Joel Armstrong (00:09)
Welcome to Between Two Joels. I'm Joel Armstrong. And sure, we may have the same first name, basically the same last name, and both our middle names were named after our fathers, but in some ways we're very different people. For example, Joel's firstborn son was born on May twentieth and my firstborn son was born May twenty first.
Joel Anderson (00:11)
And I'm Joel Anderson.
Emma Sabry (00:27)
Hey there, I'm Emma. I'm one of the producers on the show, and I'm going to be the voice behind the camera. You won't see me, but you will hear me pop in and out occasionally throughout the episodes. So let's start off with some news. The first piece I have for you is about Microsoft, Alphabet, Meta, and Amazon, four of the magnificent seven, the seven biggest technology focused companies, who are still on track to spend a combined $650 billion on AI this year.
The largest spending commitment in corporate history, but investors are split on whether the buildup is justified or inflating a bubble. What do you guys think?
Joel Anderson (01:04)
I'm pretty bullish on future AI developments and impact on economy and stocks and investments and all that. I think because right now it's very hard to know how far we are along the economic returns of AI in terms of chatbots and not just chatbots, but like agentic services and agentic tools. I would argue that we're still on the first ... towards the beginnings of that.
Joel Armstrong (01:31)
Early innings for sure, yeah.
Joel Anderson (01:34)
I think we're both on the same page that we're still relatively early on with that, but I would say that with the overall potential of AI progress, we're still at the very, very beginning of that whole thing because I'm very bullish on things like robotics having a huge impact to the economy in the future. It's not necessarily going to be in like two years, although it might start to be. Robotics could in two to three years have the sort of like ChatGPT moment that ChatGPT had at the end of 2022.
Joel Armstrong (02:00)
Yeah. I mean autonomous cars exist already. They're starting to be one of the ways people get around.
Joel Anderson (02:05)
Exactly, yeah. But just the potential impact of if you can buy a humanoid robot for $10,000 and put it in your home and help you do laundry and meal prep and dishwasher and cutting the grass. Raise your kids. All these different things. I think tons of people would be jumping on that bandwagon.
Joel Armstrong (02:15)
Yeah, yeah, I think so. I mean, that's one of the interesting things, right? Is that people kind of talk about the AI build-out, all the, you know, all this compute, all these data centers as being sort of a binary good decision or bad decision, and we'll find out soon whether it's good or bad. But we've talked about it a fair bit, that it's like a resource, a generalizable resource. And people act as though it's fully dependent on whether generative AI and these LLMs hit exact targets of being, you know, this useful by this date.
whether or not this is money well spent or just money that someone like lit on fire. But we know, and the reason like both of us took our decided to do our masters in AI before LLM's, because we both believed that AI was, you know, starting to become something that really mattered. If the industrial revolution changed the way that we use machines to do work, I think we both believe that AI is starting to be a revolution, the intelligence revolution, and will fundamentally change the way we do knowledge work. And we're just like at the very beginning of that. Totally.
Joel Anderson (03:19)
I had a really good idea and I just forgot it.
Joel Armstrong (03:24)
That's my favorite category of joke. Not overplayed. not overdone at But even if we do talk about large language models, like agents are starting to get very good and very useful. And it takes a little while for a new technology for people to figure out how to use it. And even if it just replaced this section of the economy that's like software engineering and coding, that's already having a massive impact on a major part of the economy and the way people are doing work.
Joel Anderson (03:51)
Totally agree. I don't think the economy quite appreciates the potential that agent decoding has on the whole economy as a whole.
Joel Armstrong (04:01)
But we are starting to see that kind of play out a bit, right? Like Anthropic's revenue is 10xing every year as people find more and more ways to use it. Google's earnings, I think, is probably part of the reason why this article is kind of talking about these particular findings. Like Google is, you know, six times whatever particular AI line item it was that they were measuring. But like this is getting integrated into the way people do work. It's just it it's here now. Like it's here. There's no going back. Like, I don't think. I can't imagine anyone in the knowledge worker economy, in the knowledge economy who decides AI doesn't help them and decides to, you know, going forward we're not we're not doing this.
Joel Anderson (04:39)
I saw this recently that I think is a really interesting take that if you can benchmark a capability in AI, then almost by definition, it's something that AI will saturate pretty quickly because you can benchmark it. So all the complicated things you can't benchmark. Like people's different, like better writing styles or creating new physics or creating a new genre.
All these things you can't benchmark. So AI can't grow in them.
Joel Armstrong (05:12)
Yeah. Yeah. That makes a lot of sense, both from the perspective of like because it's measurable, it allows you like it facilitates the process of optimizing it, but it also creates like a feedback loop where when you get measured on it, then people are very incentivized to try and maximize for whatever is on that benchmark because then they can advertise that they're good at that benchmark. And we know like you and I talked about a bunch about like all the different ways there are to game benchmarks ⁓ in ways that don't really translate to actual, you know, real world utility, but allow you to point to your score and say, look how good our model is.
Joel Anderson (05:41)
Totally agree. And an interesting fact I saw recently too is ⁓ only 20 % of training compute is being spent on pre-training now. The rest is being used on different forms of post-training. like reinforcement, know, human enforcement learning with human feedback. That's kind of the old framework and that was just a very small section of the overall training. now there's all these different post-training steps.
and different ways to maximize and ⁓ different domains that they have experience in partly because they want to saturate all the benchmarks to make their model look better.
Joel Armstrong (06:20)
Right. Anything that you can benchmark, you can use reinforcement learning for, which means you remove both data sets and human training as a necessary component for improvement on. You can just give them give the models end results that they're targeting and then let them iterate on themselves for a long time and figure out how to reason their way to it.
Joel Anderson (06:37)
Yeah, exactly. And that's partly why you get this jaggedness of these models sometimes where, you know, GPT 5.5 is not as good as GPT 5.4 on chemistry, for example. Right. Things like that. Because sometimes the, you know, the relative weighting ⁓ of these new post-training steps and reinforcement learning that it's doing, ⁓ they sort of, you know, don't, they don't always move up in uniform.
Joel Armstrong (07:05)
So today we're gonna be talking about something we've been working on that we're very excited about called system zero. So before we get into system zero, Joel, do you wanna give us a quick rundown of what system one and system two thinking is and how that kind of led us to system zero?
Joel Anderson (07:18)
Sure, yeah. System one and two thinking a lot of people are familiar with. It's thinking fast and slow, popularized by Daniel Kahneman in a book about 15 years ago. ⁓ System one and two thinking is about how our brain processes different information, whether we process it automatically and quickly, or we process it slowly and rationally and more deliberately, weighing pros and cons before we make a decision. System zero is this new concept of how AI mediates choices that we make, because in a lot of cases,
We type it on a keyboard or on our mobile phones, and then we ask AI, what would you do in this situation? Or I'm looking to buy a new laptop. What are the brands I should consider? Or what are the specs I should care about? And so it's no longer about how a consumer on their own is thinking about and processing the information that they're taking in through advertisement. It's also about what is AI's influence in all this?
Joel Armstrong (08:13)
Right. So basically the idea is kind of like there's a new way or a new tool that we sometimes sub in for some component of our thinking. It's not just thinking fast or slow. Sometimes we just slot in this tool and it does our thinking for us. And we think that's pretty important because
Joel Anderson (08:28)
We don't even think about it at all. We just delegate our thinking to AI. And that's important because then we need to understand, traditionally as market researchers, we care about what are the different implications of thinking fast and slow on different aspects of the work that we do. And so now with system zero, we have to think about what is this whole other external factor. So what we want to do is we want to measure AI's influences and AI's latent preferences that it may have that are nudging consumers towards or away from different factors that brands care about. We did a survey a few weeks ago of consumers and we asked them how many of you are planning on using AI for different aspects of the decision making process for buying a new product in the future. And we included options like AI does research for me, all the way to AI buys it for me. And we found that 65 % of people are planning to use AI as part of the decision process for buying things in the near future.
Joel Armstrong (09:26)
Yeah. And so the nice thing is, given, you know, the experience we have as a company and you and I as individuals in the market research industry, is that we have working frameworks, right, for understanding how people are making decisions, what the different points are along the decision process. And so over the last, you know, little while, you and I have been trying to map out like what are the elements of the decision making process that can be represented by AI or where AI can slot in, you know, something like a traditional brand funnel or something like that. We're starting to explore jobs to be done. ⁓
Joel Anderson (09:55)
Exactly,
yeah. So something that we've done at Dig for a long time is conjoint analysis. And what you can do is you can actually put an AI through a similar type of experiment and very different things like ⁓ the brand that's involved and the different features that it has, different price points and other aspects of the product, and then understand what an AI would recommend a consumer choose.
Joel Armstrong (10:17)
Right. And the conjoint structure is really important, right? Because if you just ask a model certain questions, then we know that they can be, you know, kind of obsequious. They'll tend to bend the knee and tell you what they think you want to hear. They'll be a little bit overly polite. You know, if we asked about a random brand and said, like, is this a good brand? Are they reliable? Those sorts of things. It'll give you a sort of vague yes and then hedge on a bunch of different dimensions, all that sort of stuff. But by making it a forced choice, we make sure that the model is
forced to represent something inside of it about which of these is relatively higher, which of these is relatively lower, which is the whole point of a conjoint.
Joel Anderson (10:50)
And the reason that matters is because when a consumer uses an AI, they're just going to go in and say, the AI is going to know different aspects about them. Like it might know from their history, might know how much income they have, or if they tend to choose products that are higher end or lower end, or where they live geographically, which influences the brand availability and all these things. And the features they look for, are they really tech driven or are they really kind of, they just want the simple version of everything. And so...
When that consumer then asks the AI, what should I think about or which brand should I consider or which brands would you recommend, which brand would be best for me, then the AI is going to draw on all those latent preferences that we're trying to extract with this conjoint analysis. A lot of big brands, a lot of big incumbent brands, their representation on something like a Reddit is they may have lot of complaints about them because they may have been around for a while, commanding lead in a certain sector. They might have built up various complaints over time.
Whereas a lot of challenger brands, newcomer brands, they often will have a niche audience that has a lot of loyalty behind them. And so what happens in the online discourse is that those challenger brands will often have really positive ⁓ affinity with a lot of people that are online, which is this training data that's going ⁓ into the large language models. And so it's not just a random representation, it's really this skew of
Generally speaking, this is over generalization, but what we're seeing so far as we start to explore this is that a lot of these challenger brands, these newcomer brands are having ⁓ a lot of positive brand equity in these large language models. And a lot of the incumbent brands don't have that same level of positive representation that you might expect.
Joel Armstrong (12:36)
So there certainly significant implications when it comes to, you know, individuals and how we engage in the world and what we think about and what we don't and what sort of gets made for us by AI. Do you think there are any ethical implications to system zero as we've sort of sort of mapped it out so far?
Joel Anderson (12:50)
Well, I think people don't realize the nudges and the biases that AI has and how it will influence them sort of systematically in a certain way. For example, when we were in our masters, we had to read Weapons of Math Destruction. And it was a very interesting book and it was all about different applications of AI and how they push people in certain directions. But everything, know, newspapers.
All kinds of media, books, all kinds of media over the last centuries have always affected people in different ways. But the difference with AI and machine learning now is that it scales to massive, you know, it scales massively. So now these large language models are influencing billions of people. So even a small nudge has potential huge ramifications. And so I don't think the average consumer realizes that there's all these little ways that they might get systematically affected.
For example, we know that the AIs have the sort of weird bias.
Joel Armstrong (13:48)
So we've actually done some research to try and figure out, particularly by industry or category, where it is that people are using AI and where they're not. And that's a pretty good, pretty direct ⁓ analog to when people are relying on system zero and when they're not. And so do you wanna kinda give us a quick rundown of what we found when we were looking at that?
Joel Anderson (14:06)
Yeah, so there's certain categories like areas like insurance and mobile phone plan selection that are areas that people use AI more for because it's kind of like a high stress situation, potential high impact, picking something that you'll pay monthly for for years to come. And there's also this element of fungibility, like you said before, if there's this idea that
They kind of are all replacements for each other. if I can ⁓ pay for the same service, but pay less, then great. That's what I want as a consumer. ⁓ Whereas on the opposite side of the spectrum, it's things like your personal choice, categories like what type of alcohol you want to buy, what kind of clothes you want to buy and wear, your own style. These are categories that are less susceptible to AI's influence.
Joel Armstrong (14:59)
Right. And so that kind of makes sense, right? That like in a situation where you're trying to optimize for something, the process you would normally go through is like for a cell phone plan, for example. You could go check, you know, the five cell phone providers that come to mind and say, okay, this is their price per month, ⁓ this is what you get with it, all those sorts of things. Or you could basically kind of offload that to AI and assume it's doing roughly the same thing, where it's, you know, searching the web or has that information, you know, represented internally or whatever that might be. But something like
A cell phone plan or insurance or credit cards is sort of, you know, an evaluation of a relatively objective set of measures to try and optimize what the best opportunity is for you.
Joel Anderson (15:38)
An interesting category too is ⁓ choices on healthcare because it's an area that you don't want to just delegate to AI. You want to stay highly involved in that process. I still think that people are using AI to help them make the choice, but they ultimately want to have a high level of autonomy throughout that and they're less likely to delegate as much to AI there. And so, you know, even though it's really high decision fatigue because it's exhausting to look through all the different options and they're
potentially really high ramifications on your ⁓ future. It's an area that people still want to maintain that autonomy. So it's kind of an interesting combination.
Joel Armstrong (16:19)
Yeah. If we're talking about areas of concern as well when it comes to the way people are using AI or offloading decisions, I would say health is something that I'd be particularly concerned about. I mean, there's already a tendency with Google and those sorts of things for people to, you know, think they were better doctors than doctors and then also sort of translate their subjective experience of like this doctor was a bit rude, I don't like them, they're probably not a good doctor, you know, those sorts of things. Whereas obviously it's useful to sort of gain context and understand things that are happening to you, be able to sort of
ask relevant questions if you go to a doctor or things like that. But ⁓ trusting AI to be your doctor, I think, is one of those areas where I would be cautious. That is a thing that I think is of some concern rather than trusting doctors.
Joel Anderson (16:59)
Yeah, especially when you're taking into account the psycho-fancy of AI, you know, to reinforce or echo back to what you want to hear about it. This is a dangerous slope.
Joel Armstrong (17:09)
Yeah, exactly. So if we look at ⁓ sort of what we were just talking about, like the groupings of what people use AI for and what they don't use AI for, then we find that vacation is actually up towards the, you know, heavy AI usage, more similar to some of the optimization stuff we were talking about. But in a lot of ways that would make more sense to both of us, I think, as like a more subjective experience. So why do you think vacation is kind of grouped up there?
Joel Anderson (17:32)
Yeah, think vacations are people enjoy researching their own vacations because it's fun and they have something to look forward to, so they enjoy doing it. But they still want to use AI for it often because they want to make sure that they sort of exhaustively explore the space, right? You don't want to go on a vacation and then find out when you got home that there was something amazing you could have done right beside you.
Joel Armstrong (17:53)
So it's interesting. You and I are pretty different when it comes to a lot of the ways we make decisions. This is an actual way that we are pretty different. ⁓ so you're much more of a maximizer and I'm much more of a satisficer. And so that totally makes sense. That is one of the things you wouldn't want to like come home and regret that you had missed out on something. For me, as a satisficer, I would like to use AI for vacation because it's such a giant wide open problem space of like, well, I don't know what's in Italy. I'm gonna find some things that I think would be useful. And I know that I'm not gonna come home and then Google and try and compare like, did I miss out on something?
And so I think one of the reasons for someone like me that the AI would be a good way of planning a vacation is it just like does a lot of narrowing down for you automatically when there's like no real wrong answers. Like there's a million fun things you could do on any vacation that you go to. Like last year, Laura and I went to Iceland and we bought the actual like Lonely Planet book and it was like a five day trip and we didn't know what we were gonna do. We on the plane on the way there, we took the Lonely Planet.
And we just kinda cracked it open and we're like, here's some different loops that would be available. Here's a set of things we could do in the time that we have available. What's that?
Joel Anderson (18:55)
So
stressful. It's funny. Yeah, it's very, you're right. Totally different than how I approach it. I recently went with my son down to Texas to do a big hiking trip and we wanted to maximize all the cool places that we can visit. So we were driving like two hours out of our way to do this and an hour out of our way to do that and all these different places to hit up. So I use AI obviously as part of this process to plan and trip to figure out all the different places we could go. And I want to plan it all before I even buy the tickets because I don't know ⁓ where I'm flying into, where I'm flying out of.
which the best way to do it, yeah, you're totally right. I'm more of a vacation maximizer optimizer. What about if you're buying a TV? Then how are you using system zero one two in that process?
Joel Armstrong (19:38)
Yeah. Well that's an interesting this is maybe the one instance where we'd be heavily reversed. I take TVs very seriously. So I already kind of know. I know at any given time if I were to buy a TV which one I would buy, and that's because it's almost like a hobby of mine to keep up with TVs and TV technology and know what's available and what the relative prices are and all that kind of stuff. So I'd be very system two, ⁓ very system two heavy when it came to how to choose a TV.
Joel Anderson (20:02)
This
is another funny example because I literally bought the same TV you had last time.
Joel Armstrong (20:07)
Yeah,
last time you needed a t new TV and I've had other friends do that too. They're like, What T V should I buy? I'm like, I would buy this one if you don't mind the price and then yeah, a few people, including yourself, have just bought that T V and has it been working out for you though. Great TV. It's a pretty good TV.
Joel Anderson (20:19)
I don't know if I would have noticed really, yeah, that's a good, so in that case, I'm basically just using system one because I'm saying I'm just going to take your, you're my AI. It's almost like system zero. Ooh, interesting. Yeah. It is sort of system zero, but not AI mediated.
Joel Armstrong (20:34)
Which makes sense from the like likelihood to recommend or just general recommendation. That's why we kind of map onto that part of the brand funnel is one of the ways that people make decisions is looking for recommendations from trusted sources. And so that would be just sort of sort of instance where we've actually identified system zero can slot in as a sort of trusted recommender.
Joel Anderson (20:52)
And in this case, I knew you're a trusted, because I know you do that research and I know you had just bought a TV back when I needed one a few years ago. So I went to you, but in most cases you don't have a friend who just bought the thing that you're about to buy. that's why AI kind of plays a huge role.
Emma Sabry (21:12)
Okay, so now we're gonna do a little game where you guys are gonna guess what the other person would rather do. Like a game of would you rather but like a newlywed game, but for each of you.
Joel Anderson (21:23)
Amazing. We couldn't have created a better game.
Joel Armstrong (21:25)
We're very good at the newlywed game.
Emma Sabry (21:27)
Okay. Adventure or relaxing vacation?
Joel Anderson (21:31)
Adventure. Oh. That's
right.
Joel Armstrong (21:32)
⁓ no, for the other person. ⁓ Joel it's a hundred percent adventure.
Joel Anderson (21:37)
for Joel.
Emma Sabry (21:39)
Nice.
Okay. Coffee or tea?
Joel Armstrong (21:43)
Neither for Joel. Just water.
Joel Anderson (21:45)
Yeah, and I think you would do both, but especially coffee.
Joel Armstrong (21:48)
Yeah, I drink a lot more coffee than tea. Tea's I don't mind tea, but yeah.
Emma Sabry (21:52)
Morning person or night owl?
Joel Anderson (21:55)
Okay, I'm gonna go night owl for you.
Joel Armstrong (21:56)
And morning person for Joel.
Emma Sabry (21:58)
Physical book or an e reader?
Joel Armstrong (22:01)
Physical book. That was still my answer for you, yeah. Are you? Okay. Got one wrong. I was trying to picture I mean we'd have exchanged you most of your books on eReader?
Joel Anderson (22:14)
Yeah, because it's just so convenient and also I read while my wife's sleeping most of the time when I do most of my reading.
Joel Armstrong (22:21)
Yeah. Yeah, that was just my guess 'cause we've like exchanged books and stuff. Yeah. Not the ones that you have on E Reader for obvious reasons. Yeah. I like E Reader. I definitely am not a ⁓ you know, I'm not a hardliner on it, but I do tend to read physical books, yeah.
Joel Anderson (22:26)
Yeah, only the last three years.
Emma Sabry (22:36)
a podcast or music on a commute?
Joel Armstrong (22:39)
Podcast for Joel.
Joel Anderson (22:41)
Yeah, I'm going to go. feel like 60/40 music to podcast for you.
Joel Armstrong (22:46)
More like 90/10. Music? I barely listen to podcasts.
Joel Anderson (22:49)
Yeah.
Okay. Well, I got the density in the right spot, but not the amount of...
Joel Armstrong (22:55)
Yeah, no, I try to listen to music almost all the time. And if I don't listen to music, I don't really listen to podcasts, I listen to audiobooks. But I that's essentially the same thing, I think. But
Emma Sabry (23:02)
What did you listen to this morning on your commute?
Joel Armstrong (23:05)
⁓ I read a physical book. Actually. ⁓ yeah. Speaking of, that's not only- so I warned you guys earlier that I didn't sleep very well last night. So I have to I take a train here. I live in London a couple hours away. ⁓ so I have to get up at like five forty five to make it here to make my train. And so last night I was like, I need to go to bed earlier, like, you know, ten thirty or something like that. And I ended up reading
for to like eleven thirty, 'cause I was at like the climax of my book. And I'm like, it's okay, I'll make it up, I'll sleep on the train tomorrow and then I brought the book with me and still read on the train and so I'm on like five hours sleep because I just read instead of sleeping both last night and this morning.
Emma Sabry (23:45)
What book?
Joel Armstrong (23:46)
I'm reading Oathbringer. no, that's not true. I I'm reading Words of Radiance, which is the second book in Brandon Sanderson's Stormlight Archives, ⁓ which I anticipate finishing on the train home tonight. So I also brought the third book, which is Oathbringer or Oath Keeper? Something along those lines. Yeah. So I brought that as well. I'm anticipating starting that on the train home tonight as Did you listen to a podcast on the way here today? I did.
Joel Anderson (24:02)
I did. I listened to, um, it was something about AI. was a boring anyway. forget what it was called, but yeah.
Joel Armstrong (24:22)
Something about AI. Podcast about it. That's what I would have guessed. Well, I think we're gonna end on that note. Sure. So ⁓ yeah. So thank you very much for joining us. This is our first Between Two Joels, but there's gonna be a lot more where this came from and we'll talk to you again sometime soon.
Joel Anderson (24:39)
Thank you and good night.
Joel Armstrong (24:41)
We can cut that.
Joel Anderson (24:43)
Ha!
Welcome to Between Two Joels. I'm Joel Armstrong. And sure, we may have the same first name, basically the same last name, and both our middle names were named after our fathers, but in some ways we're very different people. For example, Joel's firstborn son was born on May twentieth and my firstborn son was born May twenty first.
Joel Anderson (00:11)
And I'm Joel Anderson.
Emma Sabry (00:27)
Hey there, I'm Emma. I'm one of the producers on the show, and I'm going to be the voice behind the camera. You won't see me, but you will hear me pop in and out occasionally throughout the episodes. So let's start off with some news. The first piece I have for you is about Microsoft, Alphabet, Meta, and Amazon, four of the magnificent seven, the seven biggest technology focused companies, who are still on track to spend a combined $650 billion on AI this year.
The largest spending commitment in corporate history, but investors are split on whether the buildup is justified or inflating a bubble. What do you guys think?
Joel Anderson (01:04)
I'm pretty bullish on future AI developments and impact on economy and stocks and investments and all that. I think because right now it's very hard to know how far we are along the economic returns of AI in terms of chatbots and not just chatbots, but like agentic services and agentic tools. I would argue that we're still on the first ... towards the beginnings of that.
Joel Armstrong (01:31)
Early innings for sure, yeah.
Joel Anderson (01:34)
I think we're both on the same page that we're still relatively early on with that, but I would say that with the overall potential of AI progress, we're still at the very, very beginning of that whole thing because I'm very bullish on things like robotics having a huge impact to the economy in the future. It's not necessarily going to be in like two years, although it might start to be. Robotics could in two to three years have the sort of like ChatGPT moment that ChatGPT had at the end of 2022.
Joel Armstrong (02:00)
Yeah. I mean autonomous cars exist already. They're starting to be one of the ways people get around.
Joel Anderson (02:05)
Exactly, yeah. But just the potential impact of if you can buy a humanoid robot for $10,000 and put it in your home and help you do laundry and meal prep and dishwasher and cutting the grass. Raise your kids. All these different things. I think tons of people would be jumping on that bandwagon.
Joel Armstrong (02:15)
Yeah, yeah, I think so. I mean, that's one of the interesting things, right? Is that people kind of talk about the AI build-out, all the, you know, all this compute, all these data centers as being sort of a binary good decision or bad decision, and we'll find out soon whether it's good or bad. But we've talked about it a fair bit, that it's like a resource, a generalizable resource. And people act as though it's fully dependent on whether generative AI and these LLMs hit exact targets of being, you know, this useful by this date.
whether or not this is money well spent or just money that someone like lit on fire. But we know, and the reason like both of us took our decided to do our masters in AI before LLM's, because we both believed that AI was, you know, starting to become something that really mattered. If the industrial revolution changed the way that we use machines to do work, I think we both believe that AI is starting to be a revolution, the intelligence revolution, and will fundamentally change the way we do knowledge work. And we're just like at the very beginning of that. Totally.
Joel Anderson (03:19)
I had a really good idea and I just forgot it.
Joel Armstrong (03:24)
That's my favorite category of joke. Not overplayed. not overdone at But even if we do talk about large language models, like agents are starting to get very good and very useful. And it takes a little while for a new technology for people to figure out how to use it. And even if it just replaced this section of the economy that's like software engineering and coding, that's already having a massive impact on a major part of the economy and the way people are doing work.
Joel Anderson (03:51)
Totally agree. I don't think the economy quite appreciates the potential that agent decoding has on the whole economy as a whole.
Joel Armstrong (04:01)
But we are starting to see that kind of play out a bit, right? Like Anthropic's revenue is 10xing every year as people find more and more ways to use it. Google's earnings, I think, is probably part of the reason why this article is kind of talking about these particular findings. Like Google is, you know, six times whatever particular AI line item it was that they were measuring. But like this is getting integrated into the way people do work. It's just it it's here now. Like it's here. There's no going back. Like, I don't think. I can't imagine anyone in the knowledge worker economy, in the knowledge economy who decides AI doesn't help them and decides to, you know, going forward we're not we're not doing this.
Joel Anderson (04:39)
I saw this recently that I think is a really interesting take that if you can benchmark a capability in AI, then almost by definition, it's something that AI will saturate pretty quickly because you can benchmark it. So all the complicated things you can't benchmark. Like people's different, like better writing styles or creating new physics or creating a new genre.
All these things you can't benchmark. So AI can't grow in them.
Joel Armstrong (05:12)
Yeah. Yeah. That makes a lot of sense, both from the perspective of like because it's measurable, it allows you like it facilitates the process of optimizing it, but it also creates like a feedback loop where when you get measured on it, then people are very incentivized to try and maximize for whatever is on that benchmark because then they can advertise that they're good at that benchmark. And we know like you and I talked about a bunch about like all the different ways there are to game benchmarks ⁓ in ways that don't really translate to actual, you know, real world utility, but allow you to point to your score and say, look how good our model is.
Joel Anderson (05:41)
Totally agree. And an interesting fact I saw recently too is ⁓ only 20 % of training compute is being spent on pre-training now. The rest is being used on different forms of post-training. like reinforcement, know, human enforcement learning with human feedback. That's kind of the old framework and that was just a very small section of the overall training. now there's all these different post-training steps.
and different ways to maximize and ⁓ different domains that they have experience in partly because they want to saturate all the benchmarks to make their model look better.
Joel Armstrong (06:20)
Right. Anything that you can benchmark, you can use reinforcement learning for, which means you remove both data sets and human training as a necessary component for improvement on. You can just give them give the models end results that they're targeting and then let them iterate on themselves for a long time and figure out how to reason their way to it.
Joel Anderson (06:37)
Yeah, exactly. And that's partly why you get this jaggedness of these models sometimes where, you know, GPT 5.5 is not as good as GPT 5.4 on chemistry, for example. Right. Things like that. Because sometimes the, you know, the relative weighting ⁓ of these new post-training steps and reinforcement learning that it's doing, ⁓ they sort of, you know, don't, they don't always move up in uniform.
Joel Armstrong (07:05)
So today we're gonna be talking about something we've been working on that we're very excited about called system zero. So before we get into system zero, Joel, do you wanna give us a quick rundown of what system one and system two thinking is and how that kind of led us to system zero?
Joel Anderson (07:18)
Sure, yeah. System one and two thinking a lot of people are familiar with. It's thinking fast and slow, popularized by Daniel Kahneman in a book about 15 years ago. ⁓ System one and two thinking is about how our brain processes different information, whether we process it automatically and quickly, or we process it slowly and rationally and more deliberately, weighing pros and cons before we make a decision. System zero is this new concept of how AI mediates choices that we make, because in a lot of cases,
We type it on a keyboard or on our mobile phones, and then we ask AI, what would you do in this situation? Or I'm looking to buy a new laptop. What are the brands I should consider? Or what are the specs I should care about? And so it's no longer about how a consumer on their own is thinking about and processing the information that they're taking in through advertisement. It's also about what is AI's influence in all this?
Joel Armstrong (08:13)
Right. So basically the idea is kind of like there's a new way or a new tool that we sometimes sub in for some component of our thinking. It's not just thinking fast or slow. Sometimes we just slot in this tool and it does our thinking for us. And we think that's pretty important because
Joel Anderson (08:28)
We don't even think about it at all. We just delegate our thinking to AI. And that's important because then we need to understand, traditionally as market researchers, we care about what are the different implications of thinking fast and slow on different aspects of the work that we do. And so now with system zero, we have to think about what is this whole other external factor. So what we want to do is we want to measure AI's influences and AI's latent preferences that it may have that are nudging consumers towards or away from different factors that brands care about. We did a survey a few weeks ago of consumers and we asked them how many of you are planning on using AI for different aspects of the decision making process for buying a new product in the future. And we included options like AI does research for me, all the way to AI buys it for me. And we found that 65 % of people are planning to use AI as part of the decision process for buying things in the near future.
Joel Armstrong (09:26)
Yeah. And so the nice thing is, given, you know, the experience we have as a company and you and I as individuals in the market research industry, is that we have working frameworks, right, for understanding how people are making decisions, what the different points are along the decision process. And so over the last, you know, little while, you and I have been trying to map out like what are the elements of the decision making process that can be represented by AI or where AI can slot in, you know, something like a traditional brand funnel or something like that. We're starting to explore jobs to be done. ⁓
Joel Anderson (09:55)
Exactly,
yeah. So something that we've done at Dig for a long time is conjoint analysis. And what you can do is you can actually put an AI through a similar type of experiment and very different things like ⁓ the brand that's involved and the different features that it has, different price points and other aspects of the product, and then understand what an AI would recommend a consumer choose.
Joel Armstrong (10:17)
Right. And the conjoint structure is really important, right? Because if you just ask a model certain questions, then we know that they can be, you know, kind of obsequious. They'll tend to bend the knee and tell you what they think you want to hear. They'll be a little bit overly polite. You know, if we asked about a random brand and said, like, is this a good brand? Are they reliable? Those sorts of things. It'll give you a sort of vague yes and then hedge on a bunch of different dimensions, all that sort of stuff. But by making it a forced choice, we make sure that the model is
forced to represent something inside of it about which of these is relatively higher, which of these is relatively lower, which is the whole point of a conjoint.
Joel Anderson (10:50)
And the reason that matters is because when a consumer uses an AI, they're just going to go in and say, the AI is going to know different aspects about them. Like it might know from their history, might know how much income they have, or if they tend to choose products that are higher end or lower end, or where they live geographically, which influences the brand availability and all these things. And the features they look for, are they really tech driven or are they really kind of, they just want the simple version of everything. And so...
When that consumer then asks the AI, what should I think about or which brand should I consider or which brands would you recommend, which brand would be best for me, then the AI is going to draw on all those latent preferences that we're trying to extract with this conjoint analysis. A lot of big brands, a lot of big incumbent brands, their representation on something like a Reddit is they may have lot of complaints about them because they may have been around for a while, commanding lead in a certain sector. They might have built up various complaints over time.
Whereas a lot of challenger brands, newcomer brands, they often will have a niche audience that has a lot of loyalty behind them. And so what happens in the online discourse is that those challenger brands will often have really positive ⁓ affinity with a lot of people that are online, which is this training data that's going ⁓ into the large language models. And so it's not just a random representation, it's really this skew of
Generally speaking, this is over generalization, but what we're seeing so far as we start to explore this is that a lot of these challenger brands, these newcomer brands are having ⁓ a lot of positive brand equity in these large language models. And a lot of the incumbent brands don't have that same level of positive representation that you might expect.
Joel Armstrong (12:36)
So there certainly significant implications when it comes to, you know, individuals and how we engage in the world and what we think about and what we don't and what sort of gets made for us by AI. Do you think there are any ethical implications to system zero as we've sort of sort of mapped it out so far?
Joel Anderson (12:50)
Well, I think people don't realize the nudges and the biases that AI has and how it will influence them sort of systematically in a certain way. For example, when we were in our masters, we had to read Weapons of Math Destruction. And it was a very interesting book and it was all about different applications of AI and how they push people in certain directions. But everything, know, newspapers.
All kinds of media, books, all kinds of media over the last centuries have always affected people in different ways. But the difference with AI and machine learning now is that it scales to massive, you know, it scales massively. So now these large language models are influencing billions of people. So even a small nudge has potential huge ramifications. And so I don't think the average consumer realizes that there's all these little ways that they might get systematically affected.
For example, we know that the AIs have the sort of weird bias.
Joel Armstrong (13:48)
So we've actually done some research to try and figure out, particularly by industry or category, where it is that people are using AI and where they're not. And that's a pretty good, pretty direct ⁓ analog to when people are relying on system zero and when they're not. And so do you wanna kinda give us a quick rundown of what we found when we were looking at that?
Joel Anderson (14:06)
Yeah, so there's certain categories like areas like insurance and mobile phone plan selection that are areas that people use AI more for because it's kind of like a high stress situation, potential high impact, picking something that you'll pay monthly for for years to come. And there's also this element of fungibility, like you said before, if there's this idea that
They kind of are all replacements for each other. if I can ⁓ pay for the same service, but pay less, then great. That's what I want as a consumer. ⁓ Whereas on the opposite side of the spectrum, it's things like your personal choice, categories like what type of alcohol you want to buy, what kind of clothes you want to buy and wear, your own style. These are categories that are less susceptible to AI's influence.
Joel Armstrong (14:59)
Right. And so that kind of makes sense, right? That like in a situation where you're trying to optimize for something, the process you would normally go through is like for a cell phone plan, for example. You could go check, you know, the five cell phone providers that come to mind and say, okay, this is their price per month, ⁓ this is what you get with it, all those sorts of things. Or you could basically kind of offload that to AI and assume it's doing roughly the same thing, where it's, you know, searching the web or has that information, you know, represented internally or whatever that might be. But something like
A cell phone plan or insurance or credit cards is sort of, you know, an evaluation of a relatively objective set of measures to try and optimize what the best opportunity is for you.
Joel Anderson (15:38)
An interesting category too is ⁓ choices on healthcare because it's an area that you don't want to just delegate to AI. You want to stay highly involved in that process. I still think that people are using AI to help them make the choice, but they ultimately want to have a high level of autonomy throughout that and they're less likely to delegate as much to AI there. And so, you know, even though it's really high decision fatigue because it's exhausting to look through all the different options and they're
potentially really high ramifications on your ⁓ future. It's an area that people still want to maintain that autonomy. So it's kind of an interesting combination.
Joel Armstrong (16:19)
Yeah. If we're talking about areas of concern as well when it comes to the way people are using AI or offloading decisions, I would say health is something that I'd be particularly concerned about. I mean, there's already a tendency with Google and those sorts of things for people to, you know, think they were better doctors than doctors and then also sort of translate their subjective experience of like this doctor was a bit rude, I don't like them, they're probably not a good doctor, you know, those sorts of things. Whereas obviously it's useful to sort of gain context and understand things that are happening to you, be able to sort of
ask relevant questions if you go to a doctor or things like that. But ⁓ trusting AI to be your doctor, I think, is one of those areas where I would be cautious. That is a thing that I think is of some concern rather than trusting doctors.
Joel Anderson (16:59)
Yeah, especially when you're taking into account the psycho-fancy of AI, you know, to reinforce or echo back to what you want to hear about it. This is a dangerous slope.
Joel Armstrong (17:09)
Yeah, exactly. So if we look at ⁓ sort of what we were just talking about, like the groupings of what people use AI for and what they don't use AI for, then we find that vacation is actually up towards the, you know, heavy AI usage, more similar to some of the optimization stuff we were talking about. But in a lot of ways that would make more sense to both of us, I think, as like a more subjective experience. So why do you think vacation is kind of grouped up there?
Joel Anderson (17:32)
Yeah, think vacations are people enjoy researching their own vacations because it's fun and they have something to look forward to, so they enjoy doing it. But they still want to use AI for it often because they want to make sure that they sort of exhaustively explore the space, right? You don't want to go on a vacation and then find out when you got home that there was something amazing you could have done right beside you.
Joel Armstrong (17:53)
So it's interesting. You and I are pretty different when it comes to a lot of the ways we make decisions. This is an actual way that we are pretty different. ⁓ so you're much more of a maximizer and I'm much more of a satisficer. And so that totally makes sense. That is one of the things you wouldn't want to like come home and regret that you had missed out on something. For me, as a satisficer, I would like to use AI for vacation because it's such a giant wide open problem space of like, well, I don't know what's in Italy. I'm gonna find some things that I think would be useful. And I know that I'm not gonna come home and then Google and try and compare like, did I miss out on something?
And so I think one of the reasons for someone like me that the AI would be a good way of planning a vacation is it just like does a lot of narrowing down for you automatically when there's like no real wrong answers. Like there's a million fun things you could do on any vacation that you go to. Like last year, Laura and I went to Iceland and we bought the actual like Lonely Planet book and it was like a five day trip and we didn't know what we were gonna do. We on the plane on the way there, we took the Lonely Planet.
And we just kinda cracked it open and we're like, here's some different loops that would be available. Here's a set of things we could do in the time that we have available. What's that?
Joel Anderson (18:55)
So
stressful. It's funny. Yeah, it's very, you're right. Totally different than how I approach it. I recently went with my son down to Texas to do a big hiking trip and we wanted to maximize all the cool places that we can visit. So we were driving like two hours out of our way to do this and an hour out of our way to do that and all these different places to hit up. So I use AI obviously as part of this process to plan and trip to figure out all the different places we could go. And I want to plan it all before I even buy the tickets because I don't know ⁓ where I'm flying into, where I'm flying out of.
which the best way to do it, yeah, you're totally right. I'm more of a vacation maximizer optimizer. What about if you're buying a TV? Then how are you using system zero one two in that process?
Joel Armstrong (19:38)
Yeah. Well that's an interesting this is maybe the one instance where we'd be heavily reversed. I take TVs very seriously. So I already kind of know. I know at any given time if I were to buy a TV which one I would buy, and that's because it's almost like a hobby of mine to keep up with TVs and TV technology and know what's available and what the relative prices are and all that kind of stuff. So I'd be very system two, ⁓ very system two heavy when it came to how to choose a TV.
Joel Anderson (20:02)
This
is another funny example because I literally bought the same TV you had last time.
Joel Armstrong (20:07)
Yeah,
last time you needed a t new TV and I've had other friends do that too. They're like, What T V should I buy? I'm like, I would buy this one if you don't mind the price and then yeah, a few people, including yourself, have just bought that T V and has it been working out for you though. Great TV. It's a pretty good TV.
Joel Anderson (20:19)
I don't know if I would have noticed really, yeah, that's a good, so in that case, I'm basically just using system one because I'm saying I'm just going to take your, you're my AI. It's almost like system zero. Ooh, interesting. Yeah. It is sort of system zero, but not AI mediated.
Joel Armstrong (20:34)
Which makes sense from the like likelihood to recommend or just general recommendation. That's why we kind of map onto that part of the brand funnel is one of the ways that people make decisions is looking for recommendations from trusted sources. And so that would be just sort of sort of instance where we've actually identified system zero can slot in as a sort of trusted recommender.
Joel Anderson (20:52)
And in this case, I knew you're a trusted, because I know you do that research and I know you had just bought a TV back when I needed one a few years ago. So I went to you, but in most cases you don't have a friend who just bought the thing that you're about to buy. that's why AI kind of plays a huge role.
Emma Sabry (21:12)
Okay, so now we're gonna do a little game where you guys are gonna guess what the other person would rather do. Like a game of would you rather but like a newlywed game, but for each of you.
Joel Anderson (21:23)
Amazing. We couldn't have created a better game.
Joel Armstrong (21:25)
We're very good at the newlywed game.
Emma Sabry (21:27)
Okay. Adventure or relaxing vacation?
Joel Anderson (21:31)
Adventure. Oh. That's
right.
Joel Armstrong (21:32)
⁓ no, for the other person. ⁓ Joel it's a hundred percent adventure.
Joel Anderson (21:37)
for Joel.
Emma Sabry (21:39)
Nice.
Okay. Coffee or tea?
Joel Armstrong (21:43)
Neither for Joel. Just water.
Joel Anderson (21:45)
Yeah, and I think you would do both, but especially coffee.
Joel Armstrong (21:48)
Yeah, I drink a lot more coffee than tea. Tea's I don't mind tea, but yeah.
Emma Sabry (21:52)
Morning person or night owl?
Joel Anderson (21:55)
Okay, I'm gonna go night owl for you.
Joel Armstrong (21:56)
And morning person for Joel.
Emma Sabry (21:58)
Physical book or an e reader?
Joel Armstrong (22:01)
Physical book. That was still my answer for you, yeah. Are you? Okay. Got one wrong. I was trying to picture I mean we'd have exchanged you most of your books on eReader?
Joel Anderson (22:14)
Yeah, because it's just so convenient and also I read while my wife's sleeping most of the time when I do most of my reading.
Joel Armstrong (22:21)
Yeah. Yeah, that was just my guess 'cause we've like exchanged books and stuff. Yeah. Not the ones that you have on E Reader for obvious reasons. Yeah. I like E Reader. I definitely am not a ⁓ you know, I'm not a hardliner on it, but I do tend to read physical books, yeah.
Joel Anderson (22:26)
Yeah, only the last three years.
Emma Sabry (22:36)
a podcast or music on a commute?
Joel Armstrong (22:39)
Podcast for Joel.
Joel Anderson (22:41)
Yeah, I'm going to go. feel like 60/40 music to podcast for you.
Joel Armstrong (22:46)
More like 90/10. Music? I barely listen to podcasts.
Joel Anderson (22:49)
Yeah.
Okay. Well, I got the density in the right spot, but not the amount of...
Joel Armstrong (22:55)
Yeah, no, I try to listen to music almost all the time. And if I don't listen to music, I don't really listen to podcasts, I listen to audiobooks. But I that's essentially the same thing, I think. But
Emma Sabry (23:02)
What did you listen to this morning on your commute?
Joel Armstrong (23:05)
⁓ I read a physical book. Actually. ⁓ yeah. Speaking of, that's not only- so I warned you guys earlier that I didn't sleep very well last night. So I have to I take a train here. I live in London a couple hours away. ⁓ so I have to get up at like five forty five to make it here to make my train. And so last night I was like, I need to go to bed earlier, like, you know, ten thirty or something like that. And I ended up reading
for to like eleven thirty, 'cause I was at like the climax of my book. And I'm like, it's okay, I'll make it up, I'll sleep on the train tomorrow and then I brought the book with me and still read on the train and so I'm on like five hours sleep because I just read instead of sleeping both last night and this morning.
Emma Sabry (23:45)
What book?
Joel Armstrong (23:46)
I'm reading Oathbringer. no, that's not true. I I'm reading Words of Radiance, which is the second book in Brandon Sanderson's Stormlight Archives, ⁓ which I anticipate finishing on the train home tonight. So I also brought the third book, which is Oathbringer or Oath Keeper? Something along those lines. Yeah. So I brought that as well. I'm anticipating starting that on the train home tonight as Did you listen to a podcast on the way here today? I did.
Joel Anderson (24:02)
I did. I listened to, um, it was something about AI. was a boring anyway. forget what it was called, but yeah.
Joel Armstrong (24:22)
Something about AI. Podcast about it. That's what I would have guessed. Well, I think we're gonna end on that note. Sure. So ⁓ yeah. So thank you very much for joining us. This is our first Between Two Joels, but there's gonna be a lot more where this came from and we'll talk to you again sometime soon.
Joel Anderson (24:39)
Thank you and good night.
Joel Armstrong (24:41)
We can cut that.
Joel Anderson (24:43)
Ha!
