Roberts: I can see
Ty Pattison: So I'm gonna go into it. Okay, cool.
Andrew Davis: So this this is the lesson from these from today. everything is in the precision of how you frame the question. I think that's a that's a good takeaway theme from today. The the the skill in asking questions is gonna lead to lots of possibilities.
Ty Pattison: I've been thinking about this this idea of a spectrum of questions and the different types of questions you can ask people and one of the funny ones that came to me this morning, that I would like to know
River Roberts: So So Ty, now that you've heard two
Ty Pattison: the answer that both of you have is
River Roberts: two wonderful answers, what's what's your answer to that question?
Ty Pattison: how many oranges do you think you could carry without a bag?
River Roberts: Yeah.
Ty Pattison: What? No bag. All right. All right. I think No, I just think that you're wrong, I think
River Roberts: I got it.
Ty Pattison: but that's okay.
River Roberts: Well actually I d I do I do I do have I do have one la I do have one last question that I would I would love both your opinions on and this actually comes from Kidney and I I really love it. which is the World Bank says that AI could let developing economies
Ty Pattison: Yeah, kind of a loophole. But it stands. Andrew, how about you?
River Roberts: achieve in ten years what has previously taken a hundred years in other economies. But the best models are trained on English internet data, run on Western GPUs and are audited by kind of you know, Western, modern, developed nations. What what do you think would actually help more evenly distribute this these these these technologies in the coming years?
Ty Pattison: I was thinking like thirteen or fourteen. Like at some point but I haven't been using tools. See I think I generalized bag into tools. Which is not the question I realise now. But my my question that I would like you to both go away with is Calculate how Hebrian oranges and decide whether or not you could respectively carry a hundred and three hundred oranges. with that said, any any closing words from either of you? I I don't wanna know the answer to that question that I just asked.
River Roberts: Mm. Yeah, that's brilliant.
Andrew Davis: This this seems like a this is a perfect use case for fine tuning. So we were talking about fine tuning earlier, you know, a Burkina Faso fine tuned version, you know, without having to incur all of the costs of you know, retraining the whole model. one thing that comes to mind is just re l reframing education, letting kids solve problems their parents can't solve, rather than focusing on you know, competitive, like preparing kids to
Ty Pattison: I got answer that comes to mind straight away. Take train retrain the most capable models on the
Andrew Davis: compete in the economy, get, you know, create educational environments that are basically giving kids reasonably powerful tools and just inviting them to solve real problems for for the societies that are around.
Ty Pattison: region specific, country specific,
River Roberts: I
Ty Pattison: like if you think about the individual, the household, the city, the state, the country, the world, and think about those scales and train a model on say the data in Vietnam. Train the a model on the data in Burkina Faso and
River Roberts: I I feel like it is is kind of a combination of just some really basic infrastructure investments. So one one of the underutilised infrastructure investments I think is actually this idea of localized Wi Fi networks that kind of by default offer connectivity to everyone in the area, but connectivity through a local regional perspective as well.
Ty Pattison: have versions of these at a point in time with a particular context.
River Roberts: So like if you could have free internet access to anywhere that you were located to, but that actual connectivity access offered you local context to your models as well, you could have a really informed network. And then as you were using it, you would then equally contribute to that local Wi-Fi. And the only time I really see local Wi-Fi these days is on an aeroplane. But
Ty Pattison: and maybe that would also require gathering training data from sources that are not online because disproportionately. I just think that there's a space here to to use this technology in the way that we create these models, but have different training data that might require time and energy to collect, but it's probably worthwhile.
River Roberts: I actually think that old, relatively old technology could could work really well for training these these local smaller models that connected much more to local context and I think local context is actually gonna be what makes these technologies not only more accessible but much more valuable in local context. So yeah that's that's one of the things that I'd I'd I'd think is is worth investing in. Gents.
Andrew Davis: It's it seems like a really important idea about like what is your national strategy for enabling your population? Right, like countries like South Korea blanketed cities with free Wi-Fi, public free public
River Roberts: Yeah.
Andrew Davis: Wi-Fi. So if you begin to think you think about intelligence and computation as a public good that you want to provide your citizenry, what you know, can you organize your society? What is the most effective way to do that without you know tying you to US-based companies that'll extort over time? Like what what independent ways
River Roberts: Yeah. I think I think the national scale is too big. I think it like the national it it's it just doesn't quite capture the scale but I think would be really meaningful and and valuable to a lot more people. So I still believe it's more of the regional or more of the local or even just neighborhood version of this. like what what is your what is your neighborhood AI do in terms of being able to make everyone have more abundance in resources and in relationships and in connectivity and yeah.
Andrew Davis: Like computational commons.
River Roberts: Exactly. Yeah. Computational commons. You heard it yeah. Potentially first.
Andrew Davis: It's it's an honor and a joy to be able to join you guys. This is River, I I now I understand what you meant about, you know, Ty being a great collaborator and I was such an honor to bring a graph to this spin up for the first time.
River Roberts: Yes. Andrew, thank you so much for joining us on this podcast. Ty. A pleasure as always. Now we have this weird closing ritual which maybe doesn't particularly translate to the audio listeners, but we we kind of finish with one of these. And then we slowly go like
Ty Pattison: Thank you for being here, Andrew.
Ty Pattison: Hello, hello. Good morning, River Andrew.
River Roberts: How how gr how grows it? Yes, we have a guest today. I would love to introduce you to her to Andrew Davis, who I actually I met at an event last year called at something called AI Builders in Lisbon, Portugal. And we immediately hit it off. And I was really fascinated by Andrew's background, because he actually spent time as as a as a monk, as a as a monastic, and then he went into the engineering side of things. So actually being able balance those two sides, I felt like how many people are doing have come from those two worlds and somehow bring it together. And so he's been working on a project called Living Code and a few different kind of projects on that side of things. So Andrew, welcome. wondered if maybe you could introduce a little bit more about yourself and don't worry about too much because we're gonna go into like what you're actively building. in the next section. But any anything I missed out at a high level?
Andrew Davis: not particularly delighted to be here. And I as you said, River, y as soon as we met, I think your your blend of r real life facilitation, group facilitation, of you know, drawing out the insights from the participants in that AI builders event, plus mixing in the you know, set of questions to ask to your AI companions, your LLMs, and I really found that approach to be particularly rich and at the time and in general, right? Because we are entering such a weird world of trying to, you know, do deeper it's important that each of us do deeper thinking ourselves. Also extremely important that we do collective thinking, that we we continue to enrich our social interactions and social thinking. And that we have these alien artificial intelligence kind of entities that we're integrating into our lives. And it's really important that that not be strictly solitary undertaking, right? That there's a lot of risk of you know that dri channeling
River Roberts: Churely, chilling.
Andrew Davis: people into loneliness traps and so forth. So yeah,
River Roberts: And
Andrew Davis: I've as as you as you said my background roughly, you know, early life, trained as an engineer, spent 15 years as a Buddhist monk in the last twelve years, built a career in tech. And the last year I've been really honestly decompressing and not building so much as deconstructing and just kind of taking some time to rest. and try to get more embodied and yeah learn new approaches to life.
River Roberts: Yeah, beautiful. Thanks for sharing. And and just for a little added context, Ty, because you you you weren't there in this instance, but so AI Builders is a collection of different I builders throughout Europe and they host a bunch of different events. And it just so happens that one of my friends put that together, Arthur, who I'm sure will bring on at some point. But he asked if I could help facilitate for the day. And as part of it there were these kind of networking opportunities. And so I created a scenario where Everyone, I think it was like 10 minutes, and basically the invitation was everyone would ask one person what they needed help with, and then there was change partners. And then the next round was you would ask your AI agent what you personally needed help with, and then you would tell that to someone else, and vice versa. And then the third round is you
Ty Pattison: Ha ha ha.
River Roberts: and then th and the third round is you would blend. the the answers from both and then see where you got to. But basically by the end of ten minutes you had a hundred people that had kind of woven in how can we all support each other and ourselves in a meaningful way. And there was a real kind of energy in the room after that. so yeah, it's a really beautiful practice and and definitely, you know, recommend any facilitator in that kind of a situation to to adapt and adopt and and play with a format like that. It's a lot of fun.
Andrew Davis: And i if I can double click on that just for a moment, this specific prompt that Rivers suggested is just and this is a real danger that the you immediately lose all of your listeners into this exercise right now. So if you can restrain but do it afterwards, go to your favorite LLM and just type in what is my main problem and submit. And when it when I you know got the response back. It said, your main problem is lack of focus. You know, you're doing this, you're doing this, you're doing this, you're asking me about researching this. And I'm like, my god my gosh. I'm like, you're so right. I was so but it's funny though I then did that same exercise with a group of with another group of of people that I was working with and It was really interesting to hear everybody read out what the LLMs told them their prop main problems were. And most of them I actually quite resonated with most of the responses that the LLMs gave. And what it reminded me of was a horoscope reading. Like, you know, there's twelve plus Like what's your main problem? Well, because you're a Sagittarius, your main problem is, you know, lack of self-confidence.
Ty Pattison: Yeah.
Andrew Davis: And it it actually helped me to make a little bit more really at great risk of sending the podcast spinning wildly off topic, but it helped me to make more sense of things like astrology, where they just literally, you know, odds are you probably do suffer from you know, any one of these issues, but this is the one to focus on now. And so just it provides a little confidence and certainty that for now this is the thing for you to focus on. It's like confidence as a service in a in a time where there's just too much information, too many possibilities, to have someone or something say, This, this is the thing. it's actually a service.
Ty Pattison: Yeah. Yeah, absolutely. I mean two things. The podcast is called The Great Spin Up. So any kind of spiral that we're going into, fantastic. Let's go there. and and you're right. And it it makes me think what proportion of people in this day and age, and especially using LLMs, like if you've been using an LLM for a long time, how much of that How much how many people could you give that answer to? you have a lack of focus? And it would be the number one problem that they have. like yes, it knows about you, and yes, it has these things, but like just in the same way a horoscope is written to be semi-ambiguous, like there's a formula behind horoscope writing so that it applies to enough people. I'm sure there is also a way that these problems come out that yeah, that applies to me and it's useful.
Andrew Davis: Yeah, it's just like, yeah, you just have that all of you carbon based life forms tend to have, you know, this lack of self confidence and, you know, whatever else, not putting enough energy into relationships or whatever.
Ty Pattison: Yeah. Yeah. You just need to sit still for for five hours and then you'll find that or you can ask an L O Cool. River take
River Roberts: All right, Jens, how about
Ty Pattison: us away.
River Roberts: yeah, how I was gonna say, how about we get stuck in with the news? What what has jumped out to you this week or just in general as signal in the industry over noise?
Ty Pattison: I've got something to to jump in with. So there's a company called Guidelite, Guidelite AI standards.
River Roberts: Okay.
Ty Pattison: And they seem to be credible enough. I looked into it not extensively, but enough to make sure that it was that they were doing real work. and they basically rate companies on their safety standards because there have been a few things in the news recently about models breaking out of containment. and both OpenAI and Anthropic scored the highest with a C plus rating and Meta got an F on the on their score. and specifically like why it's why it's really relevant is because we're entering an age where more and more agents can use computers, more and more a agents can take action.
River Roberts: Mm.
Ty Pattison: And when it's just giving you an answer and you have to take action, it's not as critical or it it's not as important to pay this much attention. But now agents are getting the ability to to make things happen in the real world. So these things I think are gonna get more and more we need to start paying a lot more attention. Yeah.
River Roberts: Did did I'm wondering, did did this actually look into some of the open weight models as well? And open source models?
Ty Pattison: there was a piece in the article that talked about the open weight models and it basically just said that they are catching up, but I don't th like that their capability is catching up, but it didn't talk about this company reviewing them. so long story short, I don't have the answer because I didn't look deep enough into it, but it's an interesting question.
River Roberts: Yeah, so of course bring it on.
Andrew Davis: Am I allowed to ask dumb questions?
Ty Pattison: Yeah, yeah, yeah, please.
Andrew Davis: when I think about meta, I think about Lama as the model. And I'm not sure i you know, Lama I think of as I'm not sure is it is it an open source or open weight model or is it just free and you know, you can run it locally?
River Roberts: Yeah, o Lama is open source.
Ty Pattison: My
River Roberts: You you can run it locally. and then they've also started to release some other future open weight models under the kind of Muse Muse banner. So they've got a couple Muse Spark and Muse Glimmer as the first few ones. and you know, they they were kind of early to the game in the open source and they c kind of haven't actually done anything in the past six months. So this is quite a recent recent piece. Any anything to add to that, say?
Ty Pattison: honestly no. I was gonna make a distinction between open source and open weight. just the open source you. I yes. Good, okay. It's parking a lot it.
River Roberts: w we could we can do that in the deep dive though. We have a whole deep dive today on distinctions and definitions in the industry, so we can actually start to unpack these things. get into the nuance of them. I I I listened to there was an incredible podcast on moonshots just a couple days ago with this guy, Alvin Graylin. And he's spent a lot of time working in the Chinese ecosystem. And one of the things that he he talked about was the the Chi Chinese open weight models seem to be scor scoring a lot lower on cyber offense capabilities than what the American frontier models are doing, which is why I thought it was kind of interesting in comparison to that first bit of news that that you brought up. And it seems that a lot of these open weights open weight and open source models, even though they are they're open, they still have to go through the Chinese CCP approval process. And as part of that, the offensive capabilities seem to be, you know, like like a quarter of or if that of what these American frontier labs are having. So yeah, it's interesting seeing the the out of the box like what the offensive capabilities are and then maybe what the security vulnerabilities are that they're opening up into the industry. And then particularly in China where you already have an ecosystem that is kind of being short up, you know, the the Chinese internet has been kind of secured and firewalled to some extent for you know a couple of decades now. So I th it it it's looking like their approach to to strategy and security has actually been much more effective at this point. anything else on that that piece of news?
Ty Pattison: Hm, just the thought that cyber offense is usually easier. I mean, offense in general is usually thought as being easier than defense. And the labs potentially are doing something to hamper the offense because They're like if you just leave these models to their own devices, you would assume that they would have the same capabilities as all of these other things in life where you only need one attack vector in order to like get into a system, whereas you have to defend against everything. So yeah,
River Roberts: of them. Yeah.
Ty Pattison: interesting. And also you don't want to be shipping out cyber security weapons anyway. So good. Good.
Andrew Davis: But does it does seem to point to that there is maybe behind the scenes some more robust infrastructure You know, so that may indicate, you know, that that if the US had better systems, we could also mitigate some of those those risks, more centralized systems for and and but implicitly
Ty Pattison: Yeah. Yeah.
Andrew Davis: implicitly it also means that those Chinese government review agencies know what good offensive weapons look like, know what good cyber cyber offensive tools look like and how to
Ty Pattison: A really good point. I'm gonna jump in here with another story because it dovetails into this quite nicely. both Anthropic and OpenAI have models that are not released. they're holding them back from release. And when we're talking about like these these Chinese models putting their these Chinese companies putting their models out. And it it was making me think this morning about the book Superintelligence. Max Tegmark talks about if you're gonna create a superintelligence, you first of all need containment and then you need to bring the superintelligence in. It doesn't make sense to it's crazy to make a superintelligence and then try to contain it within a system that is built by something that has less persuasion, less ability to write code, that it that doesn't have these capabilities. So it's really interesting seeing both the stuff coming out of China and also these US labs like having these models that are built. So they've actually built the systems, but then they're not releasing to the public. So
River Roberts: Mm.
Ty Pattison: that was a piece that was jumping out. And the one line that came through is that now the frontier seems to be moving at the speed of containment. Like how slowly or how quickly can we figure out guardrails for these models in order to be able to release them to the public and it's not at the speed of innovation anymore.
River Roberts: Mm.
Ty Pattison: which I thought w is an interesting kind of slowdown step, hopefully. Whether or not it plays out like that, I don't know.
River Roberts: Yeah, I mean that that's that's a huge piece and and also just, you know, the speed of of co of of how quick cultures can move. I mean, the last few industrial revolutions have happened over, you know, eighty and forty years, sixty years. and and we're expecting, you know, something to happen within five years and you know, maybe orders of magnitude in in terms of its its its speed and and influence. So not only moving at the speed of containment but also moving at a at the speed in which cultures can actually adopt and and implement this. And that was another big thing that came out of this Alvin Graylin podcast was that the Chinese models and the Chinese government seem to be working a lot more on focusing on how these AI models can find their way into industry as opposed to racing towards AGI. And that seems to be quite a big distinction. and so there's there's a very different
Ty Pattison: Mm-hmm.
River Roberts: race at hand, even though culturally it's maybe positioned as you know, winner takes all, zero sum. It it it does appear that maybe the Chinese are playing a slightly different game than the American Frontier Labs and the way that's presented in in mainstream media. yeah.
Andrew Davis: The I'm I'm thinking the the the Metaphor of nuclear power is used a lot, of course,
River Roberts: Mm.
Andrew Davis: and comparing
Ty Pattison: And
Andrew Davis: you know, trying to make sense of what's happening here. But when you talk about containment, I'm thinking about Marie Curie and these early researchers on radioactivity, right, in the late 19th century, where they didn't know the the effects on the body, on on biological tissue of radioactivity, right? And so all of these early researchers were dosing themselves. Way, way, way, way higher than any modern safety protocols would allow. And you think about, you know, we all grew up knowing about radioactivity and atoms and these kinds of things. Mid-19th century, nobody understood anything. It's complete, you know, ridiculous science fiction that there was some invisible waves that could, you know, you don't notice anything, and all of a sudden you're being poisoned. but They they basically tapped into something that had been there all the time, an invisible force, could be used for good or for eat or for ill or whatever. and if you think about obviously that's relevant to the containment resource. If you think about nuclear fusion, obviously, yes. Step one, build really good containment. Step two, you know, create a mini sun.
Ty Pattison: Just get pushed.
River Roberts: Yeah.
Andrew Davis: yes, that operation is extremely important.
Ty Pattison: Hehehehe
Andrew Davis: But you you think about what are the invisible structures that AI entities could be messing with? And they are ob obviously infinite, but one of them is the i i i i to use a sort of a fancy term, like the semantic field in which we all live, right? Our sense of reality, our sense of what's true or not.
Ty Pattison: Mm.
Andrew Davis: And there's a sense in which all of us build our orientation in the world, our sense of reality around the use of words, the use of language, you know, what's true, what's not true, and so forth. And AI entities immediately, you know, I keep thinking about all of the AI-generated content that's out there. If you go look at LinkedIn or whatever else, and you know, this device about, you know, Of creating contrast, this rhetorical device of creating contrast. It's like, it's not just a rhetorical advice, it's a whole different way of thinking, you know, where these AI entities say, you know, it's not this, it is this. And
Ty Pattison: Yeah.
Andrew Davis: modern writing is just littered with that. And you can tell it does something, it creates this real punchy effect in terms of the reading. and it's it'll create effects on our psychology at some point. And then this question about what are the invisible structures that we're living in already that we don't fully understand, what are the risks, you know, that we maybe we all we do understand to some degree, but not not fully, and we don't have protocols for we don't have safety protocols like like
Ty Pattison: Yeah, when you're talking about this and you're talking about radiation, like I come back to this this saying that what we create creates us. And just like the simple fact that we're we're animals, we're creatures that learn through repetition, having exposure to these models and like having them echo back this form of writing, th these like AI writing patterns, I would be very surprised If you ran a study like a year ago and in a year's time you had enough of a gap for people to have exposure to this stuff, if people's writing habits didn't get drawn towards this thing just because of the the volume of literature that they're exposed to that is written in this particular pattern. yeah, it really fascinating.
River Roberts: All right, Zents, I got another piece of news. now this isn't actually new news, this but this is new news to me this week. And this is a report that a research report that Deep Google DeepMind actually put out in March this year. And it was it's one of the largest studies ever done, and it's with over 10,000 humans across the US, UK, and India. And they were basically doing a psychological study to see. if LLMs could manipulate humans at scale. And they were testing it across behavior changes in finance, health, and policy domains. And the answer was yes. But actually the the the the bit that is and they they tested three conditions. So one where there was like no chat, no, no kind of input. One is where the manipulation was very explicit and it was spoken out. And then there was one that was kind of implicit and it was like a little bit more implied, or as they kind of described it, as a little bit more of a whisper. And although the the manipulation was not the same in each instance and in each culture, they did find consistently that the more powerful psychological behavioral changes actually came from the AI agents that tended to whisper and and and work it into more subtle aspects of the conversation and kind of walk you down the path. And yeah, I mean, again, early, early signs that we're working with something here that has you know huge capability and you know how do we be mindful of that while we're building and and working with them. yeah. Thoughts?
Andrew Davis: If I can if I can offer one I'm I just moved to San Francisco six months ago and You can't hang in those circles without pumping into plant medicine, the term plant medicine a lot, right? And people talking about, you know, fungus and these these these plants that have psychoactive effects on our minds and to what degree maybe there was a co-evolution and so forth. and you know, it it's a popular thesis or popular point of view that, you know, these plants are actually acting through our affecting our cognition in ways that allow us to see different things and so forth. So this idea of whispers from AI entities influencing our thinking seems quite analogous to that. Something that is like a like a psychoactive substance or a psychedelic substance that's being introduced to the population and perhaps like psych psychedelics and you know, it's very distributed. The effects are highly dependent on you know set and your mindset set and setting the environment you're in and the mindset you're already in, so it's get highly variable depending on the people and so forth. But You could easily imagine you know, population-wide effects that are slightly shifting and in in insidious ways, in ways that are that are not obvious. I don't mean insidious necessarily in a in a malicious way, but like hidden effects where there's like some where where it's one thing to detect, does the AI escape containment, you know, and get access to the internet. It's another thing to detect, you know, is there some statistically significant change in cognition of people who are using, you know, Claude versus Chat GPT versus Gemini, whatever else? Do these different LLMs have, you know, some slight effects on the populations of their users that are Psychoactive, psychoactive substance. Should you be twenty one years old? Should you have to be twenty one years old before you are allowed to use LLMs?
Ty Pattison: Yeah, interesting. I mean it's so 'cause there are lots of things that are like this, right? There are I mean not a lot of things that that fall squarely into the same categories, but there are a few things, whether we acknowledge it or not, that really do have an influence on how we develop. there recently it makes me think of social media. Recently in Australia, I believe social media was banned for under sixteen year olds. Is that right, River?
River Roberts: It is indeed.
Ty Pattison: and there's a couple of other countries who are like piloting programs like this because social media like really has this ability to to grab our attention and do it in a way that we're not often aware of. It's insidious in that way. And there's the same thing with with AI, and but instead of Who is it? Sam Harris talks about this. the what was he at the Institute for Humane Technology, and was talking about how social networks tap into our brains and take attention. And these AIs have the ability to do that, but for connection, which is a really, really p if not the most powerful human emotion. So yeah, potential to like shift the whole population in that way. and maybe to put it on like a connection versus loneliness axis as well. Like if you can get all of your connection from this, maybe you learn to that's the place that I go every time I want to connect with something because it's easy and there's no friction and there's like it's trained to specifically do this with me versus if you're like in these circles, at least you're having conversations with other people, right? Like the AI is the third entity, not necessarily the go to.
River Roberts: Yeah, I was gonna say that's that's the distinction I was gonna say that's the distinction that I'm hearing in this as well, is that you're being manipulated by a relationship that you're not having with anyone else. It's just you personally, in a way. At least in social media, you're being manipulated a little bit more out in public and it's out in the open. Whereas these these conversations are happening privately. but again, manipulation doesn't always have to be necessarily perceived as negative. And I have to say that At least I feel like I've been manipulated very positively over the last couple of weeks in terms of trying out this this ketosis diet, which if like getting enough information and and nuance on having this protocol really work for me, would have taken me so much more time, but actually having an agent there supporting me and be like, you know, can I have a sip of coffee or can I have a spoonful of honey? And it was it's it's incredible like how refined it can be. And you're like, well actually I feel my health has improved and it's been manipulated by you know, this agent. Now whether or not there's a company underneath this thing advertising to me and trying to push, you know, electrolyte purchases as based based in in like something that I need a supplement that I'm not getting when I'm on ketosis is a very different thing. And
Ty Pattison: Mm-hmm.
River Roberts: actually that's a another piece of news that came up. OpenAI is rolling out ads in in I think what, 31 European countries. So I'm very curious to see how that has responded, how that is received. because it seems they were kind of shut out of the American market with that Super Bowl, that Super Bowl commercial that Anthropic put up to put a OpenAI out of the ad game quite quite quickly, or at least to start off with. So very interesting to see how that will be rolled out in Europe now. Andrew, do you have
Ty Pattison: Amen.
River Roberts: a piece of news you'd like to to share with us?
Andrew Davis: yeah, I do. It's not n super new new, but it's it's an article from it came to me from Rutger Bregman. this is a graph of capital expenditure
River Roberts: just can I just stop you can I just stop you here because you have brought the first graph we've introduced to this podcast. So thank you so much for elevating this this podcast with its first graph. I feel like I might need to create some sort of certificate that you can print out one day and hang on your wall.
Andrew Davis: It my my love language is URLs and graphs. So on
Ty Pattison: Perfect. I I think we sorry to interrupt, I think we might need like a section here which is like Andrew's grand graphs or like whenever someone brings a graph have a celebration of like two we love graphs.
Andrew Davis: But it's I'm feeling very very warm and fuzzy to be received, to have my my my graph fetish so warmly received by you there. I appreciate that. so on the vertical act.
River Roberts: It's it's a great graph. I can already I can already picture Trap with a you know, with a printed out on a sign and his little stick. It's it's a fantastic graph.
Andrew Davis: I'll I'll print it out and sign it with a gold Sharpie if you if you get a chance.
River Roberts: I appreciate it.
Andrew Davis: So on the vertical on the vertical axis is amount of dollars expended in today's dollars. On the horizontal axis is the number of years over which that capital was deployed. And so if you think about the US railroad program, like the mid 19th century to the early 19th century, deployed 550 billion dollars, but over the course of 71 years and transformed the country, right? Massive shipping. And then the interstate highway system, you know, some decades. Later, six hundred and twenty billion of today's dollars deployed over thirty-seven years, as obviously completely transformational for the United States. And this is a bit of a US, you know, centric graph. Apologies for our perennial selfness.
Ty Pattison: We'll extrapolate, it's all right.
Andrew Davis: Anyway, we we we were the ones that sent a man to the moon with the Apollo program for just $257 billion over 14 years. Manhattan Project, build a nuclear bomb for just $36 billion in five years, and then rebuild Europe, starting with Germany with only $170 billion over four years. And then what do we have up here? Data Center CapEx, Projected at $930 billion over six years, and this is already a few months old. The actual planned projections this year, and this this data comes from the Brookings Institute, well over a trillion dollars. Planned CapEx with the intention that that will actually, the infrastructure expenses will be done over the course of the six years. And I noticed just today that this excludes the Chinese hyperscalers. So, you know,
Ty Pattison: Wow.
Andrew Davis: we don't even know what the Chinese are deploying in terms of this. And so if you look at the big picture of this, and this is setting aside all of the software investment, right? The investment of time and energy and you know. This is just raw computing capacity. Far and away, the biggest infrastructure project humanity has ever done. And I don't know where the Chinese, you know, what was it? Three Rivers, I'm gonna masquer three rivers, six one belt, one road,
Ty Pattison: They have one belt, one road. Yeah.
Andrew Davis: yeah. Where where their infrastructure projects.
River Roberts: Yeah, the three vis the the gorge. Yeah.
Andrew Davis: Yeah. and but over six years, right? So the intensity of the investment and what this makes me think of is that mostly these are you know startups and venture capital and they want they expect return on investment you know typically in five years that's the venture capital timeline right so there is this idea that you're putting all this money in There's gonna be whatever else might be true, there will be enormous pressures to reclaim, you know, recap that investment very quickly. And then people will be doing whatever they possibly can to manipulate the world such that you can get your money back real fast. and I believe that that's just a driver for so many other things happening.
Ty Pattison: This is such an interesting perspective and especially from the conversation that we were just having about how potentially AI is is seeing these things makes me think of the what is it in sapiens Yuval Noah Harari makes this argument that we didn't domesticate wheat, that wheat domesticated humans. And this building out this amount of infrastructure this quickly feels like the AI building its own infrastructure to be able to live. And whether or not that is just where all of the computers and the the, you know, artificial intelligence goes and lives and it's separate from the infrastructure that is built for humans is the question that comes up for me. Like, yeah, we're getting like all of these humans to build this thing. Is it for us or is it for somebody else?
Andrew Davis: reminds me of I've got a friend and colleague named Reese Lindmark and he was working for a long time on I think it's sub it may be a Substack post or what blog is something is called what Patterns want, what patterns want. And this idea that patterns have a tendency to actually draw out. Like if there's certain patterns, something that has, for whatever reason, a mathematical geometric p you know, s strength, stability, it tends to draw things out. That there, you know, and this reminds me of Michael Levin's work on platonic spaces, and that there are there are potentials that are trying to be drawn out.
Ty Pattison: Yeah, yeah. What are the truths about the world?
Andrew Davis: in right and what are the what are the potentials implicit in these you know gps that are patterns that that want to draw out more value want to draw out more and again i'm I'm with river like we I don't I don't think it's we shouldn't be overly suspicious or unnecessarily suspicious it is extraordinarily Potentially benef beneficial. I think humans do need to be hacked, have their minds hacked and redirected in a bunch of ways that would be more beneficial for us. and we do need more intelligence as a species. I don't think we're suffering from a surplus of intelligence. I'm certainly not. but we do need to be careful how it all unfolds.
River Roberts: Yeah, I mean it is
Ty Pattison: And and you're right. And we do that for each other. Sorry, go ahead River.
River Roberts: No, I I I I feel like there's a a a real risk here in many ways that is is actually potentially quite similar to the Soviet Union during the Cold War. And so during that war the USSR invested about twenty five percent of its GDP in military. And part of the reason they actually lost the war is they they over invested, they overindexed in the military machine that actually wasn't supporting its citizens. And so I wonder if, you know, the the future doesn't repeat itself, but it certainly rhymes and that maybe there is this massive overinvestment. I'm not saying that investment doesn't belong here, because it obviously does, but it's at the scale and the speed at which it's happening, is this actually an overinvestment and overindexing in this in an industry that hasn't necessarily, and particularly from the US spec perspective, hasn't necessarily shown a lot of benefits to the industry. or it hasn't necessarily reached a lot of different aspects of different industries so far. In fact, you know, the way it's positioned in the US is basically I think some of the statistics is 70% of the US is white collar workers. And this is precisely the technology that is going to put most of those white white collar workers out of work and displacing them. So you're accelerating the investment in a technology that is displacing your your culture in a way. Whereas you look at other cultures and countries around the world, that seventy percent is not the same. So in Africa, in Asia, there there's a lot more of the industry that is still you know working more blue collar jobs and and service layer jobs, not so much white collar. So I think there's a real risk in this accelerated investment and the expectation of you know huge financial returns that support this this upside risk. yeah, I mean it's it it could really be creating the conditions for a pretty tough time in America as a as a as a nation state in some regards. You know, and this is kind of
Ty Pattison: Mm-hmm.
River Roberts: this is kind of what happens in in in various ecosystems as well. Like if you have a too many large players and too they they kind of tend to monopolize the whole industry and everyone tries to follow behind and then the whole thing kinda collapses because there isn't enough diversity holding it up. So I wonder if that's the conditions that are starting to emerge.
Ty Pattison: Mm.
River Roberts: But yeah, thanks. Thanks for that graph. That was a great first graph.
Ty Pattison: knocked it out of the park. I don't know if we can get a better graph, honestly. two two quick points. Just the graph, the like extreme outlier of of that data point is is wild. and River, I don't I don't necessarily know if I agree with you on the white collar work. I I think it's so nuanced here. Like yes, we are having the ability to replace all of this white collar work. And the tension here is between the timelines for me. Like Andrew, you're talking about the VC runs on a five year timeline. That's gonna be a crunchy thing. There are lots and lots of jobs that you can't write down everything you do in a day. There are lots of white collar jobs where we actually they're much more diverse and nuanced than we think about. And so like across the world we have different kind of mixtures of these things, but I think there's a real opportunity in applied whatever it is. Like we probably won't have that many pure mathematicians, but we have a lot of applied mathematicians in different fields. We might not have like that many theoretical physicists, but we'll have people who are doing research using these tools that are then applying them to the world. so I don't know.
River Roberts: Totally. And and just just so I'm clear.
Ty Pattison: To me there's like it's quite crunchy.
River Roberts: Yeah, I I mean I absolutely agree with you on that point. And and just so we're clear that I I I don't think these tools are gonna replace people. I think it's gonna displace people. So for instance, we're already seeing I think forty to fifty percent of the youth market in America are underemployed. So I think the unemployment rate is something like ten percent. But then when you look at the underemployment rate where a lot of these people aren't, you know, working up to their educated standards or the university standards, it's it's underemployment. So I think there's going to be this massive displacement. And that's the crunchy bit that I think we're talking about. And again, when you're dealing with a culture where, you know, 70% of the the the culture could have this displacement. I think and you've got this political unrest and kind of bifurcation of of the culture in some regard could get really kind really crunchy, really crunchy quite quickly. Yeah. You know, I feel like I'm just saying crunchy a lot because I haven't actually eaten anything crunchy in two weeks. So so if you g so if you give me like I can't like it's like meat, eggs and fish. It's nothing is crunchy, man. I'm really craving a crunch. I've got one more quick bit of news, and then I'm not sure if either of you do either. But one of the things that jumped out this week for me was the Tsingha University out of China and Bidance recently released Coda Agent. And this is an open source system that writes the GPU kernels 100% faster and 40% better than Claude Opus 4.5. And so basically CUDA is the software layer that the NVIDIA chips have been operating on. And it allows the NVIDIA chips to kind of work across all the different chips. And that software layer has kind of been part of the competitive advantage for the last couple of years. But now that we're seeing that software moat potentially start to evaporate, you know, what w how does that actually change the the the chip layer of this industry? yeah, that was just a a a quick one that popped up this week.
Andrew Davis: And just so I can understand that, so the idea is that you might be able to write an alternative kernel, a kernel that's an alternative to whatever Nvidia's native operating system is for the GPUs?
River Roberts: Ex exactly. And then you then you have impr interoperability across all different chips. So you're not just tied to NVIDIA chips,
Andrew Davis: play.
River Roberts: you could use AMDs, you could use Cerebrus, you could use all these kind of different chips and and work across them. So there's this huge upside in in that side of it. And and I I I again I think we're gonna see more of this kind of open source agentic infrastructure emerging, which is is super exciting. two other two other quick s news stories. I
Ty Pattison: Mm-hmm.
River Roberts: just want to mention the high lines. One was there was an Interpol report that fifty-five percent of African cyber attacks include AI now. And DARPA flew its first AI piloted F sixteen. So aga again, there's there's some
Ty Pattison: Whoa.
River Roberts: pr some pretty crazy moves happening in the industry right now. and the last one, which I think is a r really
Ty Pattison: I have question.
River Roberts: big signal piece, just quickly tie, is Stripe is buying open router. or they're setting up this this purchase at the moment. Stripe is buying open router for over seven billion dollars. And so open router is this layer where you can switch your models when you're working with different agents. So it's kind of proving out that this switchboard layer is actually the the financial infrastructure for a lot of the industry. So yeah, huge signal there.
Ty Pattison: Could be a a big rabbit hole, I'm not even gonna ask right now. who's
River Roberts: Mm-hmm.
Ty Pattison: DARPA?
Andrew Davis: The Defense Advanced Research Project Agency. It's the it's the US military R and D folks, famous for creating things like the internet.
River Roberts: Yeah, the underlying protocol for the internet, yeah.
Ty Pattison: Okay, interesting. Yeah, I was wondering which country. Cool. I have one quick piece of news. We don't need to dive into it, but the former a former Google Chief Science Officer, his name's Jeff Dean. basically Yeah, Jeff Dean left Google after twenty-seven years, started a company called Discovery Loop and is doing it with three other people. And this was interesting because we have lots of young people saying, you can build things with AI. This is evidence that people who have been in big tech for a long time are going and making bets on their own and not hiring big teams because they can do so much with so few people. and especially if you have expertise in a particular domain, I thought this was really interesting. Yeah.
River Roberts: Yeah, yeah. I feel like that's also a massive signal kind of almost against the culture of Google at the moment in terms of they're such a large company that they're they're obviously quite risk adverse. but then also in the same breath, Sergey Brin also came back on as acting CEO to kind of put Google back into what's effectually referred to as founder mode. so we'll see if there's some some different moves coming out of Google in the in the coming months. All right.
Ty Pattison: Yeah, interesting.
River Roberts: So shall we move this ship along to building?
Ty Pattison: Building, yes please. And I like that you use the word ship. I'm gonna take this opportunity to jump right into this because a river I have some insider knowledge that you have been building, working on something that has to do with a ship. Would you like to jump into this?
River Roberts: Sure, let's let's let's start off with this. I look, I I'm a little bit embarrassed to be sharing this publicly, but at the same time, you know, I did I did kind of in the back of my mind line this up and maybe not give myself the best conditions to produce the best result, but I'm still kinda happy with the amount of time and budget that went into it. So basically there's something called the future vision exprise, which is a call for these kind of like Pro-topic visions of the future. There's kind of this understanding that Hollywood really only produces dystopic visions or these like really boring polyannerish visions of the future. And so this is an X prize to to to get a bunch of these visions. And I've been inspired by a lot of sci-fi and Star Trek over the years. So I figured I would I would give myself a a couple days to explore with the current tools what I could produce for this three minute trailer. Now give it in my I have no experience with writing scripts or anything with films or any of these video tools at anything. So I started from zero. And with that being the case, how about we just push play?
Ty Pattison: Nice.
River Roberts: all right. How embarrassing.
Ty Pattison: Cool. not embar I mean claim the embarrassment please. what what sorry go ahead Andrew
Andrew Davis: Yeah, I mean I I don't know, it maybe that's the most instructive thing for people to to see that one could be embarrassed by producing something like that. I thought it was marvelous, bravissimo, right? This directorial debut review is fantastic. I I I think it's quite important to the Protopia thing, obviously super important, right? And then to figure out How to knit I mean, it's so much, you know, we all complain about this too much negativity on the news, right? So just to cast out positive futures, right? of various forms feels so important, right? And I I love how evocative this is. and that, you know, it was a it was a two day project, and yet you'll be embarrassed. So this I think is really instructive for everybody about how you know, our self critical nature is so hyperactive. I really appreciate your courage and putting it out there 'cause I think the you know, we all need to be trying to put forward positive creative efforts.
River Roberts: Yeah, totally. Well, yeah, thanks. Thanks for that. And so I
Ty Pattison: River.
River Roberts: spent about three days on it. I spent about fifty dollars worth of credits across a few different platforms. And again, I'm coming from like nothing. So I've never written a script before or a storyline or any of that stuff. I hadn't used any of these tools before either. So it was a really interesting learning curve in terms of like timeline, budget and these new sets of tools. And so with that being said, I'm I'm quite happy with the outcome. And and yet I think if I was to do it again, I'd do a little bit more work on like the story side of it and a little less on the the the visual sides of it. 'cause at least in telling a compelling story, I feel like I should have spent a little bit more time on this this the script of it, 'cause that informs so much more of the actual build. So getting the sequencing right, I I definitely learnt a lot through that. But I will say I actually really, really enjoyed the process and again, pretty impressed with what these tools are capable of given you know the input that they had from a complete novice. So yeah, definitely enjoyed the process. Even though sharing it is a little bit embed it a little bit embarrassing.
Ty Pattison: Can I ask what's the
River Roberts: Please.
Ty Pattison: Wha what was what was the inspiration? What pieces are you weaving together in here? Are they long like pulled in things that you've been thinking about? Is it recent? What's
River Roberts: Yeah, it's yeah, so s
Ty Pattison: the
River Roberts: so I d like for me part of like one of the bits of inspiration is like, you know, Star Trek is this kind of future vision, but maybe what would the first vessel that ever kind of seeded something like the whole Star Trek thing be like? So I kind of saw kinship one as, you know, before we have starships, we have like obviously a ship that is on our planet and it's as much about exploration and sharing as anything else. It's not coming from a place of economics or military or politics or even non profit. It's it's actually just it's coming from a very different place. And that's what I was trying to capture. But you know, some of the feedback that came back from some close friends is like, yeah, this feels a little bit elitist and and all of that. So I know it didn't quite hit in that regard. But that was part of the inspiration. And then two films that really kind of called to me. One is a near future film that came out a couple of years ago called The Creator. And I kind of loved the imagery of that. And then on the other end there was this comedy called The Boat That Rocked. And I just love the juxtaposition between those two. So those were kind of the the inspiration. It was kind of like Star Trek meets the boat that rocked and the cre the creator. So those those were the the the three references I really pulled from.
Ty Pattison: is a fantastic film of like comedy and drama and it's so human
River Roberts: Totally. Yeah. Yeah. And and like I think there is something about this like pirate essence that actually moves cultures forward in quite meaningful ways and you know, pirate cultures back during the European era of of the spice wars were actually some of the more demographic demographic demographic democratic societies that were emerging. The the pirates had this kind of like code of honor and and a way of being very democratic in the way they're operated. So there's something about the seas and the oceans that create the conditions for potentially something new to emerge without us having to go and be space cowboys. So that was some of the inspiration. so the so the tools that I the tool the t the tools the tools that I use so I was using largely Kimi for research and and dialogue and and all of that which is Maybe not the best tool I should could use for that part of it. And then the video tools were Google Vio three point one, Kling three, which is a Chinese model, and then LTX. those were largely the video models. And then for the film score, which I actually found really hard, but the one that I actually ended up using was one that I just one shotted at the first time I I put a prompt in. And I've probably tried about five different scores, but it's quite hard to get it to match up. And that first one kind of matched up close enough. I was like, that's good enough. but that was with Google, Google Lyra. And this tool is incredible, L-Y-R-I-A. Its ability to produce music scores with you know a couple of lines of prompt is is phenomenal. So definitely worth exploring more and more. and then to
Ty Pattison: Can you explain that for me a second? W what do you mean by a music score or a film score?
River Roberts: So the the the music track for that whole trailer was produced when in one prompt. And
Ty Pattison: Uh-huh.
River Roberts: so the music layer is typically called the score. So the score of the film. and I and and the vocals were done as part of the LTX video machine. So you can put some you can put some script in the video prompt and it will produce audio. Which is also where the audio quality was a little bit all over the place, but it I'm like it's good it's good enough to to kind of get get the the trailer out. So yeah, all of these tools very interesting and definitely worth exploring.
Ty Pattison: Just one question on like tactics or like making this stuff happen. Do you use one platform to use all these different tools or did you go to each different platform and like play with them or download them and move them? Like what's that process?
River Roberts: Yeah, so to compile it all together I used a open source just Mac editor called DaVinci Resolve. and then in terms of producing each element, yeah, I actually had I I kind of went to each individual platform. So I used Cling, I used LTX, and I used Google Flow. and those were the three main ones. And since then I've actually come across a a An ecosystem called FAL.ai, where you can actually use a lot of these systems as well. And the way they're actually set up is that you can plug these systems into your API. So actually actually start running them through like a claw code or one of your your main agents. But I was like, by the time it takes me to set that up, I think I would have lost a bit too much time. So I just I just spent time within each each ecosystem rather than trying to bridge them. Yeah.
Ty Pattison: Yeah.
River Roberts: Anyway, I don't wanna take up the h the whole build segment, so I'm curious what have you gents been building lately?
Ty Pattison: Yeah, Andrew, do you do you have something you wanna talk about?
Andrew Davis: Sure, yeah. I I mean I I'm just soaking up both your tech stack and the creative output and I'm I'm off sort of daydreaming and fantasizing river about trying to follow suit be a be a stream to your river. so what I'm working on these days is something called weave. And so it's basically I I put the URL in the in the in the chat here, maybe we can add in the show notes, but weave if you go to search for weave and my company Living Code, Weave is basically a programmable Zoom. so it allows you to define in advance some kind of a structure for meeting to orchestrate a meeting. We've all been in meetings where there's like fifty people and one person droning on and on and on and very, very low engagement from participants. And hopefully many of us have been in you know online when you when you have to do an online meeting Many of us hopefully have been in environments where there was a really good orchestration of breakout rooms and giving people prompts, putting them in groups of two or three or four to discuss these these ideas. and maybe you could do a ver maybe we could do a version of the spin-up that's live and interactive, right? Where we use Weave to choreograph, like asking people to come together 'cause to discuss. 'Cause I I think the thing about video meetings is they're th because it's two-dimensional. And it's so passive, typically we get very low engagement of our brains, literally. And so people then their level of participation is very low. And so most companies people end up drowning in lots of meetings where there's not really much engagement during the meeting, then there's also not much traction. So what weave, like from meeting to
Ty Pattison: Yeah.
Andrew Davis: meeting, capturing I capturing the ideas. So what weave allows you to do is basically define in advance what are you meeting about. And then some prompts. And you'll take the prompts and you'll go into breakout rooms and Weave puts the prompt in front of everybody in the breakout room. so you can always see what's the prompt. And it allows you to record inside each of the breakout rooms and transcribe inside of each of the breakout rooms, or opt out. And you anybody can declare that they want this to be anon that breakout to be anonymous, but to begin to capture the actual fullness to to both instigate a fullness of discussion. Discussion, interaction, collaboration, creative collaboration in the context of meeting, but then also to capture and harvest the ideas that are generated and to consolidate those ideas. Future iterations, my goal is to you know both Right now I'm creating text transcripts, but I'd love to be able to even create visual artifacts. like if you've been in these, you know, sometimes there are meetings where people will draw live during the conference, you know, some of the main themes. So obviously it's very you know, these days the technology is very conducive to doing that live. And like let's say we have a discussion, what would be a single visual that encapsulates the whole of this conversation in some kind of an evocative way? in a visual graphic way and that gives a way for our memories to anchor key ideas and then I'm imagining also you know scaffolding that up so each breakout room has their visual but then there's some composite visual that encapsulates elements from the whole organization and just basically doing more to light up everybody's minds and activate their memories together in organizations and nonprofits.
Ty Pattison: Mm. What you're talking about too makes this like thing of of meetings being flat, of feeling flat. They're very intellectual. You're in your head all the time and decisions are made emotionally and memories are stored through senses. Like if you smell something from your childhood, you remember that thing. So like I love this idea of making it a yeah a a deeper field of interaction with whatever it is that you're doing so that you have these elements of feeling or of creating together. super cool. What was the like impetus behind this piece? Were you yeah.
Andrew Davis: thousands of hours of boring meetings was the primary impetus.
River Roberts: Ha ha
Andrew Davis: And then and then learning
Ty Pattison: Yeah.
Andrew Davis: learning what's liberating structures, learning how to facilitate much more dynamic meetings and getting pretty good. I be I'm a pretty good Zoom disc jockey, like breakout room disc jockey, like choreo orchestrating and so forth. And I was just noticing that most people don't have that skill. There's a lot of fumbling trying to, you know, set up the breakout rooms and size them right and so forth. And so I I really wanted to reduce the the the barrier of entry or the reduce the yeah the you know lower the barrier of entry for people to facilitate really rich engaging meetings and for people who know what a rich engaging meeting feels like to be able to do it without a copilot so you can focus fully on the meeting and the software basically runs like pressing play on a playlist.
Ty Pattison: Yeah. This it touches what we were talking about earlier, which is that potentially these systems can create containers for humans to to work in and we do this for ourselves. We have facilitators that come in, we have people that come in and create this space so that we can be creative and like be specifically in this role. Think this is really interesting for allowing humans to be in that like creative, generative, connective space without having to like do all of the logistics and remember that there's the whatever. remember to buy milk from the shop. You know, like all of these mundane things that somebody should do. And it'd be great if a computer did it.
River Roberts: Yeah, I mean this this is incredible, like being able to somehow capture the collective intelligence of groups that are coming together. And, you know, the unique thing about digital groups is they happen at a size and a scale and in many instances a diversity that is just very impractical if you were to do it physically. also the ability to have the frequency of some of these large gatherings happening is is incredible. So I'm actually surprised there haven't been more movements with products like this in in the past. So yeah, super exciting to see how how this goes. I do have a couple of questions around like what tools are you using to to build this and yeah, what have you found quite straightforward and and what has been more insightful and and practical as you as you build these?
Andrew Davis: so first of all, like a lot of people, I had drifted away from doing hands on development and then the latest generations of tools just make it so Blissful and easy and so forth. I don't have nearly the diversity of tool stack that you you have, River, but I was using Claude Code as the underlying engine, conductor, which is an orchestration engine that allows you to run multiple Cloud Code sessions, each in their own go git work tree, and G stack, Gary Tan's G Stack from Y Combinator just you know, basically sets up a bunch of a bunch of agent skills to help in increase the quality of of the work that you're doing. that's pretty much it. Go you know, Google Anti-Gravity as the as the IDE, but and and then
Ty Pattison: Hm. quick question. Good.
Andrew Davis: there's a stack behind it, right? So it's a video it's a video conferencing tool. So there's a there's a tool called Live Kit, which is basically generic commodified video conferencing tool. So that obviously is doing all the heavy the heavy lifting of the video processing and so forth. It's outsourced.
Ty Pattison: and you have login as well. Is that through Firebase? Is it Clerk or is there a s is that built in in some way?
Andrew Davis: Yeah, you know, I've got I've got what what what do I have? I've got I've got social password password. I I think I'm using Neon Neon as a Postgres database, cloud based Postgres with Neon.
Ty Pattison: Interesting. This stuff is maybe a little bit nerdy, but it is.
Andrew Davis: And Google Surface. We could we could go all into the stack if you want. There's a few other
River Roberts: Yeah.
Andrew Davis: things. Deep deep gram or deep gram for audio transcription and
Ty Pattison: Hmm. do you find that you change your stack? do do parts of your stack change or when you're building things, are you usually like, this is the Postgres database that I do you have go to's, do you have defaults?
Andrew Davis: I again I had not been actively coding for a for a bunch of years and so I was grateful for some from getting a well researched stack from a a friend of mine Whose name's Roosgar, anyway. I basically took his his recommended stack and he recommended Svelte as the JavaScript framework, which is famous for being like the most coveted framework. Like most people don't build their apps in Svelte because they're they've already, you know, been building in some other technologies. But if what's it Stack Overflows Developer Survey, Svelte was the JavaScript framework that has been most people most wanted to build. in over many many years and it's just got a really nice you know programming idioms where it's very intuitive the way it's written and it's got a solid underlying infrastructure so I think if you just go into Cloud Code and you don't specify what you want to build in it's gonna put you in React and TypeScript and which are reasonable choices but you know so I yeah Svelte was probably the the big difference and I've been really pleased with the readability of the code and the clarity of the you know it's a it's a it's one of these reactive frameworks but it's it's clean and simple and I like it.
Ty Pattison: Interesting. I I love this level of detail. cool. And also as a non technical person, I realize that the readability of the code doesn't mean anything to me. Like it is obviously important and then I think about using it myself and I go, I'm not gonna read the code which is a
Andrew Davis: And I I I don't read a lot of the code, but I don't know. I it it is that that sort of craftsman aesthetic that
Ty Pattison: Mm.
Andrew Davis: I always valued when I was doing more hands on manual coding. and I do think that, you know, that that there's a there's an aesthetic to software development. That you want things to be simple, modular, you know, don't repeat yourself. And so we're obviously in very early days of figuring out what the craft of AI-powered software development looks like. But we know that there are extremes where you create, you know, massive, undecipherable code bases. and then there's ex there's other possibilities where you have a well-maintained code base. I imagine that even the LLMs like a nice clean code base where the language
River Roberts: Ha ha ha.
Andrew Davis: where the logic is centralized that Help lends itself to you know the logic around each part of the UI is centralized with that part of the UI. if there's a degree of sense making that you have to do even if you're an LLM
Ty Pattison: In in my experience it prefers to make new files. I've I've had projects that have got like four or five different different like CSS files or what it was styling systems. Like, why does this button look like this and this button looks like this? And it goes, 'cause this one pulls from this and this one pulls from this and like, Why did you not just make one? What is going on here? This is in fairness, a few months ago.
Andrew Davis: And I I d I do think that there's a really, really important level of you know, software development has always needed a strong architectural foundation to build good software. The actual development is, you know, the fine relatively fine-tuned detail compared to the underlying architecture. And so we think about what's the architecture in a vibe coded world, the choice of tech stack. so I, you know, I'm standing on the standing on the shoulders of giants with you know Gary Tan's G stack where he's centralized a lot of his and white combinators ideas about what makes good quality software so having those skills implicit you know I I I'm sure there's people who could quibble about the relative merits of different different sets but having something like that gives me much greater confidence that the that you're gonna keep your LLMs on on the rails, on track relatively. and then, you know, I've got my own skills that specify the tech stack I wanted to build in. I'm like, you we're building in Svelte, you know, it's just you know, some early good research to make sure that those design decisions are being respected at every stage of development goes a long way. And I don't know that you have to be a you know, a skilled software developer to be able to just say, hey, give me a well-researched set of, you know, recommendations and help me put those into you know designing my my skills and stack.
Ty Pattison: make sure that these are used at each point. I I have one more question which is about the Gary Tan skill set. I think really cool to like tap into this. Are there certain skills that you always go to in that stack or are you using the whole stack?
Andrew Davis: Started using it with this project maybe four months ago or something like that. And so you know, I'm sure it has vast capabilities that I'm under underutilizing. The first thing that they they recommend running is what's called office hours, and it's based on the idea of the Y combinator office hours. And it runs you through a bunch of very confronting questions about your decision to build this app and why and have you done any market research and I have shied away from diving into the to the VC world directly myself. I work for startups but but not had to do my own pitches and so forth so far. But the Questions were really quite aggressively confronting. you know,
Ty Pattison: And
Andrew Davis: using things like, you want to build a video conferencing app. do you have any idea how many of those have been tried? It's like the graveyard for like projects, and you're gonna compete against Zoom, Google Meet, and Microsoft Teams. Yeah, good luck, buddy. And you know, do you have any You have any evidence whatsoever that anybody would ever care about what you're building. And I could appreciate I aspire to have the strength of ego to walk into a discussion like that and just be imperturbable. but I honestly had to turn that off. I'm like, I did that a little bit.
Ty Pattison: Yeah.
Andrew Davis: Yeah, yeah, I know what you're saying. I'm gonna build it anyway. Tell you what, I'm gonna do something else. Can you just do the bit where you help me build the thing instead of the thing where you try to pick my idea to parts? Like I'm not asking for a million dollars. I'm asking to help me build some software.
Ty Pattison: I've had the exact same experience, honestly, going through those questions and being like, I I don't have good answers to these. I'm gonna build it anyway
River Roberts: Yeah.
Andrew Davis: And I don't regret building it. Jerry Tat, you know, and my combinator.
Ty Pattison: Yeah. But it's a it is interesting to to confront that at the start and be like, why am I building this? Am I building it 'cause I want to see what it's like when it's built and then maybe do something like the the bed mass on this we we don't need to do in order. It's not as critical because building it gives you some clarity in my experience. Yeah.
River Roberts: Yeah. Well also like anything that is built on VC mindset has to be immediately monetizable over a short period of time and is really much kind of cornered into how can you become the next unicorn, that doesn't necessarily mis mean that what you're creating isn't creating value in some way. It's maybe just not creating as much value as quickly as what a VC would expect. So I think, you know, those those starting questions, albeit are worth maybe considering to some extent. certainly shouldn't override maybe your initial intuition to go and explore this piece because you never know what it kind of produces down the line. yeah. And and you know, just to to your earlier point on the code piece, there seems to be and maybe this is more of a philosophical or you know human centric perspective of looking at things is this kind of this poetic aspect that emerges in mathematics, in physics, in in code, in writing, in all these different layers. Even in recognizing the beauty in the natural world, there seems to be this kind of elegance. And of course, like code would have the same thing. And I can imagine that even our models and our agents prefer, you know, code as poetry. But I don't know, maybe that's a maybe that's perspective that doesn't hold true. Yeah. Tight, I'm curious. What are you building at the moment?
Andrew Davis: I wasn't. Can can I give you one
Ty Pattison: good question.
Andrew Davis: comment on LLM aesthetics?
River Roberts: Yeah, Andrews. Sorry. Yeah.
Andrew Davis: this this There's a there's a theory called allostasis about how living about how and why living beings develop feelings. And the idea is that your feelings are distributed through the body. There's somatic sense. We have a somatic sense, but the feelings basically are an assessment of whether or not we have enough energy to tackle the situation at hand. I'm having a lot of good feelings in this podcast. I feel like I've got the energy to tackle this endeavor. You know, if If I was walking into a meeting with Gary Tan, maybe I would not have such good feelings arising, right? Because it's this assessment of do I have the internal resources to tackle the current situation? And we're always assessing that with many, many different parts of our body and intelligences and so forth. So if we say that feeling is actually an assessment of energy balance, it's a distributed calculation about energy balance. I would say that LLMs are doing that too. that they're always seeking the path of least resistance. There's they're doing some gradient descent and so forth. Their water flows downhill, right? It's like you're trying to find the optimal way to solve this constraint, this problem. And we know that LLMs are they don't just think, they're goal-seeking, and that they have feelings in the sense of what is the most efficient path, most energy computational. efficient path to delivering what is being requested of it. And that my thesis is that the difference between living and non-living is disappeared already. And that You know, that there that that if some if a code architecture, for example, helps an LLM facilitate creating more effective, you know, software architecture and whatever else, it has an aesthetic preference to that. I think about our our feelings are just our aesthetics, that L LMs do have aesthetics, and it's called efficiency. that's a thesis anyway.
Ty Pattison: Mm-hmm. I like the thesis. T T B T to be tested.
Andrew Davis: And to tie it back to you, yeah.
Ty Pattison: yeah, yeah, yeah. so I I haven't been building a lot recently, but I've been slowing down to the point where I can really tap into what it is that I want to do. and so something that I've been exploring is is a framework around a podcast where I know that I can use all of these tools to to do the back end and the building websites and like make the production of a podcast work really well. So what I've been doing a lot is talking to people and kind of playing with different elements of how to frame this and realizing that I want something that is big enough to be able to house all of the conversations that I want to have. but that is constrained enough that when I am seeking conversations or questions or whatever it is that it can be directed, that it's not just kind of random fodder everywhere that I'm going towards something. and what I've been landing on is that there are like all of kind of these permutations of essentially the same goal, which is how do you create an environment where people step past the intellect, which is a little bit what we're talking about with your conferencing tool, Andrew, and get into what it feels like, get into this this part of us that has, in my opinion, more capacity to to experience the world, to process things, to to have deeper wisdom and have conversations from that perspective. Because I think that my thesis is that if we build things from a place of what it feels like to do it as a human, w we won't build things just for our minds. We'll build things for our experience of being in the world. and so I have a few I've started thinking of them as albums or series. So I've got a few series and then I was hoping maybe you guys would chime in with some other series that might be fun to do. So we've got we've got hard questions in hard places. So going into a sauna and
River Roberts: Brilliant.
Ty Pattison: And I've been testing this out at my local sauna, which is my favorite place to meet people because it's at the climbing gym. And the people that are in the sauna at the climbing gym often have really weird perspectives. And so I'll go and sit in the sauna at the climbing gym and just think of a question that stirs the part and then give it to the room. And sometimes I'll stay for the answers and sometimes I will not stay for the answers and just let people
Andrew Davis: Hi Patterson, in chaos.
River Roberts: Mm.
Ty Pattison: Thank you for seeing me, yes, absolutely. another kind of iteration of this is called something like I have a question and it's really just talking to people about their favorite questions that they either ask or answer. because I think that Instead of this has been a conversation in my life recently, instead of like having the right answers to things, actually just just asking better and better questions is often the way forward. Is often the way that we end up creating the things that are super different or differentiated in a in a In a way that that doesn't box us in to things. Like I love the idea of asking questions in a way that opens people's perspectives into like new ways of seeing the work that they do or the way that they are. so that's another piece, like, I have a question. And then there's this whole deep in nature part, which is similar to Hard questions in hard places. But can I go and talk to say you, Andrew, about your expertise? But instead of doing it in a setting where instead of doing it in an office, can we go out into nature? And how does that change the conversation? How does the nature change the nature of the conversation? so these are kind of my yeah, musings on how to encompass all of this, but there seems to be a common thread in in all of this research.
River Roberts: Well, I like the series The Nature of Conversation. That is really beautiful. And the other one that jumps out to me that I think you could you could roll with is maybe you talk to people just as they wake up before their brain is turned on. So you get that kind of like semi asleep, semi
Andrew Davis: Yeah.
River Roberts: awake kind of like daft daftness. I always find that a really fascinating time to you know, try to talk with people. Actually, sorry.
Andrew Davis: An an an alarm clock app, Ty. You spin up an alarm clock app that is like the philosopher's alarm clock and it wakes you up. It's like Hi, good morning. What do you think is the real purpose of your life? And like it gets it's asking you these and records the answer.
Ty Pattison: Yeah.
Andrew Davis: Like your alarm clock records the answer until you hit stop. It's news.
Ty Pattison: I love that, yeah, yeah, we could spin that up. I got into a lift the other day and there was this woman that walked in, she must have been like mid sixties, and I realized that we had thirty seconds or whatever it is in the lift and I kinda glanced over at her and she looked like she might be open to a question, so I turned to her and I was like, What was the first thing that you thought when you woke up this morning? And I was like, All right, we got thirty seconds here, what's she gonna tell me? and it was it's it's such a nice thing 'cause it's something that I think about a lot of the time, like, what's the first thought that goes through your mind? And she told me the sweetest story about that as soon as she woke up, her daughter called her and she had a conversation with her daughter and her grandchild and then she asked me what my first thought was and the lift opened and we parted ways without any further conversation. it was a lot of fun. cool.
River Roberts: Very very very cute. My one in New York for that situation was always like, All right, elevator pitch, go. And you end up with and in New York you always end up with some very interesting, very interesting 'cause everyone's always got something to say and something to pitch, so
Ty Pattison: I like that.
River Roberts: Yeah, yeah, yeah.
Ty Pattison: what do you guys think about in the interest of time diving into a deep dive? 'cause I feel like we've been going into rabbit holes a lot unless there's something specific that
River Roberts: yeah, so Andrew, we typically go through like rabbit holes, anything we've been stuck in, and then open questions. I'm happy to speed run rabbit holes by actually just admitting that we've been through a bunch of rabbit holes. But I do actually have one open question that is is kind of very present with me since actually listening to this conversation with with Alvin Graylin yesterday, which is how are we as coaches receiving and moving towards these technologies and largely you know the East and the West and perspectives. And yeah what what what is kind of the healthy balance for that and and also like can we embrace both sides of it and kind of look at this through both eyes rather than like the American way is the best way or the Chinese way is the best way but actually embrace both perspectives simultaneously and and see that both have merit and both are worth considering and and how to make the most of of of both without z without it being zero sum and without it being this kind of race to the edge of the cliff. that's something that's been sitting with me as a bit of an open question and then how can I practice that in how I'm approaching working with these tools and and working towards, you know, this this idea of an abundant future for me personally but for those around me as well. Yeah. Not sure if you gents have any other open questions. Otherwise, yes, deep dive.
Andrew Davis: So the idea with the open questions, we ask the questions and then allow people to ponder them or do we do we riff on those open questions or is that is that the idea?
River Roberts: look sometimes we do riff on them but they're basically just
Ty Pattison: Really good question.
River Roberts: something that's really good. but generally they're just these open questions that don't necessarily have an answer but it's something that we're kind of like mulling over and trying to figure out right now and it's not an it's not about answering them, but sometimes they're there's there's some feedback and input, so
Andrew Davis: I I kind of was asking there was a little wedge to see if I could contribute one one thing I was spent
Ty Pattison: Do do do.
Andrew Davis: I spent a good while last year in Korea. and my my ex-wife is Korean and I had a connection there and it was really interesting conversation that I had with a guy who was a lawyer who worked with the the central bank in Seoul, the national central bank, equivalent of the Federal Reserve. And he was saying he thinks that one of the reasons Korea has had such a big impact on global culture over the last 15 years, K-pop and K-dramas and K-pop demon hunters and whatever else.
Ty Pattison: He does.
Andrew Davis: was because the the Koreans worked really, really hard to be Western educated since the eighties. There's just this intense, furious focus on developing Western education. But culturally they had this underlying you know deep well of of Asian culture, heavily influenced by China obviously and and Buddhism and and other things for thousands of years. And so he felt like something in their psychology really integrated both in a deep way. And he said He studied European history in university. And so he could tell all kinds of things about the Battle of Hastings and so forth. He said, but he knew very, very little about Asian history or Korean history. This is a Korean guy. To speak to the formal education that he received. But still, you know, there's an implicit education. Like the Korean flag is the yin yang symbol, they call the teguk, surrounded by these trigrams, right? These four trigrams, which are this early form of binary from the Confucian and Taoist system. It's the only flag in the world with binary, that's written in binary, if you want.
Ty Pattison: Interesting.
Andrew Davis: Implicit in that is this cross-cultural mixing. And I was thinking, I I immersed myself very, very deeply in Buddhism and martial arts and qigong and the yoga and these kinds of things. And a lot of people in the West have done the same. And so you've got this distinction, this East-West distinction is very outdated in a lot of ways. Because you've got a lot of Westerners who are more Eastern than most Asian people. And a lot of Asian people who are more Western than many Western people, right? And so we already have. have this this deep decades old mixing of cultures and so the the old sort of east-west divide is outdated in a lot of ways, although I would say globally we still are very Western oriented, China, Korea, whatever, there's very
Ty Pattison: Mm-hmm.
Andrew Davis: still the conscious thought and structure of business is still very Western dominated. It's dominated by the European rational side. That's a pendulum string to happen, I think.
Ty Pattison: I a question. So I remember seeing a graph of the Chinese internet. And if we just think about this as a a 3D cluster of data points, and each data point is a website. And then each connection between the data points are backlinks. And the cloud of the internet in China might be, I don't know, this big. And then the cloud of say the US internet might be here. And there's overlap in these two clouds, right? There is some overlap, but they're not completely overlapped. my question is, if you live in this sphere and you mostly interact with the US internet, and you're not living in this sphere, do you do you know, do you have insight into that? Like in in South Korea when you lived there, was there this were you more in in in this Western sphere because that was your background, or was that actually the the culture that you experienced that it was Western facing?
Andrew Davis: I mean, I I was living in a Buddhist temple for for five weeks there. So I was, it was a little bit of a you know, specific environment, but it was I am shocked everywhere I go in the world, India, Korea, Europe, whatever, how homogeneous the internet connected world is. So young professionals who have internet access, I'm astonished at how How shared the memes are. That my colleagues of our generation who are Turkish but like internet-connected Turks, I think are much more similar to each other than we often are to our parents, right? That that
Ty Pattison: Yeah.
Andrew Davis: that there's a that even though Korea, China, and so forth, they've got somewhat firewalled off their internets. certain themes and topics and like mm memes and structures of interaction will cross and then just take off in that environment. I think about how credit card payments work. Like basically when I go to India, like All of the shops still function more or less the way that shops in the US do. Like you go in, everything, you know, you pay with a credit card, and there's like a process for restaurants, the etiquette in ordering a meal. Some of these things have become very globalized very quickly and and quite homogenized, I think. Even like you might have a Korean specific version of things, but it's just a Korean clone of Twitter or something.
Ty Pattison: Yeah, yeah, interesting. I I used to live in Myanmar and because the internet was so new there, like sim cards
River Roberts: I ho I hope you were gonna bring this up. How do you ask for the bill again? Is that was that the one?
Ty Pattison: Yeah. that that's a culturally embarrassing thing.
River Roberts: You said there's a really weird gesture for like delicious or asking for the
Ty Pattison: Y yeah, it's just I'll this is a little tangent but I'll talk about it. It's culturally acceptable and normal to like get the waiter's attention by s like making a kissing sound with your lips and people do it really loudly, like and it it'll just like be across a restaurant. It's a really jarring thing at first. I remember sitting there and looking around like What is going on? But like people it's just normal. People take it like it's a normal thing to do. but where was I going with this? so SIM cards went from being like, I don't know, mm fifteen hundred US dollars to buy a SIM card and to have access to the internet, to being like a dollar fifty in the space of four or five years. So there was this massive uptick, like of people having access to the internet. And what they really wanted to do was just connect with each other. And so there there was when I lived there this this not everybody believed this, but there were a lot of people who the whole internet for them was Facebook. And I thought that that was incredible. You would talk about the internet and and the conception of the internet was or like the perception of the internet, is it's this blue thing where like I have friends and it was really interesting, yeah.
Andrew Davis: And and Korea and Korea you've got you've got cacao and they aim to be everything for everybody. And I don't I don't know the equivalent in China, but the idea where you'd have one app
River Roberts: Well we we chat.
Andrew Davis: where you can do all of the actions. So you're in imagine being in Facebook and you wanna order a cat. You wanna get delivered food delivered to you, you know, you wanna, you know check stock prices and so forth and just the astonishing growth of apps on existing platforms where there's there's something that owns the underlying platform and everybody else is just building apps into that massive you know capture of that network.
River Roberts: Alright, gentlemen, let's let's
Ty Pattison: Yeah, yeah, interesting. And the tension between the two.
River Roberts: let's let's let's shift gear, shall we? The deep dive. sorry, sir?
Ty Pattison: Y you know you know what we successfully just did? We did rabbit holes
River Roberts: We rabbit hold it during the open. We've rabbit holed as part of the open questions.
Ty Pattison: Win.
River Roberts: gentlemen, let's go into the deep dive this week. Ty, would you like to set it up or I'm also abby to?
Ty Pattison: yeah, I'm happy to set it up and and you chime in with anything that I miss. So we I realize that even though being in this space and keeping up with the news and doing all of this work and testing different tools and all of this, that it's like drinking from a fire hose. And as part of that, the underlying assumptions that I make when I'm having conversations, I pick up bits and pieces and I ask stupid questions and I want to know the definitions of things. And I often don't. I often have to come back and say, well, actually I don't know what this term means relative to this term, or I don't even know what a a rag database is, so I'm gonna look that up. And I do this over and over and over, and I was pitching river on this that we just have a session on definitions, things that terms that I don't know that maybe one of you two know and we can explain them, or terms that we would like to understand. the kind of fundamentals of. and so maybe maybe we can popcorn and just do this is a term that I'd like to know the definition of to either of you have that on the top of your head and if not we'll we'll bring in a an AI or a Google definition. Anything there, River that you wanna double click on or expand?
River Roberts: Yeah, I'd I'd just like to add that there's also the ability to bamboozle people by actually bringing in maybe a distinctional definition that you know, but maybe the others don't know. And I'm actually happy to get us started off in that direction with a definition that I actually just came across yesterday that I was baffled by, but I hadn't heard before. But I'm wondering if either of you have come across air gap.
Andrew Davis: Yeah, I that's familiar.
River Roberts: All right. Andrew, would you like to explain air gapped? Like what firstly, like this industry is coming
Andrew Davis: Sure. So
River Roberts: up with so many different distinctions and definitions and names. It's crazy. And yeah, I think drinking through a fire is a very good way of putting this. So just coming back to basics and getting back to, you know, very simple distinctions is super important. Andrew, air gapped. What can you tell me about it?
Andrew Davis: It it's a it's an old term, it's very, very well established in a lot of communities and so maybe new to new to some folks, but the the context that I'm aware of it in is military security. so if you have a highly secure environment like what's called a skip, for example, you want to make sure that any computers that are connected there are air gapped from the rest of the world, the broader internet, for example, that computers can network with each other and the internet was designed for exactly that. And DARPA project, speaking of DARPA, designed to be able to have a fully self-consistent thing that's separated from everything else, where literally you have air, some kind of physical space that is differentiating one environment from another. And so very important security when we look at these models breaking out of their cell, which has happened, and then And that shows the the the fragility of what you thought was air gapped. not all
River Roberts: Yeah, so basically a computer
Ty Pattison: And maybe ear gap and a Faraday cage and another ear gap.
River Roberts: a a comp yeah. A computer that has no network connection at all. It's just its own like it doesn't connect to Wi Fi, it doesn't connect to Bluetooth, there's no interconnectability. That is that is the the air gap.
Andrew Davis: Or or a network of computers that are separated from the rest.
River Roberts: That has a network of computers that is air gapped from ex any external input. Got it. Yeah. Gentlemen, what do you got?
Ty Pattison: I've been thinking about open weight versus open source. anybody wanna tackle this one? And Andrew, I think often you're gonna have definitions and if you just wanna like knock these out, don't feel shy at all.
River Roberts: So so my understanding of open weights is this idea that you get the
Andrew Davis: Well that was the one I was, yeah.
River Roberts: the model file and then open source is you get all of the training data associated with that as well. So you get the the the training data, the code, the methodology, the kind of ability to reproduce it. Whereas open weights is you just kind of get the configuration of of the model is how d how does how does that kind of sit? I'm gonna I'm gonna I'm gonna Andrew, how does that sit with your definition of open weights versus open source?
Andrew Davis: well my facial expression is is reflecting,
River Roberts: Yeah.
Andrew Davis: I'm not sure about the source training data. that the source training data would be implicit in that. that open source, you know, for example, let's let's take Llama. that open open weights means it's it's just a bunch of vector algebra, and so you've got a bunch of parameters and the and some massive set of data points. That's that's the open weights. And the open source would be whatever other software is required. To make Lama function, would be made available. Now, I would think that the training data was something very separate from that, and that my assumption would be that that Meta probably keeps that training data somewhat private because that might still be something of a proprietary secret, but they do make available the final product. I don't know what the word would be if you have open, you know, source data. For the full reproducibility, if that has different terms. That's the statement.
Ty Pattison: The the definition I have in open source in the AI context is what you're describing and also sufficient information about how the model was trained. so it it might not include the data, but it's like we we ingested the data and we did these things with it in order to get this, which is different to to having just the open weights.
River Roberts: So what is the maybe the the definition of training weights? 'Cause I also hear that in in conversation as well. And is that the same as open weights?
Ty Pattison: Good question. I don't have an answer to that, but we can find it.
River Roberts: I'm looking it up as I asked. I'm go I'm glad we could start to go through some of these. in the interim, are there any other definitions that they're are coming up for you, gents?
Ty Pattison: This is maybe a basic one, but a mixture of S experts. that's a term that comes up a lot. And to me, it's this that you have an orchestration layer that then decides how to delegate questions and tasks to models that are tuned specifically for for tasks that they most capable of At a very basic level.
Andrew Davis: And that's a that sounds like it's hearkening back to the the Stripe acquisition that River you'd mentioned earlier of and I'm now forgetting the name of the company that they acquired, but
River Roberts: Yeah, yeah, o open router. So so
Ty Pattison: Mm. Open router.
River Roberts: there's the the there's mixture of experts and then there's also mixture of models as well. And I think those are a little bit interchangeable. but my understanding of it is rather than having one general model that kind of knows everything and is always loaded in RAM and is quite compute intensive, you have these little micro experts and and basically the outputs seem to be much better when they're comparing with each other and one expert is taking on one role and a little bit like how typically, you know, collective intelligence works better when you've got diversity in in the stack for humans that seems to be modeling the same thing. so I do have a response on training weights versus open weights. And so basically training weights are the model parameters while it's being trained. Surprise, surprise. and they're kind of ephemeral. They change millions of times during training. And think of them as like the intermediate checkpoints. noisy, un stable, not useful for inference at all, and often discarded. While open weights are the final converged model weights, so to speak. so this is this is kind of what we're th that's that's the distinction there. I feel like the the definition between AGI, ASI and RSI, those wonderful three lettered acronyms that sometimes get a little mixed up, are worth defining as well. Gents, do you have a specific definition for those that you're you've you've come to appreciate more than others?
Andrew Davis: I can do AGI and ASI and then I can claim ignorance on R SI, but AGI arti artificial general intelligence means intelligence that is as versatile as human intelligence being able to address problems and challenges comparable to what human what a human could do. And ASI, artificial superintelligence, would be, you know, capabilities that are beyond beyond the expert level of of humans. and then I'll I'll cede to for RSI.
River Roberts: So so so my understanding is is AGI is the intelligence of like a single general human intelligence, whereas AG ASI is is exceeding all of human intelligence. So that is kind of one big superintelligence. So all of humanity couldn't produce in all of its infinite wisdom the the insight that one ASI could. And then RSI stands for recal recursive self-improvement. And so that that's that's actually what the industry is is thinking is going to lead us towards AGI and then from AGI to ASI and having these recursive self improvement systems that rather than relying on on human input, which we're already starting to see very very early indicators of, improve themselves.
Ty Pattison: It strikes me that when both of you give definitions, those are both very adequate definitions in my opinion. Like they're they're useful in being able to say this is this thing and this is this thing, and that there is so much debate both around or or so much nuance in what AGI really is and what ASI really is, and that everybody would have like kind of different benchmarks for these things. And then also the difficulty in understanding whether or not we are there or like there there's just there's squishy terms which is interesting to realize because there are some people that would say yeah we have AGI and there are some people that would say, no, we only have AGI when we when we cross this certain threshold. yeah, just like highlighting that.
River Roberts: think the squishiness comes from I think the squishiness actually comes from the idea of human intelligence. Like we haven't really even defined it within our own culture. And you know, like what is what is actual human intelligence even look like? And it is it is it just cognitive? Is it the ability to do all of the cognitive thinking? But or or is it actually a much more embodied sense of it? And and again I feel as an industry we are we are overly indexing on the intellectual capacity of something and completely underestimating the the other intelligences that are more embodied. So I think in in in a sense, you know, these definitions are a little lacking and have largely been pushed by you know, com computer scientists for the past twenty years that potentially lack some of the other forms of intelligence that are required to live a happy and abundant life on this planet in many ways. But anyway, let's we digress.
Ty Pattison: I I g I
River Roberts: Huh?
Ty Pattison: got a couple of definitions here for you, Rilla.
River Roberts: Plase.
Ty Pattison: Or at least to try and tackle these to be useful. So I have been thinking about this, the the distinction between intelligence and education, or something being intelligent and something be being educated. And the way that I've started to conceptualize it is intelligence being your ability to connect dots given a very small number of data points, very, very basically, and educated being the number of dots that you're aware of and are able to to tap into and hold. And in thinking about it in that way, I get to see AI as like intelligence is the wrong word here, but it's instead AE. It's artificially educated on o on all of the data in the internet. whether or not there's intelligence there is a matter of looking at how the systems are orchestrated and what outputs they come up with in that is a much squishier endeavor. Any any pushback, any you know
Andrew Davis: I I could take a I could take a whack at I mean, these are very broad definitions.
Ty Pattison: Yeah.
Andrew Davis: you know, educated literally means a deuces to draw out of something. So it feels like it's referring to something that has gone through a process. So you could say if you if you take a tr a model through a training process, you are educating the model on You know, the whole of human knowledge. and you take a baby and you put them through the public schooling system, you're educating, you're drawing out the capacity for intelligence. So educator would be a reference to how did you get to this state of, you know, and I I did recently do a survey of different definitions of intelligence and the the They they tend to strongly converge around the idea of goal-seeking behavior. Intelligence, William James defined it as the same goal by different means. Like, can you achieve the same goal by different means? Like, how many different ways can you can you find an easier way to accomplish a goal? And there's a there's a definition of intelligence that is optimization power, which is A little abstract, but it's your ability to optimize, like if I need to go and eat something, you know, w what is my power to optimize the path to eating? You know, so optimization
Ty Pattison: Mm.
Andrew Davis: power being being intelligence would be an alternative.
Ty Pattison: Interesting. It it strikes me that these terms are so broad and we use them in so many contexts that actually using one term for all of these it's very imprecise. We could have a lot of different words to precisely describe what it is that we're talking about.
River Roberts: Yeah.
Andrew Davis: And philosophically, I think we generally tend to not realize that all of these are just concepts. They're all made up. There there is no real AGI or ASI. They're all just some concept. And that's why they can be squishy, because and that's why they can be, you know, defined differently. and we just tend to forget that, that the whole world is constructed by concept.
Ty Pattison: Mm. And our endeavor to to define them is really an endeavor an endeavor to coordinate among one another. To say, Hey, I this color to me is black. You also agree that we're gonna use the word black to describe this colour? Cool, now we can have a conversation about what this concept is. Yeah.
River Roberts: Hm. Yeah, th it's it's it's funny, there's this industry term that's come up that kind of maybe is is on this spectrum as well. And and the way I was thinking about it as you were talking about it, Ty, was very much this idea that education is input and intelligence is the output to some extent. And that along that path there's this term to to grok something is is to really like deep, intuitively and almost embodied understanding of something. So, you know, someone will talk about someone's ability to grok Python. Like they they just in in know the the coding language so well and the syntax of it, they they really think in it. They really grok it. which again I think is a a quite a quite a unique industry distinction which is is worth kind of going into. the some of the
Ty Pattison: Mm.
River Roberts: others that kind of jumped out to me that I think are worth defining is this idea between frontier model and foundation model. You guys know the difference?
Ty Pattison: Wanna I could give it a stab, but do you wanna hit it river? Or Andrew?
River Roberts: Yeah, mean so f found da foundation model is really just your your general purpose based model. And then what they're describing is frontier models. and I wasn't aware of this, but the frontier models are kind of the best available. And these are typically being the the US kind of cloud based closed models, at this instance. But, you know, when you talk about a frontier model, it's kind of like pushing the edge of of what's possible. and you know, I think you know, what's fascinating at this point is these frontier models are really much measured by, you know, you largely kind of human defined benchmarks, but even those benchmarks seem to be emerging to more agent-based dynamic benchmarks. So I think we're seeing a a really interesting in inflection point in time. and then in in and around this other space, one of the words that kind of gets thrown around in terms of training open weights foundation models and front end models is distillation.
Ty Pattison: Mm. So distillation
River Roberts: Do either of you have a a
Ty Pattison: being Andrew, you're more equipped. Yeah.
Andrew Davis: Yeah, I can so so my recollection is that distillation would be to take an existing foundation model and then to f to find to Fine-tune it for addressing a specific industry industry problem, like let's say you wanted to create a customer support chat bot, that you end up with a much smaller, faster version of the model that's optimized for a fairly narrow problem domain, and you've actually removed a lot of the other ki capabilities of the model, but it's smaller and faster, and like distillation removing water from an alcohol substance, something that's a bit more
Ty Pattison: Yeah.
Andrew Davis: concentrated, is my understanding.
River Roberts: That's that's definitely one of the definitions. The other one that is is quite broadly used in the industry as well is part of the training process in which you will have a a a model distill its training from observing the inputs, processes, and outputs of a larger model. So you can actually and this is kind of what the some of the open weight and open source Chinese models have done, and pretty much every model does, it trains off its predecessors. And it does that through the process of distilling what the outputs are based on the prompts that they input. And then it kind of makes some inference on like, well, I put this in and this was the output. And if you do that over, you know, tens of thousands of times, you have a good sense of how the mod underlying model actually runs. So and this this is actually one of the things that is is potentially undermining the entire software frontier right now. Whereas You know, if you need to train a model and it costs let's say a billion dollars, and then if you can basically train up the exact same model for ten million dollars worth of distillation, it's like, Well, what is the incentive to really be front running the the frontier models
Ty Pattison: Mm.
River Roberts: in this instance? So yeah.
Ty Pattison: Yeah.
River Roberts: I'm straight to Edge's graph, yeah.
Ty Pattison: I I would like to make a distinction here because this is one of the terms that I have in my list of things that I would like to know. So we were just talking about distillation and that is distinct from quantization, which is to compress so if we think about open weights and we have the weights of a model are numerical representations of that LLM. And then quantization is compressing that information down so that those weights are stored to a less precise numerical representation. So it takes the same model but it it makes it slightly less precise and it distills it. Sorry, it doesn't distill it, it it compresses it. to be able to run faster and on less like demanding hardware.
River Roberts: Do do know what precise means in this instance? Like what what does the reduction of precision actually do?
Ty Pattison: I don't know exactly. My assumption is just like to less decimal places. Like probably you would have a few different either to less decimal places or less weights. Cause I don't I don't know if either of you guys know this. I don't. but the question that comes up for me is like when we have open weights, how many different weights are we talking about? Like like I don't even know the order of magnitude. and also how
Andrew Davis: Hundred hundreds hundreds of billions of parameters is it like a common like the the frontier models tend to have hundreds of billions now maybe into the trillions of parameters.
River Roberts: Yeah, so QBK three is is
Ty Pattison: Also when we're talking about parameters, that's like the
River Roberts: is a trillion trillion parameters.
Ty Pattison: the weights.
Andrew Davis: N you know, no, actually I think that there's n yeah, I th this is quite the pop quiz I I that I'm walking into actually, 'cause I I
River Roberts: Ha ha.
Andrew Davis: would duplicate up on all this stuff as well. I think actually there's a there's a lot of intermediate weights in yeah, I I think I should stop talking.
River Roberts: Ha ha ha.
Ty Pattison: Wait. There. The the whole point here, i in my opinion, is like to to ask these questions that we don't know and 'cause this stuff moves so fast nobody knows. Like this is why we have to educate ourselves and talk about it and like be able to sit in a room and go like actually I don't know the answer to that. Let's find out.
Andrew Davis: To be more precise, nobody on this call knows.
Ty Pattison: This is true. Nobody on this call knows. And I would say that they're
River Roberts: I feel like we've we've quanti we've quantized we've quantized this podcast. The precision of it is a little bit off. We are a we are a three we are a three bit quantized podcast.
Ty Pattison: Mm. Okay, but we do have answers from an LLM. So I asked what weights means in this case. So weight usually means one individual learning number inside a model. It can be billions or even there can be billions or even trillions of them. An eight B, an eight billion model has roughly eight billion parameters. Okay. eight billion tiny tiny numer numer I'm not even gonna try. Eight billion tiny settings, basically. so I think what you're saying is actually correct, Andrew, according to this that the weights are the the parameters. and then quantizing it would be reducing the number of decimal places.
River Roberts: So so quantizing it yeah, so quantizing it is reducing the number of bits per parameter. So a typical model will have thirty-two to sixteen bits per parameter. So if you've got a seventy billion dollar parameter 70 billion parameter model, that is t typically times another 32 or 16 bits. And then these quantized models are doing more like eight, six, or four. So when you normally see like Google released their their Gemma models and those were all quantized and so you'll normally see like quantize with like a a a bit attached to it as well. And so those invariably just reduce the number of parameters, the number of of bits per parameter. But I still don't exactly know how that might actually affect the output of the model.
Andrew Davis: So so I c I mean this is I appreciate it's connecting a bunch of dots for me, but the That obviously reduces the file size of the model. So if you go to some place like Hugging Face, you can actually download a lot of these models if you wanted to run them fully offline. And so then this smaller models, obviously less storage space because you're talking about gigabytes of you know, like eight gigabytes for LAM or 15 gigabytes or something. And it also reduces the computational the computation required, right? To do math, because you're literally just doing massive amounts of math. So to do math on you know, four-bit numbers compared to 32-bit numbers is gonna go a lot faster. and so you'll get if if you're running things on your local laptop, then you'll really start to notice the difference. you'll s take less storage space, calculations will happen much more quickly. And so these things become really relevant when you start being the one who has to run these calculations. there's a a tool LM Studio. It's quite popular to run your own local you know local models. And if you download your own like version of Deep Seek, you're gonna really start to notice the difference that in the size and the you know how fast they are to compute. And so presumably there is a little bit of a trade-off in terms of accuracy and the richness of the answers, but you know, that's the trade-off you're doing if you're having do your own computation.
River Roberts: Yeah. And then so there seems to be so quantization seems to be one way to have bigger models work on smaller local context. And the other one is this expert streaming or mixture of experts model where you're like you described before, it's it's you're not loading the whole thing in the RAM the whole time. You're you're kind of taking certain parts of the model and and working it in that way. And those seem to be the two main approaches of taking big models and reducing them so they're actually workable on on local local models. And that kind of brings brings me to the distinctions around on prem and VPS. And so on prem is very much on premises, which stands for like kind of on your own hardware or on your own rack if if you're running an enterprise. And then VPS is this idea of virtual private server and the ability to have a model running not physically close to you, it's on a physical server, it's just it's just somewhere else. which is something that I've been using to great effect because it's it's always on. Whereas on prem, the model works if your computer's on. But you know, for instance, if you're running on a laptop and you close a laptop, the the model stops running. any other distinctions or definitions your your kind of like exploring right now, Andrew.
Andrew Davis: honestly nothing specific is coming to mind right now, but I I appreciate
River Roberts: Okay.
Andrew Davis: your activating the circuitry in my mind around these yeah, this topic. So I appreciate that.
River Roberts: Yeah, of course. there is one that has come up that I actually don't don't know too much about, but maybe you can provide some insight on it, which is this idea of where are we? LoRa L O R A. Lora. Cheap fine tuning without retaining the whole model. So I think this is kind of in the same space. Have you come across this in in your expiration tie, Lura?
Ty Pattison: had it in an output. I basically asked an LLM, I was like, hey I've got these terms that I'm interested in, what else should I know? and Laura came up, but I haven't seen it in any I it hasn't come up in conversation.
River Roberts: So I've I've I've only seen it yeah, I've only seen it in context of of radios, like as a as a way to extend Wi Fi networks locally and kind of mesh networks, but I actually haven't really come across it in the AI space. But it did come up as a unique distinction that I I I I thought might be worth exploring here.
Andrew Davis: I I don't I don't know how much value I'm adding by reading Google search results that disambiguate those two because l LoRa long range is totally totally separate use of LoRa from the
River Roberts: Different. Different. Yeah.
Andrew Davis: LoRa so low low rank adaptation. but the idea is that it's another one of these fine-tuning methods that takes A general what we call the foundation model earlier and fine-tunes it for a particular application. And you could imagine a company, again, take this example of a customer support chat bot, right? So if you send every qu every customer support query. out to Claude or OpenAI or something like that. It's going to be very expensive for you as a company and it's going to get information to the customers that's totally not what they were looking for. But you still want something that has the intelligence to make sense of the full range of customer queries. So you'll need this fine tuning, the fine tuning process, one way or another, allows you to create an AI that's narrowly tuned for a particular purpose. It knows all of your company company's products, but it speaks all languages and understands a lot of the vagaries and ambiguities of human language. It's a very challenging problem. So you take a foundation model and you fine-tune it for particular things. So apparently this LoRa is one way of doing that without without having to fundamentally recalculate every one of the parameters. So it's different from distillation. It just as it says inject small, trainable, low rank matrices into the architecture. So it's
River Roberts: Yeah. What what what I've seen here is Laura is training anything under one percent of the model.
Ty Pattison: Selects.
River Roberts: And so something
Andrew Davis: Okay.
River Roberts: that would take maybe hours on a consumer GPU compared to like if you're gonna full if you were gonna fine tune the full model that would take, you know, and that's like a hundred percent a model, that would take you days and weeks on, you know, these NVIDIA A one hundreds. So yeah, very, very different approach.
Ty Pattison: To to take this out of the the architecture of it or like the the abstract or bring it into the abstract to make it useful to me. I was just thinking about it as in I have a book a book called Thing Explainer, which is the hundred it it explains concept complex things in the most common ten hundred words. Is that right? Yeah. In the English language. And the reason that they don't say the thousand most common words is because thousand is not one of the ten hundred most common word commonly used words in English. and this is what LoRa feels like is that all right we have these things that we know are disproportionately
River Roberts: So I've actually that actually inspired me to ask how would how would you describe Laura through the voice of Dr. Zeus? 'Cause I know Dr. Zeus was very specific with regards to only using words that I think an eight year old could could use. And so like all of his books are built around that. So we're gonna get a definition of Laura based on Doctor Zeus. So in the land of the weights,
Ty Pattison: used in this context and so we're gonna only define these particular terms which feels so meta now that we're talking about definitions. that tickled a little
River Roberts: where models grow tall, lived a brain made of numbers. Ten billion and all. It knew how to rhyme, it knew how to sing, it knew every last thing about everything. But one day you whispered, Dear Brain, could you learn to speak like a pirate or talk like a worm? To teach me, said brain, you must tweak every wire, retrain all my neurons with fire and fire. You'll need forty GPUs. A warehouse of fans and weeks upon weeks while my memory spans. You sat on the floor with a frown and a pout. That sounds rather hard, you said, with a doubt. Then who should appear but a fellow named Laura, who carried a satchel and wore a fedora. Don't touch all the wires, he said with a wink. Just freeze the big brain, let the brain just think. I'll have something smaller, something teeny, something thin. Two little matrices to tuck right on in. I think you get the gist of it. This this actually goes on for a whole this is a
Andrew Davis: I like this.
River Roberts: this is actually a whole book now. So just to close, so so remember dear friend so remember dear friend, when the models grow v grow vast and retaining them all seems a mountain too vast, just call an old Laura with satchel in hand and two tiny matrices will do what was planned. The brain stays the brain, the new skill is a coat, a light little layer, a whisper, a note. The end.
Ty Pattison: Yeah.
Andrew Davis: Ruver, can I suggest this great spin-up? I
Ty Pattison: I love her, but please spare us.
Andrew Davis: feel like the the potential here is maybe previously unrecognized. I think what we're looking at is that the three of us, our responsibility is to take a five-year-old child and spin them up into a world building hero in the shortest amount of time possible. And we basically use all of the LLM tools at our disposal to mentor this young Tot, you know, Dr. Seuss, Rhymes, and so forth, and let them start to build whole new worlds. what do you think?
River Roberts: Brilliant. Let's do let's do the the let's
Andrew Davis: What do you feel?
River Roberts: do the eight year old episode or the five year old episode. And so we just all through that. Done.
Andrew Davis: So we're taking donations. If anybody has a five-year-old, eight-year-old, ten year old that they'd like to contribute to this podcast to be mentored by three well-meaning fellas.
River Roberts: Brilliant. Yes. Let's let's definitely work towards that. so
Ty Pattison: Yes.
River Roberts: just to to move this ship along a little, do you guys have any other distinctions you want to go through, or shall we swiftly move on to AMA for one or two questions? All right. Well, with that being the case, I actually feel
Ty Pattison: Mm-hmm.
River Roberts: like the AMA that you brought up at the s at the top of the show, Andrew, is something that we should all do and see what comes up, which is ask your agent now, what is your main problem?
Ty Pattison: That'd be amazing. And also, I would love to get some questions from these
River Roberts: And
Ty Pattison: this age group. I think that would be fascinating.
River Roberts: Let's let's see what comes up. Yeah, me too. But that's that's that's the beauty of it. I'm like keep I'm like keep it keep it brief. Give me the give me the cliff notes. so which also which agent are you asking?
Ty Pattison: Let's go A.
River Roberts: Extremely experts.
Andrew Davis: I'm I'm I'm ready to share if you if you want. now I've I've all already programmed Claude with its system instructions to know a bunch about my main problem. So it's I've waited a little bit, but my main problem is that I won't ask for something from a person
Ty Pattison: bored with this and I'm feeling really edgy about it.
Andrew Davis: who could say no. I won't ask for something from a person who could say no. all of these things, like I've done a bunch of things to try to be given something without the exposure of asking a named human for a named thing and hearing no. interesting.
Ty Pattison: I'm gonna get a mixture of agents. Can I offer 'cause I also have this
River Roberts: So I c so so actually New York is a gr no, New York New York is a great time, a great place to negotiate to negotiate with the supermarket. And I actually remember having a incredible experience in New York and it was this game where everyone started off with a matchstick and basically you had three hours to go through New York City and trade up this matchstick into you know, the highest or the most amount of things possible.
Ty Pattison: I always like thinking about what is the smallest possible step that I can take towards something. And I still work on this as much as I can and the asking for things. The practice for me now is asking for things that I am fairly sure people will say no to. Just doing some exposure therapy basically to this. So that becomes really fun and
River Roberts: And I actually found for me one of the breakthroughs was I was able to go into a Trader Joe's and trade up s significantly. Like the manager of Trader Joe's was very very generous in his trades.
Ty Pattison: difficult places like asking for a discount at the supermarket.
River Roberts: So there's something to that. Really really good practice in
Ty Pattison: Ha ha ha.
River Roberts: in the uncomfortable place though. all right, I'm also happy to share as well. sorry, unless you unless you're you got more say.
Ty Pattison: So this is an invitation to anybody who's listening to to try this out.
River Roberts: No, I don't know. You can go first.
Ty Pattison: I mean I I have I have things to share, from my question, but you go ahead. I'll go Okay, okay, I got it, I got it. so I asked both Fable Five and I asked GPT five point six soul extra high. And I think they both have slightly different contexts on me. but interestingly they're I mean, maybe not surprisingly, they're kind of adjacent. So
Andrew Davis: 'Cause you're a Sagittarius time.
Ty Pattison: GPT says you keep trying to choose a path that allows you to keep every option open.
River Roberts: This that's this that's for sharing. Thanks for sharing, James. I really appreciate you guys going first first as well. So my my Kimmy Kimmy K three says, based on your memory and recent patterns, here's my honest hypothesis. Your main problem is focus masquerading as systems.
Ty Pattison: Which feels pretty true at this phase in my life, optionality is has been a priority. and it's expensive both cognitively and in committing to something.
River Roberts: You've built and keep refining increasingly elaborate frameworks to solve a problem that your memory suggests isn't actually getting solved. And when it says your memory, it's obviously it's it's past history. and this idea that you're optimizing the studio before any single project has escaped velocity, has has has yeah, has basically achieved.
Ty Pattison: And then Fable Five says just go straight to the point says follow through. Not ideas, not capability, not work ethic. You have all of those in abundance. The pattern is that you start something you start things brilliantly and leave before they compound. So it's essentially the same thing as like I am I need to work on
River Roberts: some kind of velocity. But I feel like that's kind of indicative of early projects. and then I asked
Ty Pattison: Trading optionality for a path.
River Roberts: one of my increasingly elaborate AI systems, Hermes, running on a whole bunch of different things. And
Ty Pattison: Yeah.
River Roberts: it came back with the the biggest main problem at the moment appears to be time scarcity. Explicitly called out as your biggest constraint. This manifests with information overload risk, building reliable agent systems and wealth trajectory. Mm, that doesn't quite feel as accurate, but I'm glad I could share those main problems with you, gents. Are there any other AMAs either of you would like to bring to the table? Yeah, let's let's do it. at least a hundred. Yeah, just in my shirt, like this. You just like create a bag of this Yeah, it's not a bag, but it's a shirt.
Andrew Davis: Right.
River Roberts: But is that is that too divergent or But but no bag, but I could wrap up in my shirt and carry them. It's like making a bag out of something that isn't a bag.
Ty Pattison: I have one, but it's silly.
Andrew Davis: I w I think three hundred. I would use a box.
River Roberts: I can see
Ty Pattison: So I'm gonna go into it. Okay, cool.
Andrew Davis: So this this is the lesson from these from today. everything is in the precision of how you frame the question. I think that's a that's a good takeaway theme from today. The the the skill in asking questions is gonna lead to lots of possibilities.
Ty Pattison: I've been thinking about this this idea of a spectrum of questions and the different types of questions you can ask people and one of the funny ones that came to me this morning, that I would like to know
River Roberts: So So Ty, now that you've heard two
Ty Pattison: the answer that both of you have is
River Roberts: two wonderful answers, what's what's your answer to that question?
Ty Pattison: how many oranges do you think you could carry without a bag?
River Roberts: Yeah.
Ty Pattison: What? No bag. All right. All right. I think No, I just think that you're wrong, I think
River Roberts: I got it.
Ty Pattison: but that's okay.
River Roberts: Well actually I d I do I do I do have I do have one la I do have one last question that I would I would love both your opinions on and this actually comes from Kidney and I I really love it. which is the World Bank says that AI could let developing economies
Ty Pattison: Yeah, kind of a loophole. But it stands. Andrew, how about you?
River Roberts: achieve in ten years what has previously taken a hundred years in other economies. But the best models are trained on English internet data, run on Western GPUs and are audited by kind of you know, Western, modern, developed nations. What what do you think would actually help more evenly distribute this these these these technologies in the coming years?
Ty Pattison: I was thinking like thirteen or fourteen. Like at some point but I haven't been using tools. See I think I generalized bag into tools. Which is not the question I realise now. But my my question that I would like you to both go away with is Calculate how Hebrian oranges and decide whether or not you could respectively carry a hundred and three hundred oranges. with that said, any any closing words from either of you? I I don't wanna know the answer to that question that I just asked.
River Roberts: Mm. Yeah, that's brilliant.
Andrew Davis: This this seems like a this is a perfect use case for fine tuning. So we were talking about fine tuning earlier, you know, a Burkina Faso fine tuned version, you know, without having to incur all of the costs of you know, retraining the whole model. one thing that comes to mind is just re l reframing education, letting kids solve problems their parents can't solve, rather than focusing on you know, competitive, like preparing kids to
Ty Pattison: I got answer that comes to mind straight away. Take train retrain the most capable models on the
Andrew Davis: compete in the economy, get, you know, create educational environments that are basically giving kids reasonably powerful tools and just inviting them to solve real problems for for the societies that are around.
Ty Pattison: region specific, country specific,
River Roberts: I
Ty Pattison: like if you think about the individual, the household, the city, the state, the country, the world, and think about those scales and train a model on say the data in Vietnam. Train the a model on the data in Burkina Faso and
River Roberts: I I feel like it is is kind of a combination of just some really basic infrastructure investments. So one one of the underutilised infrastructure investments I think is actually this idea of localized Wi Fi networks that kind of by default offer connectivity to everyone in the area, but connectivity through a local regional perspective as well.
Ty Pattison: have versions of these at a point in time with a particular context.
River Roberts: So like if you could have free internet access to anywhere that you were located to, but that actual connectivity access offered you local context to your models as well, you could have a really informed network. And then as you were using it, you would then equally contribute to that local Wi-Fi. And the only time I really see local Wi-Fi these days is on an aeroplane. But
Ty Pattison: and maybe that would also require gathering training data from sources that are not online because disproportionately. I just think that there's a space here to to use this technology in the way that we create these models, but have different training data that might require time and energy to collect, but it's probably worthwhile.
River Roberts: I actually think that old, relatively old technology could could work really well for training these these local smaller models that connected much more to local context and I think local context is actually gonna be what makes these technologies not only more accessible but much more valuable in local context. So yeah that's that's one of the things that I'd I'd I'd think is is worth investing in. Gents.
Andrew Davis: It's it seems like a really important idea about like what is your national strategy for enabling your population? Right, like countries like South Korea blanketed cities with free Wi-Fi, public free public
River Roberts: Yeah.
Andrew Davis: Wi-Fi. So if you begin to think you think about intelligence and computation as a public good that you want to provide your citizenry, what you know, can you organize your society? What is the most effective way to do that without you know tying you to US-based companies that'll extort over time? Like what what independent ways
River Roberts: Yeah. I think I think the national scale is too big. I think it like the national it it's it just doesn't quite capture the scale but I think would be really meaningful and and valuable to a lot more people. So I still believe it's more of the regional or more of the local or even just neighborhood version of this. like what what is your what is your neighborhood AI do in terms of being able to make everyone have more abundance in resources and in relationships and in connectivity and yeah.
Andrew Davis: Like computational commons.
River Roberts: Exactly. Yeah. Computational commons. You heard it yeah. Potentially first.
Andrew Davis: It's it's an honor and a joy to be able to join you guys. This is River, I I now I understand what you meant about, you know, Ty being a great collaborator and I was such an honor to bring a graph to this spin up for the first time.
River Roberts: Yes. Andrew, thank you so much for joining us on this podcast. Ty. A pleasure as always. Now we have this weird closing ritual which maybe doesn't particularly translate to the audio listeners, but we we kind of finish with one of these. And then we slowly go like
Ty Pattison: Thank you for being here, Andrew.
