Kevin Williams: Hey, AI at work. Welcome to a new week. this week we have a special guest because our buddy Matt is heads down in yet another hacker house. He seems to just like rove the world doing that now and again. I think he he might be in Austin, Texas. So he is he has conveniently offered up one of his partners in crime, a Vlad Larin. And Vlad, maybe you can introduce yourself for us. I I'm excited about this conversation because as as much as Matt and I love to geek out, our geekery has a certain limit because we're not like true devs. Like I like to think that I'm like the the business guy who's pretty dang technical, but I have limits. Whereas Vlad, you are an actual AI developer. So I think this could be a really fun conversation. for the audience to s to sort of hear from a commercial perspective of how you're thinking about this. And I suspect I'm gonna learn quite a bit from the conversation as well. So I'll be I'll be scribbling some notes down. So Vlad, welcome and maybe introduce yourself a bit.
Vlad: Thanks, Kevin. Yeah, here's hoping that this all comes amounts to something. So for sure. Hey everybody. my name is Vlad. I have been in tech for the last almost 20 years in various roles and applications. I do have a technical background, although it is somewhat diverse in the sense that I have been a software engineer back in the day, actually around 15 years ago. Then I transitioned mostly to the project management and technical project management. And for the last ten years or so, I've been mostly managing teams, engineers, and other various aspects of the business, but probably more focused on the technical perspective of things. That's why it's actually a very one of the most seamless transitions probably there is out there for people who are not necessarily like hands-on developers all all day, every day, but are closer to that part on the spectrum of The technicality or whatever you want to call it. So definitely these days I am what people would call an AI developer for sure. but this is actually very recent. This is a new career branch for me. So, and I feel like this is exactly a natural outcome of our experience, our aspirations, our preferences, kind of finding the right way of expressing it all. using the new tools, which is the AI and everything connected to it. So it's it's pretty awesome. So a year ago I was managing a team of 250 people doing delivery, doing like everything starting from pre-sales, design, mock-ups, technical scoping, and then subsequently the design, the developer, user user testing and all that stuff. and now I'm actually probably on the other side of it while I'm doing the the thing with the AI and testing it with people, testing it myself and kind of really reducing the the amount of people who are involved into building something real. So definitely a very, very exciting change of scale and change of career. Of sorts for me. And I feel this is super natural. I feel this is like where a lot of people are going actually, regardless of your technicality. Previously, I've seen a lot of examples of people being completely hands off and now all of a sudden like being super technical builders like in a day or or week's time. So yeah, pretty exciting.
Kevin Williams: that's that that's great to hear. and I think we're all sort of feeling it, that the line between what is a citizen developer slash vibe coder and what is an
Vlad: Definitely.
Kevin Williams: actual developer is is definitely blurring from a skill set perspective.
Vlad: Mm-hmm.
Kevin Williams: but maybe speak a little bit to the mindset around it though. I I'll say that that you know, any any Tom Dick or Harry can can subscribe to Codec Codex or have a Claude Code subscription and have an idea and start building something. But what
Vlad: Mm-hmm.
Kevin Williams: what do you think differentiates a a a quality approach to citizen development, again, keeping it for sort sort of mere mere mortals versus a more sophisticated approach?
Vlad: Yeah, that's a great question, actually. I think the same thing that did differentiate good coders from bad coders before, like the vision, the architecture, if you will, and the business acumen, meaning what is this thing is actually supposed to be doing? Because yeah, technicality is obviously an important component, but if we're speaking about the business and the product, it's all about whether it's doing its job and people like it and they want to pay for it or whatever. I mean, it's not a paid product, that's fine too. Whether it's accomplishing the goals that are set in front of that product or not. And I have a ton of experience working with these specific problems, meaning, I'm the person who was dealing with the angry customers who said this is completely different from what I had in mind, and you guys you know need to rework the whole thing from the ground up. Or the opposite side of things, dealing with a developer that tells me this is impossible because these requirements do not make any sense. They are like contradictory to one another. And this wouldn't work like at all if we build it the way the customer wants it to be built and wants it to be looking and working like. So I was, I'm very proficient at untangling these from both the perspectives, the technicality and the the actual business part of things. That's why I feel I'm right in the middle to be talking to the AI, basically in the same way I would be talking to my technical team, saying, hey, here's the objective, here's what we know. Here's what how I think we can hypothetically structure that system. Here's the components. I'm not sure about this one. To please provide the you know some ideas in terms of how we can facilitate this, but I'm like very sure that this is supposed to be the outcome. And it kind of gives the AI a good enough frame for it not to go wild at things. Same as it did for the developers, by the way. I mean, I think the approach is almost exactly the same. and and even the exceptions are pretty much similar. So meaning. some developers previously they didn't need that much of an instruction. They kind of understood this, they may have had some some more experience, some more perspective from the business side, and you didn't need to kind of like figure out every single thing with them before they built something meaningful. And the same with the AI. Sometimes it just gets it. It even brings up some things that you didn't think of or you know, that you didn't think were important, but actually he the the AI emphasizes them being important and it kind of does it. But a lot of times if you don't give it the right framework or the guardrails, that's the new word that's in our vocabulary now every day, it it does something crazy or it tries to do something in a way that wouldn't make sense for you, you know, at all. Without any technicality of it, to just like logically is supposed to be making sense or not. And you're you're kind of the person who is the judge of that. And I I've kinda feel like all my technical experience actually translates except that I now I'm yelling at the AI versus previously I've been yelling at people for actually the same same reasons exactly.
Kevin Williams: So, I mean, one that fully resonates. So I'm always telling people that 80% of the battle is in the planning and understanding what are the things you're trying to solve. And this is where I think senior business people have an advantage. Have enough miles under your belt to understand the business process and being able to now have the technical solution within grasp is the smaller part of the problem at the end of the day. I will push back just a little bit because I think
Vlad: Mm-hmm.
Kevin Williams: what I see is the cost of iteration is so much lower than it used to be that if you had
Vlad: Definitely.
Kevin Williams: gone down the wrong rabbit hole on your spec in you know three, four years ago, it could have been tens of thousands of dollars of development and time
Vlad: Yes, we'll
Kevin Williams: and all of these things. Whereas the speed to production and the speed to a minimum viable product that you can play with is such that to some degree, I'm I'm guilty of this as well. I
Vlad: Yeah.
Kevin Williams: do my 80% of the time planning, but then I get to this feeling of I'm close. Now I need to see what this looks like and play with it a little bit and and learn from it. And then maybe I scrap the whole thing. But at least
Vlad: Yeah. Mm-hmm. Mm-hmm.
Kevin Williams: in the current environment of tokens being cheap and You know, being able to mo run parallel processes and things like that, it can be okay to do that. W where it's not okay, and this is like a takeaway for for my, you know, my my hands-on keyboard people out there, is in security controls, the governance aspects, like are you creating risks in your organization, et cetera? But
Vlad: Yeah yeah yeah.
Kevin Williams: the feature set can move. You also said something kind of interesting here, which I I guess I knew, but Like the rest of us, you're clearly using AI to plan the projects. And in my world, I've built projects, you know, projects as an artifact within GPT
Vlad: Mm-hmm.
Kevin Williams: and within Cloud. And those projects are equipped with my standards, my documents, et cetera. There's a GitHub linkage and whatever, such that I can build on a consistent basis and do that
Vlad: Mm-hmm.
Kevin Williams: 80% of back and forth there, such I can get a spec. So What is your tooling when you're you're you're, you know, Kevin's the client and we're going to build something new? What does that look like? What do you turn to? What's your app? What's your setup, et cetera, to do that conceptual design?
Vlad: Right. for me, I'm a codex guy, open
Kevin Williams: Mm-hmm.
Vlad: AI. I am working with a few different ones as well just to keep myself informed about how things are developing, because you know, it there's always the best model that's being launched every day with a different provider, and then it gets to be the best right until the next one, which maybe within the same day. And I I actually saw the news that the cycle of the models is actually decreasing, meaning New things are being launched a lot faster than they have been previously. So it's a pretty interesting statistics. and if we look at the the latest launch of Chat GPT six and now Chat GPT GPT actually six point one, I think it has been like within a week or or two. yeah, yeah, yeah. So previously it has been like months for the the first
Kevin Williams: Days. Uh-huh.
Vlad: version of the GPT and then the subsequent ones. Anyways, back to the to the question GitHub all the way, for sure.
Kevin Williams: Mm.
Vlad: I'm using this at both the the repository to keep the code consistent and track the changes as well as the documentation. It is so much easier also to switch the models working with the same repository in case you need to. Like I've mentioned this today to you when you're running out of credits and you know you need to do something urgently and maybe you have a different model, like maybe Claude or maybe other ones that you want to hook up. If you do not have a repository with good docs, you're in for a world of pain because you you will be gambling on the fact whether AI will understand what is it that you were up to or not. Because if you do not have the context set up for the thing, it may just go wild. So
Kevin Williams: So so let me let me interject because you're one step removed from I think the piece that I'm after. That yes, absolutely. Like a friends out there on the internet, like use GitHub, have repositories, they're standardized. You can then you can then if you need to move between codecs or cloud code or grok or whatever you want to do, the repositories there, you can work from it. But go up a level as far as that planning
Vlad: Mm-hmm.
Kevin Williams: piece, like.
Vlad: Right.
Kevin Williams: Are you are you do you have like a planning project or do you do it directly into codecs? Do you reference a repository set?
Vlad: yeah.
Kevin Williams: Like what does that look like?
Vlad: For sure. So I think thank you for clarifying that. definitely like setting things up before you actually get to the building part. I think one of the great improvements overall in how AI works and specifically the latest models is how it works with the examples. It's fascinatingly good these days. For both like the visual aspect and the technical aspect of things. It can really get into whatever you're referring to as an example or even a partial example. And it helps you to do 80% of the legwork before you even get to the real quote unquote planning of the thing. but it also kind of brings up the the the the thing that you mentioned. It's so much easier to get to 80 slash 90% now in both the planning and the building. And sometimes, you know, actually, even even previously, obviously the the path from zero to eighty has been a lot longer, but it has been the easier one too. meaning that it's always the last 10% is where things start to crumble and and actually you discover the complexity when you're into that thing. But yeah, for planning for sure, the examples, the references, the screenshots. I'm also a professional yapper now, so I do not I mean, I do type certain things in, but mostly I just yap with my codecs all day, every day. My wife is like, Okay, what's going on here? Like, who are you talking to? And I'm like being emotional and everything. And I'm like, what is it that you're doing? Are you crazy? Are you how did you even think of that? And it's so much easier to interact with it like I would with a with a person. And I I think that, you know, I I I don't I c we can't say for sure, but I went I'm really mad with it not getting me. I'm start I'm starting to gaslight it. I'm like, dude, you know, my friend told me that Claude did this so much better. Like at
Kevin Williams: Yeah.
Vlad: one tenth of the speed. Could you could can you kind of let's beat that guy? I mean, let's prove you're better. And I can't say for sure because we don't see what's happening under the hood, but in my experience, I have been able to solve quite a few problems, especially in the last days when the mo the new smarter model came out. Much better. faster than I would without the the gaslighting thing. So I think it actually works. So it's like it's called like the
Kevin Williams: Ha ha ha.
Vlad: unofficial turbo mode. It's like when you gaslight codex. But you know, you have to be careful because if you are being too harsh, like my friend tried the same thing and he's like and he probably was using a a different way to formulate the thought. He's like, if you won't do this, I will I will go to Claude or something like that. And it actually said, okay, you know, fair enough. And it gave it the prompt. Like exactly. So you have to
Kevin Williams: Yeah take your bags and go.
Vlad: know when to stop, you have to be balanced. So you have to be a a you know, a professional L L gaslighter to use that feature. So yeah that
Kevin Williams: So Anthropic actually has an internal index for frustration and they track expletives. So
Vlad: Yeah.
Kevin Williams: there is a dashboard in Anthropic and they're probably a GPT as well that somebody looks
Vlad: Probably, yeah.
Kevin Williams: like as far as the sort of the per capita or per prompt incidence of cussing. And the more that people drop like the F bomb,
Vlad: Mm-hmm.
Kevin Williams: the more frustration there is in the system. And apparently that's part of how they figured out that GPT-5.0 was just like horrendous for
Vlad: Yeah.
Kevin Williams: a lot of people, is because the cussing just went off the charts. But I'm gonna push just
Vlad: Yeah, that's a good metric.
Kevin Williams: a little bit harder. Like, like as a dev, when you open this up, are you just opening up the Codex app, like the desktop app? you know, now with cursor is off of Codex land because of Grok and you know, et cetera.
Vlad: Yeah.
Kevin Williams: I imagine that's changed your behavior. But is your blank screen, is that like a project in Codecs that's like this is my project starting point?
Vlad: Exactly. So a new project, every time you actually are building something new, even if it's adjacent to something existing, I still prefer to build it as a new project because of just the context and some things that definitely should be separate. So our our friend actually doesn't do that and he doesn't get like the whole project thing. So it's it's all over the place where in his workspace, and I'm like always due to have to change that because you're confusing the hell out of everybody, yourself included. So definitely start with the project. I usually connect it directly with a GitHub instance. And most of the time, the things that I'm working also require a backend, a database, let's call it that. So I'm using Superbase for the most of my projects. Even before I actually am done with anything, I usually spin up the instance of Superbase. I connect it to GitHub. Both are free. And I also use Vercel for front end, which is also free up until very like late in the game. So it's costs you literally just the subscription to get. thing off the ground and actually publish it in front of people. but yeah the whole planning goes in there. So I always use plan mode when I'm kind of figuring things out. It's also a hack that I don't know it's pretty plain in sight, but I think some people are overlooking it. previously it has been more pronounced in terms of how the Chat GPT interface itself was saying hey you can plan things. whenever you mentioned plan in your prompt It actually highlighted okay, there's a planning feature. Have you heard that? Claude actually is being somewhat better and also somewhat more annoying with this whole thing. It it doesn't ask you whether you want to plan or not. It always kind of is in that mode. And that actually freaks me out sometimes because it's like, okay, it's dude, it's like so obvious. You didn't even need
Kevin Williams: Ha ha ha.
Vlad: to ask me. That's such a weird other option that you're proposing to me that I, you know, do we really need to ask? But I think field codex is a lot more balanced in that regard. So I plan things. Even if my idea is perfectly validated and calculated, I always start with the planning mode. And a lot of times codex does ask me good questions. Okay, how about this one? What about this edge case? Should we do it this or that? And you know, it kind of helps me both to think about this better myself, as well as for the codex to do a better job in the end of the day. So And yeah, so once once that is done, once the planning phase and the whole back and forth figuring out the questions is there, I just shoot it and we're off to the races.
Kevin Williams: So the this audience knows that I've been very Claude Code centric until just the last about six weeks or so. And now I am extremely codecs focused. but I do have both of them up, and part of it is just actually right this second. This for this show I've I stopped everything. I've learned hard lessons about having like agents running on my aging Mac M1 while doing a podcast. Like when you see when you see it freeze on YouTube, that's usually because Kevin's run out of processor. but you know going back and forth between them is interesting, and there are certain things that I'm very comfortable launching Claude Code for, but I'm finding the utility of Codex and its extensibility to just be so much better. And I don't want to go down this rabbit hole quite yet, but we we will in a minute. One is computer use and the ability for Codecs to independently be able to do things like configure Google Cloud. you know, you wanna create an API, it's gonna make it so. Like it's kind of amazing. But but from a more pedestrian perspective, there are a couple of little features that I I'm curious how much you use. So one
Vlad: Mm-hmm.
Kevin Williams: in Codecs is the side chat. So my MO and Claude Code, it's sort of similar to what you were describing. I would do my planning, but I would have this project. in Cloud browser or in the chat function that
Vlad: Yeah.
Kevin Williams: was connected to my GitHub repositories and then I could plan there and I'd kind of bounce between the two. Codex has a side chat functionality and it's not brilliant in that it doesn't know your whole GPT universe, but it can look at what you're doing and it can dispatch agents and it can do things on your behalf. So
Vlad: Mm-hmm.
Kevin Williams: that's one I use very, very frequently. But there are also a couple of other features in there, like there's a browser feature, there's a preview feature, et cetera.
Vlad: Yeah. Mm-hmm.
Kevin Williams: Maybe talk to how you think about those those ancillary new features or newer features for codex.
Vlad: Right. So I'm actually not using this side chat feature all that much, exactly for the reason that you mentioned. I I love to be more in control about the context. And usually when I'm asking something, it's related to something else that I previously asked, or I just want to play around a little bit. So rarely do I have the question that is completely disconnected from like everything. And if I do, I I usually do need to do something with the output of that query. So I just do it directly in codex. And one
Kevin Williams: Okay.
Vlad: of the things that contributed to me kind of not really using these all that much is the about a month ago, maybe a couple months ago, there was like a major update in terms of the codex and the chat GPT app. Previously those were separate applications, and Chat
Kevin Williams: Mm-hmm.
Vlad: GPT as a desktop application was used by a hundred million people or whatever, a lot of folks, and the browser version, while codex was used by just five by five million, I think. And pro mostly there were developers or People who were curious about the building part of things. And I think they kind of tried to flip the switch on that and let more people be familiarized with the power of codecs and building. This is where they did the chat work thing. And I honestly I understand why they did this, but I don't love the change because it kind of
Kevin Williams: Mm.
Vlad: it messes things up, it kind of puts together things that I'm working on and things that I'm just you know want to talk about. together and then I mean I I get the utility, but I just want to keep it separate. So whenever I just want to talk about something or do a research or query or like do a plan, I usually probably go to my browser, same as you do with your claude.
Kevin Williams: Mm.
Vlad: But when I'm actually into the building mode, I I go back to my codex and I'm just I'm I keep on using it just the way it was before. So maybe it was just, you know, me being rigid and non-flexible with the whole change thing. And I maybe I'll I'll it will grow on me, but so far I'm kind of keeping things separate. and yeah but in terms of the The other like features, the browser thing, I'm using it all the time. I'm always
Kevin Williams: Okay.
Vlad: using it to because I do a lot of front end work as well as the back end work.
Kevin Williams: Mm-hmm.
Vlad: Before I publish it actually to the front end, I start a lot of things on my local host and I use that. That is actually very helpful. So whenever I'm in the build mode, there's like the whole context right in front of me. And I said, okay, before you push this any further, open this in the side tab, show it to me how it looks like, validate this and this and that, and I kind of If something is off or there's like a big red error like all over the screen, I can just directly in there looking at it, say, hey, here's what happened. Let's figure it out. This is super convenient and helpful. So you and for this, actually, I I see this as very valuable because it keeps like the context shared, visible, and from both like a user perspective and also the system perspective. So it's definitely a winner. But I think mostly old data coding agents do that. I Cloud Code also does pretty much the same thing.
Kevin Williams: So a c a couple of different observations. and I think we're we're logically going to get into new agent forms and dot in just a second, because
Vlad: Mm-hmm.
Kevin Williams: of the but one is it's very obvious to me that that the the platform universes are drawing together, and pretty soon you will have that shared context, like across different apps and applications and things like that, that I think will make that side chat a lot more useful. the other corollary is in just business productivity. Claude is quite strong in Microsoft. it's actually, I think it's quite a bit better than GPT in the Microsoft apps, but you could have multiple instant multiple instances up and they're all baby clods. They can't actually talk to each other. They don't have well, they can talk to each other from a document perspective, but
Vlad: Mm-hmm.
Kevin Williams: they can't actually reference another chat. So that stuff's all happening and and This is sort of perfect to queue up what happened even just this week or in the last little bit with these new agent forms. And Matt and I talked about this a little bit last week. with Grokbot and then playing with Meta's Muse. personally,
Vlad: Yeah.
Kevin Williams: I I tell you, hey audience, you want an unlock, you want to earn your time with us. I set Muse off. So this is Meta's product. It's very friendly. we put a link in the last show with a with with like a Code for a billion tokens. We'll put it in this one as well, just to experiment with it. But I sent it to my auto insurance and had it log in to GICO, pull down my policy documents, and then had it go out and do comparative quotes. And I felt like a total schmuck because apparently I was leaving thousands of dollars on the table. And it literally saved me $3,000 on my auto insurance. I sound like an ad, but like for for no work. Like I I
Vlad: Yeah.
Kevin Williams: I'm gonna save three thousand dollars this year at a church. Thank you very much. I have never said thank you to Mark Zuckerberg ever in my life, but thank you, Mark Zuckerberg. That's a first, not my favorite person.
Vlad: Yeah, yeah, yeah.
Kevin Williams: so okay. And then just three days ago, OpenAI had its big dev day, whatever the heck they call
Vlad: Mm-hmm.
Kevin Williams: it, and they they launched a couple of things. we're probably not gonna go into spaces right here, but that the maybe that's something for Matt and I to cover in the future. I think that's interesting. But the killer unlocked for a lot of us was GPT's version of I'm gonna call it its version of Grockbot versus its version of Muse because they are a little bit different in their functionality. And you and I were talking about this a little bit pre-show, and
Vlad: Mm-hmm.
Kevin Williams: I am completely freaking obsessed with this thing in
Vlad: Me too.
Kevin Williams: that it is and it's a different form factor. Than I'm accustomed to. And I I'm it's kind of wrecking a lot of my workflows. And now I have so many different things that are going on at once. Like it's absurd. I have like literally 25 different subagents that are working on different things. I'll pause right now. actually
Vlad: Yeah.
Kevin Williams: it's in the cloud, so I they don't actually have to be paused. They could they could be doing their thing. So maybe describe
Vlad: Right. Just being careful.
Kevin Williams: a little bit on on on what what dot is and What the heck is the name dot? But whatever. It's fine. Every time I I
Vlad: Yeah.
Kevin Williams: try a transcript dot, by the way, using WhisperFlow, it puts a period in place, which you say I have to spell it. I'm like,
Vlad: Yeah, not very AI transcript friendly. Yeah.
Kevin Williams: like, come on, guys. Like you you need to be AI friendly with your naming. So there are lot of periods that are around that represent
Vlad: Yeah.
Kevin Williams: dot as I've been talking about it. But anyway, explain to us a little bit what dot is and your initial impressions and how you're using it.
Vlad: Right. In Lame Wayne terms, essentially that's your virtual assistant. So think about the task that you would naturally give to your assistant. Like like you said, check out the insurance, look through the docs, check out my mail and calendar for the the meeting spots or the opportunities or what I said that I do and I didn't do, and and stuff like that. So I mean obviously there's a lot more to it, but that's probably the easiest way to describe the The primary utility of this thing. So for me, the the the the aha moment was that I I didn't like set it up or something. It just came with the new update and I just you know gave it a name and I left it at that. And my Chat GBT had the access to my mail, to my account, or to a couple of other things that I'm using. And you know, out of nowhere it just pops in and says, Hey, you had a call yesterday with this guy and You said that you're gonna invite that person to that meeting today, and I don't see a new person on that invite. Are you sure you you're still on track here? I'm like, dude, I didn't even ask for it. Like,
Kevin Williams: Ha ha ha.
Vlad: thank you. I appreciate it. I mean, that was a small gesture, but it was so thoughtful. I don't know, it probably it it it's one of the things that it kind of came out from experience and it didn't just, you know, happen. And maybe just like somebody else training it to do that kind of thing, but it was like super, super convenient. And the the two hours later. It pops into my inbox and said, Hey, I saw a couple of bills in your inbox with a due date of today. just making sure you took care of this ones. I'm like, Wow, wow, wow. I mean, I appreciate this. you know, whatever. I mean, it's like naturally out of the box without you needing to do anything at all. It acts as your assistant. You can obviously train it and explain it and then set up the routines to do a lot more. but you know, this is essentially in TLDR. This is what the OpenClaw was supposed to be.
Kevin Williams: Mm-hmm.
Vlad: And this is what people expected from OpenClaw to be like a long time ago. And it it's sort of like it's a bit of a different breed. Obviously, people will say OpenClaw is a lot more configurable and flexible. You can set up so much more in OpenClaw, and I agree. I use OpenClaw every day. but this thing is I I heard this description of people saying this about the the grok butt, but I think it's very applicable. Here as well. It's like if Apple would do open claw, this is how it would look and work like. So very hands-off. It kind of knows the best way to do things, needs very little guidance. You can set it up the way you want to, but you don't need to like get into the code or under the hood or like explain every part of the way for it not to crumble down in the first unexpected scenario, stuff like that. So yeah, that's that's pretty cool. I'm I'm very impressed. And basically this is like a AI agent that is with you that works in the same place. You know, speaking about the previous topic of having the shared context between your like working space and your personal like chats and and whatnot. So this thing is everywhere. So it knows exactly what's going on. So it you can talk to it like with a person who would be like literally sitting next to you and that is aware about what's going on with both your like life or whatever else you're exposing to Chat GPT, so which is very convenient.
Kevin Williams: So I've been calling it just that promise of an executive assistant. So it is
Vlad: Yeah, yeah, yeah, exactly.
Kevin Williams: you know, it's the Jarvis idea from Iron Man. And
Vlad: Yeah.
Kevin Williams: just a I'm gonna wrap on a couple of other things that maybe you haven't touched yet. one is it has this very odd interface for an audio function. It's actually a phone call, and I can't really figure out why it's a phone call, but it
Vlad: Yeah. I don't know.
Kevin Williams: will trigger it. And I found it to be too slow, so I had it speed up. But most
Vlad: Mm-hmm.
Kevin Williams: mornings now I'm I'm kicking back in in in our sponsored chair, the the Liber Novo chair, which you know, they they they kindly sent us. And I'm just talking to it and having a conversation as if I were having a conversation with Jarvis of like bouncing ideas off of it and whatnot. And that's that's an interesting experience. I'm not sure that it's fast enough for me in terms of
Vlad: Mm-hmm.
Kevin Williams: like it has to it the the latency is isn't isn't long, that's not the issue, but I want to be doing 10 things at once. And
Vlad: Right. Yeah.
Kevin Williams: really that that ends you end up on a single track, but it's it's a pretty cool experience. The
Vlad: For sure.
Kevin Williams: parts that I stumbled on that I found really novel and I can't figure out quite yet as far as my workflow is it's actually directing codecs for me. And it started doing that sort of unexpectedly. where I was I basically was recognized that one of my apps it was a like a marketing dashboard and it had a YouTube connection and the MCP wasn't working the way it should be. And I basically made myself a to do. And as opposed to just making itself a to do, it made it a to done as in it just like it it just like launched a codec session
Vlad: Yeah. Just did it.
Kevin Williams: and it went off and it fixed it. And then
Vlad: Mm-hmm.
Kevin Williams: it took over my browser. And did the confusion the the not nothing the the connections that it needed to make from a browser perspective as I was doing something else. And I went whole hog into it yesterday and like the amount of stuff that I did, I I I I mean, okay, we're like all productivity n ninjas and we're like super involved in this. I did a month's worth of work yesterday. Like I
Vlad: That's pretty cool.
Kevin Williams: absolutely did a month's worth of work. And I like I it it's just chewing on it just happily. I have no idea how expensive this is going to be. Like at the moment, they're clearly giving you some room to play with it because my
Vlad: Yeah. For sure. Yeah.
Kevin Williams: my my I think I'm down to like 17% on my week or something like that. But given the amount of power use I've tossed at it, I think I think they're subsidizing it just so they can give people like experiences around it.
Vlad: For sure. The whole thing. Yeah.
Kevin Williams: but it's a different experience. And and you know, previously GPT had had agents, which it's funny, I'm looking at my browser GPT, and agents are still sitting there in the left bar, but the agent function in the desktop app has actually gone away. And I think that they're mol they're like melding in agents into this new dot type functionality. And what I I more than think, I know that's what's going to happen next is going to be an open clause style orchestration layer where your main dot. So right now everybody only has one dot, and that main
Vlad: Mm-hmm.
Kevin Williams: dot does everything from managing, in my case, like my kids' social calendar to using codecs to connecting to email and parsing and whatever. But it's pretty obvious that you're soon going to have multiple dots. You're going to have like you're going to assign a dot that is like your dev dot. And your main
Vlad: Mm-hmm.
Kevin Williams: dot will talk to the dev dot and be like checking in on it through the day through heartbeats. And it will start managing that on your own and on its own. And I just I honestly I'm not quite sure what to make of this because the The productivity jump is just it's so big. Like it's the biggest
Vlad: Mm-hmm.
Kevin Williams: productivity jump that I've seen yet in all of this. and as more and more people start using it, I I I I I just don't even know what what work it's gonna start to look like, to be honest.
Vlad: Yes. this is a question that we all ask every day. I wonder if they actually reuse some of the technology and and the knowledge and the the thinking of the person that did create the open claw in the first place because they hired the guy like six months ago. And and everybody
Kevin Williams: yeah, of course they did. Yeah. Yeah.
Vlad: was like all excited about this, but then it kind of went silent. He went on a few podcasts and whatnot and six months later or whatever that is. Yeah, yeah, exactly. And
Kevin Williams: Took his billion dollars and
Vlad: And but looks like he was busy or or people that, you know, took over whatever he provided were busy. So this is looking pretty good.
Kevin Williams: It it is a I have to say it is a little bit frustrating as this marches around. Not necessarily frustrating, but the evolution is so fast that you
Vlad: Mm-hmm.
Kevin Williams: establish like say primitives around what I was talking about with, you know, my approach. I'm I'm still, even after our conversation, I'm still pretty comfortable with my approach to to dev design. I don't think it's very different from yours. So that was very validating for me. but it is like if I were to go into a fresh person. If I was to go to
Vlad: Mm-hmm.
Kevin Williams: my 17 year old and be like, all right, this is the way that you should build. I'm not sure that I would necessarily build all of those building blocks today. It might be a little bit more just conversational and like there
Vlad: Yeah.
Kevin Williams: may be a there may be like a a skipping here that a lot of people are going to get the benefit of.
Vlad: Mm-hmm.
Kevin Williams: And yes, you and I and people listening to this show have a lot of hard fought and sometimes expensive lessons that I think are going to serve us in good stead going forward. But all of this is clearly just going to get much more accessible as time goes on.
Vlad: Definitely. I mean for me the the change with the the whole like call your your dot thing wasn't that radical because I I'm telling you, I've been yapping with my codecs and all my models for the last two, three months at least because it I I feel like I convey so much more and so much faster than I would do in typing. I mean I I've been speaking with my developer friends and they're like, No, no, no, we actually stick to typing. We have like the feeling of control or whatever. I'm like, dude, let it go. I mean, this is I mean we can so much better like articulate what we actually have in mind. We can I mean for for conciseness and for actually texting a person, I'm all for the text messages. I I'm not really a fan of voice messages and all that. I'm that guy who ignores your voice messages or transcribes them in instead of listening. Yeah, yeah, definitely. That is the voice message guy and but luckily
Kevin Williams: Yeah, yeah, tell Matt that. Matt is a bo Matt is a voice message guy.
Vlad: for for me and for the rest of the folks like me. The transcribes works pretty good. I mean, these days, so you don't have to worry too much about that. But
Kevin Williams: I didn't want to I don't want to say that I I am 100% dictation. I'm over two million words on
Vlad: Yeah. Yeah, for sure.
Kevin Williams: on WhisperFlow. And and I joke that like when somebody makes me type, I type like a cat with peanut butter on its paws. I'm like, I don't want to do
Vlad: Yeah.
Kevin Williams: this. So the conversational aspect I think is well established. that I'm fine with. So I've always been been managing that. I think it's the it's this this orchestration. the the lower friction in dot
Vlad: Yeah. That's new.
Kevin Williams: style or grokbot style orchestration than we had with Hermes or or or open claw and whatever. And again, you know, hard fought systems, they do what they're supposed to do, like they're great, they're highly configurable, whatever. Kinda so I actually haven't used Hermes and a and a buddy of mine was was like, when are you gonna give Hermes a try? And I literally responded with never. Not going to because
Vlad: Yeah.
Kevin Williams: There's there's really like, okay, I get it, and there there are there's a limit to the stuff that I can experiment with. but sort of waiting for the easiest technology, the most accessible technology, I think is gonna be the right bet for a lot of people. And we're right
Vlad: For sure.
Kevin Williams: at the cusp of it.
Vlad: Yeah, I mean, unless you have like an urge and or an urgent requirement to do things right now. So we just took the open claw and built one of the features that we had to have to facilitate certain like interactivity with our system. We just took what we we knew was available. And probably there were a couple of alternatives, including Hermes, that were already available at that point. But it did the job, it works fine, so no complaints. But in terms of which technology to invest in, I mean it's getting scary. As I said, the the the the development cycles between the next best thing and and the other next best thing are getting super short. I mean, like literally within a week. So you don't even know. So it's like, whoa, what's going on here? So should I even get started?
Kevin Williams: Well that was a comment from OpenAI Dev Day, which was they're six months ahead on their own feature production because they're using their own tools. And they themselves are amazed. They're like, Wow, that turned out to be really easy to do because we were just using Codecs with a billion tokens a minute. you know, using Codecs eight point or whatever the heck they're using internally to to to
Vlad: Whatever they're using, yeah, for sure.
Kevin Williams: to do the thing and just, you know, rampaging ahead. Yeah. Okay. So the
Vlad: You never know with these guys.
Kevin Williams: so the the the piece we we we alluded to a little while ago, but I d I bet that most of the audience has not really leaned into computer use. So computer use has been around for a while as as far as the ability to move files around, access your downloads file, access your desk desktop, et cetera. Totally fine. one of the big unlocks for for most of us with Astra in particular. And now Sol 6.1, et cetera,
Vlad: Mm-hmm.
Kevin Williams: is computer use as in the ability for the LLM to use your browser, to interact with your computer, et cetera. And the example I gave earlier was like interacting with Google Cloud, which anybody who's had to set up like APIs in Google Cloud knows that it is the most Google of Google environments, which means it's really finicky and it has a lot of like
Vlad: Yeah, for sure.
Kevin Williams: levers and dials and things that you have to deal with. And now at this point, I'm just like make it so, like straight out of Star
Vlad: Mm-hmm.
Kevin Williams: Trek. That's really my thing. I literally say make it so because I'm a dork. but it's like,
Vlad: Yeah.
Kevin Williams: you know, like talk to that a little bit because it is a little terrifying to like let it loose that way. But I've had a couple of weird experiences, but by and large, it just freaking works. Like I launched
Vlad: Yeah.
Kevin Williams: a A hundred ad meta campaign yesterday, like h completely hands off with a mixture between the MCP and you know, using computer control.
Vlad: Yeah, for sure. I mean, definitely I I think the most staggering thing about this whole development is like how far and how how much this has developed from from day one. Because as as you said, the computer use is one of the actually the the early features of AI. And I think this is the one feature that most people who are not necessarily doing this every day were expecting. They were thinking that AI is gonna come like literally at their computer, whatever they're using. take their software and just do this thing better, faster, no mm, making no mistakes and kind of like 10x the outcome without changing any of the foundational work. We have been into quite a few conversations with customers, and this literally what they wanted. So when we said AI consulting, they began by showing us their random software from 2004 that barely runs and crashes, that has its own local database that doesn't know what cloud is, and you know. And it provides some sort of like a weird CSV that is the basis of their business. and I'm not trying to kind of poo-poo on them. Those are people from the industries which are not like very tech enabled from the construction and whatnot. And they said, Okay, guys, you probably need a new app because what you need, what you're describing as a problem, is not necessarily an AI problem. That's just a software problem. So your software is
Kevin Williams: Mm-hmm.
Vlad: outdated and your your needs are completely different from what it does. And they're like, no, no, no. We actually love our software. We just want the AI to do this for us in that program from 2004. And like, okay, we're gonna have a problem here. So let's build a harness around it and whatever. So we kind of like were I was we were sidetracking that conversation to go to a different place. Let's start somewhere, and then slowly the plan was to convince them to actually redo the whole thing because did they did not it do AI. But if we would have that conversation today, I said. I would probably say sure, we can probably give it a shot because as you said, Astra and and the the new generation of the LMs is pretty damn good at this and they're not spending all that much tokens on the whole thing. So in in a lot of cases it's actually a better strategy to try and just, you know, give it your computer for a ride and w rather than building the whole like piece of software. Looking for MCPs, configuring CLIs and you know, dealing with expired keys and whatever. I mean, just let it let it take the wheel. And the last bit of it is that it's just I think maybe this is kind of the the first thing also about it, is because it's getting so much better, it's actually much more efficient. I mean previously I think Claude was a lot better and a lot more advanced in in the browser use and the computer use. But one problem that I continuously ran into is that it spent my whole credit limit like on a super simple operation, like literally within tens of minutes. Like I worked for 15 minutes, claw does the computer use for me, and then it expired. I'm like, dude, I'm I'm
Kevin Williams: Yep.
Vlad: not I'm jump what's going on here. So and now it actually can work for hours and apparently it stays still under limit, which tells me that it's probably just a lot more efficient. So faster, cheaper, just Gets the stuff done.
Kevin Williams: So I think there's a point here that's a little bit under the hood, which is under the faster and cheaper bid and the exigentic functions. so Muse. If you want to strip back what Muse is, you know, we're not supposed to talk about models on this show, but Muse is built on Meta's Spark 1.3 model, which is a thoroughly okay model. It is not a dumb model, but it's not a frontier model. Muse
Vlad: Mm-hmm.
Kevin Williams: is freaking brilliant. And when you interact with it, it's basically batting outside of its weight class. for its model capability, because the model has been tuned to this sort of activity in a way that it doesn't need the frontier intelligence. I think
Vlad: Mm-hmm.
Kevin Williams: Opus 5.5 or even Luna 5.5, not Luna Luna's on the I'm gonna confuse myself, Sonnet 5.5 was was yesterday. Like these are being tuned for efficiency and specific use. And you're going to start seeing tuning around computer use as well that's going to
Vlad: Yeah, definitely.
Kevin Williams: make them much more efficient. And I I'm fascinated by that because, particularly when you add like decision models, like a JEV in, and there's actually a new decision model that's clearly a JEV ripoff for for open AI, etc. We're in a really, really good place. And I I have been an optimist on pricing for a long time, sort of a dubious optimist. And because everybody's like, the subsidy era is gonna end and all of this stuff's gonna get heinously expensive in a hurry. And you know, we were talking before this about what happens when you go off of your plan and you use tokens and it's like crazy expensive. But
Vlad: Yeah. Not a good place. Yes.
Kevin Williams: There's a there's a balance for the model companies, for the frontier companies of adoption and practical use versus the ROI. And they
Vlad: Mm-hmm.
Kevin Williams: know they're not dumb. They know that if they increase these prices too fast, they're gonna chase people away from things like being able to use your dot for you know hours and hours and hours and enculturating
Vlad: Mm-hmm.
Kevin Williams: that into your business. And I I think that there's an argument to be made that. People are once it gets a little bit more advanced, people are going to be reasonably willing to pay rates of like $500 a month for a subscription. And there is a new $500 a month GPT subscription out there. And I could probably justify that, but if it was $5,000 a month, it's going to be very
Vlad: Mm-hmm. Yeah.
Kevin Williams: hard for small businesses or individuals to justify it. And I it's going to r sort of stunt the market. So these
Vlad: Mm-hmm.
Kevin Williams: efficiency efforts that are happening under the hood. That are orienting to like least common denominator models, I think is what's going to give us the staying power. Like it may not be quite as fast, it may not be quite as brilliant, but if you're not doing protein folding, like you don't need that brilliance so much. And that that's going to keep costs under reasonable control, particularly when combined with market pressures and you know their need to keep adoption high.
Vlad: Definitely. There's also this whole big thing going on with the the open source models. So there's a whole like war in that specific sector that it's apparently building up for that the frontier companies such as OpenAI Anthropic, obviously, Groc maybe to a certain extent, they d do not openly voice their concern about the specifically the open source models, but it's very obvious that this is what they're Trying to kind of throttle or at least diminish or like hold back a little.
Kevin Williams: Mm-hmm.
Vlad: Because one of the options of how this can
Kevin Williams: Yeah.
Vlad: unfold is that we all know, and this is not a secret, that it costs a lot more for OpenAI or Anthropic to run their models than what you're reasonably able to pay for it. So I I somewhere I heard that you can Theoretically, I mean not everybody does that obviously, but you can theoretically use up to twenty thousand dollars of company costs for the the cost of the two hundred dollar subscription. like I don't know to which extent that is true. That's probably hypothetical, but I'm I'm guessing it's like somewhere out there. and th that's probably one of the reasons why are they also like upping their amount. And I think Anthropic has also been playing around with the amount of credits and subscription costs a little also for that purpose. Obviously the They wanna grow, but you know, it's it it it stinks, so they burn a lot of money. So
Kevin Williams: So just super quick, because we're we're gonna run out of time here and I wanna I wanna get into MCPs for just a second, but what is your go-to open source model?
Vlad: I'm yet to pick one, honestly.
Kevin Williams: Okay.
Vlad: yeah. Because I think I'm kind of I'm too busy playing with the new things that are being launched by OpenAI and Anthropic every day. But to one of your points made previously, I also had a M1 up until last week and it was it was getting to me because it was getting a little slow. So I actually got myself a new Mac that is very much local LM ready. It has forty eight. gigs of RAM, the M4 Pro. So I'm bringing on. I can probably launch OpenAI here locally, if it would be available. But you know, whatever is the next thing, I'm ready for it and I'm definitely gonna take it for
Kevin Williams: So there's a there's a whole rabbit hole there with outsource processors as well
Vlad: For sure.
Kevin Williams: of that that people are doing. But I I think in sort of the last bit of technical bits I'd love to leave people with are the the increasing flexibility of MCPs. And MCPs are model context protocols. And these are basically flexible gateways to sources of data. And they're so easy to spin up now. And I don't think that people really realize that. That if you have a data store somewhere that's accessible, slapping an MCP on the front of it, it you can do this with greater or lesser efficiency. But you know, people need to know that they can do that. Like I think one that that is just anybody listening should do is anybody listening to this show is probably doing a lot of transcripts. And your transcript provider probably offers an MCP that you can connect to and you can connect it to Claude and then be like, hey, remind me of that last conversation I had with Vlad and It it will pop up. But I encourage people to retain their own transcripts because it's one of the pieces of a business that is just so valuable. And
Vlad: Mm-hmm.
Kevin Williams: being able to put an MCP on top of it and basically using MCPs to collect information. So, like something I was working on yesterday with Dot. was again the marketing dashboard and that my my internal marketing dashboard has probably eight or nine different APIs or MCPs that are connecting to it and it's pulling in all of the data from all of the sources. I can put an MCP on top of that entire app that has the collected data and the collected intelligence of what's going on with my ads and what's going on with SEO and AEO and whatever. And not have to rebuild all of those pieces individually on the like the chat bot side. So I guess sort of speak a little bit to how you are you are deploying MCPs. And there's this this idea of interrelation between MCPs that I don't think a lot of people are thinking about yet, that they think of them as a single point. I'm gonna go get my, I'm gonna use a connection to Google, it's probably an API connection to get SEO data from Search Console. I'm gonna pull that one piece in. But the inner relation between all of those pieces of data is like the next step of complexity.
Vlad: For sure. And I think that honestly, my take on this whole MCP thing is that we're getting on a track where most people will not even realize that they're using an MCP. And they probably shouldn't care at all about that. But you know, all these fancy terms like okay, spin up your open AI, the droplet, the whatever. That's
Kevin Williams: Yeah.
Vlad: most people are just gonna be saying, I I just say, you know. I want it so, and then it it it's off to the races and it's just getting things done. So I this is sort of how I work with the MCPs even now. So I obviously connected to whenever and whatever I'm using and interacting with the most, like super base for my transcript. I use Fireflies. I I'm you know, all day, every day connected to my MCP. This is how the dot reminded me about that guy not being invited to the meeting from today because it pulled my transcript from Fireflies and kind of connected the dots. No pun intended. So and but it's a a lot of times I'm actually not using MCP because of for example I'm you I'm working with multiple instances of the one thing like for example multiple instances of super base and the MCP like out of the box usually only lets you connect to one. So again my specific use case probably for most people it's not even going to be relevant but there are so much more ways to access certain data. And sometimes AI is just able to figure this out on your own. So you and I'm I'm sure that a lot of people may be using MCPs or CLIs or APIs without even realizing at what exactly did they use at this particular moment. As long as it gets things done, it probably doesn't even matter. So I guess that, you know, definitely it's a powerful tool. Definitely should be out there and you should know that this is a thing. And if not, you know, ask your agent, it will describe it to you. But I think very soon, may I dunno, very soon nowadays is maybe two weeks from now or a month from now, this
Kevin Williams: Yeah.
Vlad: whole thing may not be even be relevant. It will just, you know, work. So I don't know if it's a good thing or a bad thing, but we're definitely moving in that direction.
Kevin Williams: Well, Vlad, this has been great. I seriously being able to geek out a little bit more on this. I really enjoyed it. we should we should kick Matt off occasionally so you can you you and I can can can do this. let him roam. Let him roam. It's all right.
Vlad: Yeah. Let him roam. So yeah, the man likes to roam. Yeah.
Kevin Williams: but yeah, sir, if people wanted to I don't know, do you have socials or anywhere where people might be able to follow you? If you're not into that thing, don't worry about it.
Vlad: Yeah, I'm on Twitter. I'm on LinkedIn. Yeah, yeah. So yeah, hit me up. I'm I'm happy to geek out more about all this whole thing.
Kevin Williams: Awesome, man. All right. We'll talk soon.
Vlad: Thanks, Kevin. Talk to you later.
Vlad: Thanks, Kevin. Yeah, here's hoping that this all comes amounts to something. So for sure. Hey everybody. my name is Vlad. I have been in tech for the last almost 20 years in various roles and applications. I do have a technical background, although it is somewhat diverse in the sense that I have been a software engineer back in the day, actually around 15 years ago. Then I transitioned mostly to the project management and technical project management. And for the last ten years or so, I've been mostly managing teams, engineers, and other various aspects of the business, but probably more focused on the technical perspective of things. That's why it's actually a very one of the most seamless transitions probably there is out there for people who are not necessarily like hands-on developers all all day, every day, but are closer to that part on the spectrum of The technicality or whatever you want to call it. So definitely these days I am what people would call an AI developer for sure. but this is actually very recent. This is a new career branch for me. So, and I feel like this is exactly a natural outcome of our experience, our aspirations, our preferences, kind of finding the right way of expressing it all. using the new tools, which is the AI and everything connected to it. So it's it's pretty awesome. So a year ago I was managing a team of 250 people doing delivery, doing like everything starting from pre-sales, design, mock-ups, technical scoping, and then subsequently the design, the developer, user user testing and all that stuff. and now I'm actually probably on the other side of it while I'm doing the the thing with the AI and testing it with people, testing it myself and kind of really reducing the the amount of people who are involved into building something real. So definitely a very, very exciting change of scale and change of career. Of sorts for me. And I feel this is super natural. I feel this is like where a lot of people are going actually, regardless of your technicality. Previously, I've seen a lot of examples of people being completely hands off and now all of a sudden like being super technical builders like in a day or or week's time. So yeah, pretty exciting.
Kevin Williams: that's that that's great to hear. and I think we're all sort of feeling it, that the line between what is a citizen developer slash vibe coder and what is an
Vlad: Definitely.
Kevin Williams: actual developer is is definitely blurring from a skill set perspective.
Vlad: Mm-hmm.
Kevin Williams: but maybe speak a little bit to the mindset around it though. I I'll say that that you know, any any Tom Dick or Harry can can subscribe to Codec Codex or have a Claude Code subscription and have an idea and start building something. But what
Vlad: Mm-hmm.
Kevin Williams: what do you think differentiates a a a quality approach to citizen development, again, keeping it for sort sort of mere mere mortals versus a more sophisticated approach?
Vlad: Yeah, that's a great question, actually. I think the same thing that did differentiate good coders from bad coders before, like the vision, the architecture, if you will, and the business acumen, meaning what is this thing is actually supposed to be doing? Because yeah, technicality is obviously an important component, but if we're speaking about the business and the product, it's all about whether it's doing its job and people like it and they want to pay for it or whatever. I mean, it's not a paid product, that's fine too. Whether it's accomplishing the goals that are set in front of that product or not. And I have a ton of experience working with these specific problems, meaning, I'm the person who was dealing with the angry customers who said this is completely different from what I had in mind, and you guys you know need to rework the whole thing from the ground up. Or the opposite side of things, dealing with a developer that tells me this is impossible because these requirements do not make any sense. They are like contradictory to one another. And this wouldn't work like at all if we build it the way the customer wants it to be built and wants it to be looking and working like. So I was, I'm very proficient at untangling these from both the perspectives, the technicality and the the actual business part of things. That's why I feel I'm right in the middle to be talking to the AI, basically in the same way I would be talking to my technical team, saying, hey, here's the objective, here's what we know. Here's what how I think we can hypothetically structure that system. Here's the components. I'm not sure about this one. To please provide the you know some ideas in terms of how we can facilitate this, but I'm like very sure that this is supposed to be the outcome. And it kind of gives the AI a good enough frame for it not to go wild at things. Same as it did for the developers, by the way. I mean, I think the approach is almost exactly the same. and and even the exceptions are pretty much similar. So meaning. some developers previously they didn't need that much of an instruction. They kind of understood this, they may have had some some more experience, some more perspective from the business side, and you didn't need to kind of like figure out every single thing with them before they built something meaningful. And the same with the AI. Sometimes it just gets it. It even brings up some things that you didn't think of or you know, that you didn't think were important, but actually he the the AI emphasizes them being important and it kind of does it. But a lot of times if you don't give it the right framework or the guardrails, that's the new word that's in our vocabulary now every day, it it does something crazy or it tries to do something in a way that wouldn't make sense for you, you know, at all. Without any technicality of it, to just like logically is supposed to be making sense or not. And you're you're kind of the person who is the judge of that. And I I've kinda feel like all my technical experience actually translates except that I now I'm yelling at the AI versus previously I've been yelling at people for actually the same same reasons exactly.
Kevin Williams: So, I mean, one that fully resonates. So I'm always telling people that 80% of the battle is in the planning and understanding what are the things you're trying to solve. And this is where I think senior business people have an advantage. Have enough miles under your belt to understand the business process and being able to now have the technical solution within grasp is the smaller part of the problem at the end of the day. I will push back just a little bit because I think
Vlad: Mm-hmm.
Kevin Williams: what I see is the cost of iteration is so much lower than it used to be that if you had
Vlad: Definitely.
Kevin Williams: gone down the wrong rabbit hole on your spec in you know three, four years ago, it could have been tens of thousands of dollars of development and time
Vlad: Yes, we'll
Kevin Williams: and all of these things. Whereas the speed to production and the speed to a minimum viable product that you can play with is such that to some degree, I'm I'm guilty of this as well. I
Vlad: Yeah.
Kevin Williams: do my 80% of the time planning, but then I get to this feeling of I'm close. Now I need to see what this looks like and play with it a little bit and and learn from it. And then maybe I scrap the whole thing. But at least
Vlad: Yeah. Mm-hmm. Mm-hmm.
Kevin Williams: in the current environment of tokens being cheap and You know, being able to mo run parallel processes and things like that, it can be okay to do that. W where it's not okay, and this is like a takeaway for for my, you know, my my hands-on keyboard people out there, is in security controls, the governance aspects, like are you creating risks in your organization, et cetera? But
Vlad: Yeah yeah yeah.
Kevin Williams: the feature set can move. You also said something kind of interesting here, which I I guess I knew, but Like the rest of us, you're clearly using AI to plan the projects. And in my world, I've built projects, you know, projects as an artifact within GPT
Vlad: Mm-hmm.
Kevin Williams: and within Cloud. And those projects are equipped with my standards, my documents, et cetera. There's a GitHub linkage and whatever, such that I can build on a consistent basis and do that
Vlad: Mm-hmm.
Kevin Williams: 80% of back and forth there, such I can get a spec. So What is your tooling when you're you're you're, you know, Kevin's the client and we're going to build something new? What does that look like? What do you turn to? What's your app? What's your setup, et cetera, to do that conceptual design?
Vlad: Right. for me, I'm a codex guy, open
Kevin Williams: Mm-hmm.
Vlad: AI. I am working with a few different ones as well just to keep myself informed about how things are developing, because you know, it there's always the best model that's being launched every day with a different provider, and then it gets to be the best right until the next one, which maybe within the same day. And I I actually saw the news that the cycle of the models is actually decreasing, meaning New things are being launched a lot faster than they have been previously. So it's a pretty interesting statistics. and if we look at the the latest launch of Chat GPT six and now Chat GPT GPT actually six point one, I think it has been like within a week or or two. yeah, yeah, yeah. So previously it has been like months for the the first
Kevin Williams: Days. Uh-huh.
Vlad: version of the GPT and then the subsequent ones. Anyways, back to the to the question GitHub all the way, for sure.
Kevin Williams: Mm.
Vlad: I'm using this at both the the repository to keep the code consistent and track the changes as well as the documentation. It is so much easier also to switch the models working with the same repository in case you need to. Like I've mentioned this today to you when you're running out of credits and you know you need to do something urgently and maybe you have a different model, like maybe Claude or maybe other ones that you want to hook up. If you do not have a repository with good docs, you're in for a world of pain because you you will be gambling on the fact whether AI will understand what is it that you were up to or not. Because if you do not have the context set up for the thing, it may just go wild. So
Kevin Williams: So so let me let me interject because you're one step removed from I think the piece that I'm after. That yes, absolutely. Like a friends out there on the internet, like use GitHub, have repositories, they're standardized. You can then you can then if you need to move between codecs or cloud code or grok or whatever you want to do, the repositories there, you can work from it. But go up a level as far as that planning
Vlad: Mm-hmm.
Kevin Williams: piece, like.
Vlad: Right.
Kevin Williams: Are you are you do you have like a planning project or do you do it directly into codecs? Do you reference a repository set?
Vlad: yeah.
Kevin Williams: Like what does that look like?
Vlad: For sure. So I think thank you for clarifying that. definitely like setting things up before you actually get to the building part. I think one of the great improvements overall in how AI works and specifically the latest models is how it works with the examples. It's fascinatingly good these days. For both like the visual aspect and the technical aspect of things. It can really get into whatever you're referring to as an example or even a partial example. And it helps you to do 80% of the legwork before you even get to the real quote unquote planning of the thing. but it also kind of brings up the the the the thing that you mentioned. It's so much easier to get to 80 slash 90% now in both the planning and the building. And sometimes, you know, actually, even even previously, obviously the the path from zero to eighty has been a lot longer, but it has been the easier one too. meaning that it's always the last 10% is where things start to crumble and and actually you discover the complexity when you're into that thing. But yeah, for planning for sure, the examples, the references, the screenshots. I'm also a professional yapper now, so I do not I mean, I do type certain things in, but mostly I just yap with my codecs all day, every day. My wife is like, Okay, what's going on here? Like, who are you talking to? And I'm like being emotional and everything. And I'm like, what is it that you're doing? Are you crazy? Are you how did you even think of that? And it's so much easier to interact with it like I would with a with a person. And I I think that, you know, I I I don't I c we can't say for sure, but I went I'm really mad with it not getting me. I'm start I'm starting to gaslight it. I'm like, dude, you know, my friend told me that Claude did this so much better. Like at
Kevin Williams: Yeah.
Vlad: one tenth of the speed. Could you could can you kind of let's beat that guy? I mean, let's prove you're better. And I can't say for sure because we don't see what's happening under the hood, but in my experience, I have been able to solve quite a few problems, especially in the last days when the mo the new smarter model came out. Much better. faster than I would without the the gaslighting thing. So I think it actually works. So it's like it's called like the
Kevin Williams: Ha ha ha.
Vlad: unofficial turbo mode. It's like when you gaslight codex. But you know, you have to be careful because if you are being too harsh, like my friend tried the same thing and he's like and he probably was using a a different way to formulate the thought. He's like, if you won't do this, I will I will go to Claude or something like that. And it actually said, okay, you know, fair enough. And it gave it the prompt. Like exactly. So you have to
Kevin Williams: Yeah take your bags and go.
Vlad: know when to stop, you have to be balanced. So you have to be a a you know, a professional L L gaslighter to use that feature. So yeah that
Kevin Williams: So Anthropic actually has an internal index for frustration and they track expletives. So
Vlad: Yeah.
Kevin Williams: there is a dashboard in Anthropic and they're probably a GPT as well that somebody looks
Vlad: Probably, yeah.
Kevin Williams: like as far as the sort of the per capita or per prompt incidence of cussing. And the more that people drop like the F bomb,
Vlad: Mm-hmm.
Kevin Williams: the more frustration there is in the system. And apparently that's part of how they figured out that GPT-5.0 was just like horrendous for
Vlad: Yeah.
Kevin Williams: a lot of people, is because the cussing just went off the charts. But I'm gonna push just
Vlad: Yeah, that's a good metric.
Kevin Williams: a little bit harder. Like, like as a dev, when you open this up, are you just opening up the Codex app, like the desktop app? you know, now with cursor is off of Codex land because of Grok and you know, et cetera.
Vlad: Yeah.
Kevin Williams: I imagine that's changed your behavior. But is your blank screen, is that like a project in Codecs that's like this is my project starting point?
Vlad: Exactly. So a new project, every time you actually are building something new, even if it's adjacent to something existing, I still prefer to build it as a new project because of just the context and some things that definitely should be separate. So our our friend actually doesn't do that and he doesn't get like the whole project thing. So it's it's all over the place where in his workspace, and I'm like always due to have to change that because you're confusing the hell out of everybody, yourself included. So definitely start with the project. I usually connect it directly with a GitHub instance. And most of the time, the things that I'm working also require a backend, a database, let's call it that. So I'm using Superbase for the most of my projects. Even before I actually am done with anything, I usually spin up the instance of Superbase. I connect it to GitHub. Both are free. And I also use Vercel for front end, which is also free up until very like late in the game. So it's costs you literally just the subscription to get. thing off the ground and actually publish it in front of people. but yeah the whole planning goes in there. So I always use plan mode when I'm kind of figuring things out. It's also a hack that I don't know it's pretty plain in sight, but I think some people are overlooking it. previously it has been more pronounced in terms of how the Chat GPT interface itself was saying hey you can plan things. whenever you mentioned plan in your prompt It actually highlighted okay, there's a planning feature. Have you heard that? Claude actually is being somewhat better and also somewhat more annoying with this whole thing. It it doesn't ask you whether you want to plan or not. It always kind of is in that mode. And that actually freaks me out sometimes because it's like, okay, it's dude, it's like so obvious. You didn't even need
Kevin Williams: Ha ha ha.
Vlad: to ask me. That's such a weird other option that you're proposing to me that I, you know, do we really need to ask? But I think field codex is a lot more balanced in that regard. So I plan things. Even if my idea is perfectly validated and calculated, I always start with the planning mode. And a lot of times codex does ask me good questions. Okay, how about this one? What about this edge case? Should we do it this or that? And you know, it kind of helps me both to think about this better myself, as well as for the codex to do a better job in the end of the day. So And yeah, so once once that is done, once the planning phase and the whole back and forth figuring out the questions is there, I just shoot it and we're off to the races.
Kevin Williams: So the this audience knows that I've been very Claude Code centric until just the last about six weeks or so. And now I am extremely codecs focused. but I do have both of them up, and part of it is just actually right this second. This for this show I've I stopped everything. I've learned hard lessons about having like agents running on my aging Mac M1 while doing a podcast. Like when you see when you see it freeze on YouTube, that's usually because Kevin's run out of processor. but you know going back and forth between them is interesting, and there are certain things that I'm very comfortable launching Claude Code for, but I'm finding the utility of Codex and its extensibility to just be so much better. And I don't want to go down this rabbit hole quite yet, but we we will in a minute. One is computer use and the ability for Codecs to independently be able to do things like configure Google Cloud. you know, you wanna create an API, it's gonna make it so. Like it's kind of amazing. But but from a more pedestrian perspective, there are a couple of little features that I I'm curious how much you use. So one
Vlad: Mm-hmm.
Kevin Williams: in Codecs is the side chat. So my MO and Claude Code, it's sort of similar to what you were describing. I would do my planning, but I would have this project. in Cloud browser or in the chat function that
Vlad: Yeah.
Kevin Williams: was connected to my GitHub repositories and then I could plan there and I'd kind of bounce between the two. Codex has a side chat functionality and it's not brilliant in that it doesn't know your whole GPT universe, but it can look at what you're doing and it can dispatch agents and it can do things on your behalf. So
Vlad: Mm-hmm.
Kevin Williams: that's one I use very, very frequently. But there are also a couple of other features in there, like there's a browser feature, there's a preview feature, et cetera.
Vlad: Yeah. Mm-hmm.
Kevin Williams: Maybe talk to how you think about those those ancillary new features or newer features for codex.
Vlad: Right. So I'm actually not using this side chat feature all that much, exactly for the reason that you mentioned. I I love to be more in control about the context. And usually when I'm asking something, it's related to something else that I previously asked, or I just want to play around a little bit. So rarely do I have the question that is completely disconnected from like everything. And if I do, I I usually do need to do something with the output of that query. So I just do it directly in codex. And one
Kevin Williams: Okay.
Vlad: of the things that contributed to me kind of not really using these all that much is the about a month ago, maybe a couple months ago, there was like a major update in terms of the codex and the chat GPT app. Previously those were separate applications, and Chat
Kevin Williams: Mm-hmm.
Vlad: GPT as a desktop application was used by a hundred million people or whatever, a lot of folks, and the browser version, while codex was used by just five by five million, I think. And pro mostly there were developers or People who were curious about the building part of things. And I think they kind of tried to flip the switch on that and let more people be familiarized with the power of codecs and building. This is where they did the chat work thing. And I honestly I understand why they did this, but I don't love the change because it kind of
Kevin Williams: Mm.
Vlad: it messes things up, it kind of puts together things that I'm working on and things that I'm just you know want to talk about. together and then I mean I I get the utility, but I just want to keep it separate. So whenever I just want to talk about something or do a research or query or like do a plan, I usually probably go to my browser, same as you do with your claude.
Kevin Williams: Mm.
Vlad: But when I'm actually into the building mode, I I go back to my codex and I'm just I'm I keep on using it just the way it was before. So maybe it was just, you know, me being rigid and non-flexible with the whole change thing. And I maybe I'll I'll it will grow on me, but so far I'm kind of keeping things separate. and yeah but in terms of the The other like features, the browser thing, I'm using it all the time. I'm always
Kevin Williams: Okay.
Vlad: using it to because I do a lot of front end work as well as the back end work.
Kevin Williams: Mm-hmm.
Vlad: Before I publish it actually to the front end, I start a lot of things on my local host and I use that. That is actually very helpful. So whenever I'm in the build mode, there's like the whole context right in front of me. And I said, okay, before you push this any further, open this in the side tab, show it to me how it looks like, validate this and this and that, and I kind of If something is off or there's like a big red error like all over the screen, I can just directly in there looking at it, say, hey, here's what happened. Let's figure it out. This is super convenient and helpful. So you and for this, actually, I I see this as very valuable because it keeps like the context shared, visible, and from both like a user perspective and also the system perspective. So it's definitely a winner. But I think mostly old data coding agents do that. I Cloud Code also does pretty much the same thing.
Kevin Williams: So a c a couple of different observations. and I think we're we're logically going to get into new agent forms and dot in just a second, because
Vlad: Mm-hmm.
Kevin Williams: of the but one is it's very obvious to me that that the the platform universes are drawing together, and pretty soon you will have that shared context, like across different apps and applications and things like that, that I think will make that side chat a lot more useful. the other corollary is in just business productivity. Claude is quite strong in Microsoft. it's actually, I think it's quite a bit better than GPT in the Microsoft apps, but you could have multiple instant multiple instances up and they're all baby clods. They can't actually talk to each other. They don't have well, they can talk to each other from a document perspective, but
Vlad: Mm-hmm.
Kevin Williams: they can't actually reference another chat. So that stuff's all happening and and This is sort of perfect to queue up what happened even just this week or in the last little bit with these new agent forms. And Matt and I talked about this a little bit last week. with Grokbot and then playing with Meta's Muse. personally,
Vlad: Yeah.
Kevin Williams: I I tell you, hey audience, you want an unlock, you want to earn your time with us. I set Muse off. So this is Meta's product. It's very friendly. we put a link in the last show with a with with like a Code for a billion tokens. We'll put it in this one as well, just to experiment with it. But I sent it to my auto insurance and had it log in to GICO, pull down my policy documents, and then had it go out and do comparative quotes. And I felt like a total schmuck because apparently I was leaving thousands of dollars on the table. And it literally saved me $3,000 on my auto insurance. I sound like an ad, but like for for no work. Like I I
Vlad: Yeah.
Kevin Williams: I'm gonna save three thousand dollars this year at a church. Thank you very much. I have never said thank you to Mark Zuckerberg ever in my life, but thank you, Mark Zuckerberg. That's a first, not my favorite person.
Vlad: Yeah, yeah, yeah.
Kevin Williams: so okay. And then just three days ago, OpenAI had its big dev day, whatever the heck they call
Vlad: Mm-hmm.
Kevin Williams: it, and they they launched a couple of things. we're probably not gonna go into spaces right here, but that the maybe that's something for Matt and I to cover in the future. I think that's interesting. But the killer unlocked for a lot of us was GPT's version of I'm gonna call it its version of Grockbot versus its version of Muse because they are a little bit different in their functionality. And you and I were talking about this a little bit pre-show, and
Vlad: Mm-hmm.
Kevin Williams: I am completely freaking obsessed with this thing in
Vlad: Me too.
Kevin Williams: that it is and it's a different form factor. Than I'm accustomed to. And I I'm it's kind of wrecking a lot of my workflows. And now I have so many different things that are going on at once. Like it's absurd. I have like literally 25 different subagents that are working on different things. I'll pause right now. actually
Vlad: Yeah.
Kevin Williams: it's in the cloud, so I they don't actually have to be paused. They could they could be doing their thing. So maybe describe
Vlad: Right. Just being careful.
Kevin Williams: a little bit on on on what what dot is and What the heck is the name dot? But whatever. It's fine. Every time I I
Vlad: Yeah.
Kevin Williams: try a transcript dot, by the way, using WhisperFlow, it puts a period in place, which you say I have to spell it. I'm like,
Vlad: Yeah, not very AI transcript friendly. Yeah.
Kevin Williams: like, come on, guys. Like you you need to be AI friendly with your naming. So there are lot of periods that are around that represent
Vlad: Yeah.
Kevin Williams: dot as I've been talking about it. But anyway, explain to us a little bit what dot is and your initial impressions and how you're using it.
Vlad: Right. In Lame Wayne terms, essentially that's your virtual assistant. So think about the task that you would naturally give to your assistant. Like like you said, check out the insurance, look through the docs, check out my mail and calendar for the the meeting spots or the opportunities or what I said that I do and I didn't do, and and stuff like that. So I mean obviously there's a lot more to it, but that's probably the easiest way to describe the The primary utility of this thing. So for me, the the the the aha moment was that I I didn't like set it up or something. It just came with the new update and I just you know gave it a name and I left it at that. And my Chat GBT had the access to my mail, to my account, or to a couple of other things that I'm using. And you know, out of nowhere it just pops in and says, Hey, you had a call yesterday with this guy and You said that you're gonna invite that person to that meeting today, and I don't see a new person on that invite. Are you sure you you're still on track here? I'm like, dude, I didn't even ask for it. Like,
Kevin Williams: Ha ha ha.
Vlad: thank you. I appreciate it. I mean, that was a small gesture, but it was so thoughtful. I don't know, it probably it it it's one of the things that it kind of came out from experience and it didn't just, you know, happen. And maybe just like somebody else training it to do that kind of thing, but it was like super, super convenient. And the the two hours later. It pops into my inbox and said, Hey, I saw a couple of bills in your inbox with a due date of today. just making sure you took care of this ones. I'm like, Wow, wow, wow. I mean, I appreciate this. you know, whatever. I mean, it's like naturally out of the box without you needing to do anything at all. It acts as your assistant. You can obviously train it and explain it and then set up the routines to do a lot more. but you know, this is essentially in TLDR. This is what the OpenClaw was supposed to be.
Kevin Williams: Mm-hmm.
Vlad: And this is what people expected from OpenClaw to be like a long time ago. And it it's sort of like it's a bit of a different breed. Obviously, people will say OpenClaw is a lot more configurable and flexible. You can set up so much more in OpenClaw, and I agree. I use OpenClaw every day. but this thing is I I heard this description of people saying this about the the grok butt, but I think it's very applicable. Here as well. It's like if Apple would do open claw, this is how it would look and work like. So very hands-off. It kind of knows the best way to do things, needs very little guidance. You can set it up the way you want to, but you don't need to like get into the code or under the hood or like explain every part of the way for it not to crumble down in the first unexpected scenario, stuff like that. So yeah, that's that's pretty cool. I'm I'm very impressed. And basically this is like a AI agent that is with you that works in the same place. You know, speaking about the previous topic of having the shared context between your like working space and your personal like chats and and whatnot. So this thing is everywhere. So it knows exactly what's going on. So it you can talk to it like with a person who would be like literally sitting next to you and that is aware about what's going on with both your like life or whatever else you're exposing to Chat GPT, so which is very convenient.
Kevin Williams: So I've been calling it just that promise of an executive assistant. So it is
Vlad: Yeah, yeah, yeah, exactly.
Kevin Williams: you know, it's the Jarvis idea from Iron Man. And
Vlad: Yeah.
Kevin Williams: just a I'm gonna wrap on a couple of other things that maybe you haven't touched yet. one is it has this very odd interface for an audio function. It's actually a phone call, and I can't really figure out why it's a phone call, but it
Vlad: Yeah. I don't know.
Kevin Williams: will trigger it. And I found it to be too slow, so I had it speed up. But most
Vlad: Mm-hmm.
Kevin Williams: mornings now I'm I'm kicking back in in in our sponsored chair, the the Liber Novo chair, which you know, they they they kindly sent us. And I'm just talking to it and having a conversation as if I were having a conversation with Jarvis of like bouncing ideas off of it and whatnot. And that's that's an interesting experience. I'm not sure that it's fast enough for me in terms of
Vlad: Mm-hmm.
Kevin Williams: like it has to it the the latency is isn't isn't long, that's not the issue, but I want to be doing 10 things at once. And
Vlad: Right. Yeah.
Kevin Williams: really that that ends you end up on a single track, but it's it's a pretty cool experience. The
Vlad: For sure.
Kevin Williams: parts that I stumbled on that I found really novel and I can't figure out quite yet as far as my workflow is it's actually directing codecs for me. And it started doing that sort of unexpectedly. where I was I basically was recognized that one of my apps it was a like a marketing dashboard and it had a YouTube connection and the MCP wasn't working the way it should be. And I basically made myself a to do. And as opposed to just making itself a to do, it made it a to done as in it just like it it just like launched a codec session
Vlad: Yeah. Just did it.
Kevin Williams: and it went off and it fixed it. And then
Vlad: Mm-hmm.
Kevin Williams: it took over my browser. And did the confusion the the not nothing the the connections that it needed to make from a browser perspective as I was doing something else. And I went whole hog into it yesterday and like the amount of stuff that I did, I I I I mean, okay, we're like all productivity n ninjas and we're like super involved in this. I did a month's worth of work yesterday. Like I
Vlad: That's pretty cool.
Kevin Williams: absolutely did a month's worth of work. And I like I it it's just chewing on it just happily. I have no idea how expensive this is going to be. Like at the moment, they're clearly giving you some room to play with it because my
Vlad: Yeah. For sure. Yeah.
Kevin Williams: my my I think I'm down to like 17% on my week or something like that. But given the amount of power use I've tossed at it, I think I think they're subsidizing it just so they can give people like experiences around it.
Vlad: For sure. The whole thing. Yeah.
Kevin Williams: but it's a different experience. And and you know, previously GPT had had agents, which it's funny, I'm looking at my browser GPT, and agents are still sitting there in the left bar, but the agent function in the desktop app has actually gone away. And I think that they're mol they're like melding in agents into this new dot type functionality. And what I I more than think, I know that's what's going to happen next is going to be an open clause style orchestration layer where your main dot. So right now everybody only has one dot, and that main
Vlad: Mm-hmm.
Kevin Williams: dot does everything from managing, in my case, like my kids' social calendar to using codecs to connecting to email and parsing and whatever. But it's pretty obvious that you're soon going to have multiple dots. You're going to have like you're going to assign a dot that is like your dev dot. And your main
Vlad: Mm-hmm.
Kevin Williams: dot will talk to the dev dot and be like checking in on it through the day through heartbeats. And it will start managing that on your own and on its own. And I just I honestly I'm not quite sure what to make of this because the The productivity jump is just it's so big. Like it's the biggest
Vlad: Mm-hmm.
Kevin Williams: productivity jump that I've seen yet in all of this. and as more and more people start using it, I I I I I just don't even know what what work it's gonna start to look like, to be honest.
Vlad: Yes. this is a question that we all ask every day. I wonder if they actually reuse some of the technology and and the knowledge and the the thinking of the person that did create the open claw in the first place because they hired the guy like six months ago. And and everybody
Kevin Williams: yeah, of course they did. Yeah. Yeah.
Vlad: was like all excited about this, but then it kind of went silent. He went on a few podcasts and whatnot and six months later or whatever that is. Yeah, yeah, exactly. And
Kevin Williams: Took his billion dollars and
Vlad: And but looks like he was busy or or people that, you know, took over whatever he provided were busy. So this is looking pretty good.
Kevin Williams: It it is a I have to say it is a little bit frustrating as this marches around. Not necessarily frustrating, but the evolution is so fast that you
Vlad: Mm-hmm.
Kevin Williams: establish like say primitives around what I was talking about with, you know, my approach. I'm I'm still, even after our conversation, I'm still pretty comfortable with my approach to to dev design. I don't think it's very different from yours. So that was very validating for me. but it is like if I were to go into a fresh person. If I was to go to
Vlad: Mm-hmm.
Kevin Williams: my 17 year old and be like, all right, this is the way that you should build. I'm not sure that I would necessarily build all of those building blocks today. It might be a little bit more just conversational and like there
Vlad: Yeah.
Kevin Williams: may be a there may be like a a skipping here that a lot of people are going to get the benefit of.
Vlad: Mm-hmm.
Kevin Williams: And yes, you and I and people listening to this show have a lot of hard fought and sometimes expensive lessons that I think are going to serve us in good stead going forward. But all of this is clearly just going to get much more accessible as time goes on.
Vlad: Definitely. I mean for me the the change with the the whole like call your your dot thing wasn't that radical because I I'm telling you, I've been yapping with my codecs and all my models for the last two, three months at least because it I I feel like I convey so much more and so much faster than I would do in typing. I mean I I've been speaking with my developer friends and they're like, No, no, no, we actually stick to typing. We have like the feeling of control or whatever. I'm like, dude, let it go. I mean, this is I mean we can so much better like articulate what we actually have in mind. We can I mean for for conciseness and for actually texting a person, I'm all for the text messages. I I'm not really a fan of voice messages and all that. I'm that guy who ignores your voice messages or transcribes them in instead of listening. Yeah, yeah, definitely. That is the voice message guy and but luckily
Kevin Williams: Yeah, yeah, tell Matt that. Matt is a bo Matt is a voice message guy.
Vlad: for for me and for the rest of the folks like me. The transcribes works pretty good. I mean, these days, so you don't have to worry too much about that. But
Kevin Williams: I didn't want to I don't want to say that I I am 100% dictation. I'm over two million words on
Vlad: Yeah. Yeah, for sure.
Kevin Williams: on WhisperFlow. And and I joke that like when somebody makes me type, I type like a cat with peanut butter on its paws. I'm like, I don't want to do
Vlad: Yeah.
Kevin Williams: this. So the conversational aspect I think is well established. that I'm fine with. So I've always been been managing that. I think it's the it's this this orchestration. the the lower friction in dot
Vlad: Yeah. That's new.
Kevin Williams: style or grokbot style orchestration than we had with Hermes or or or open claw and whatever. And again, you know, hard fought systems, they do what they're supposed to do, like they're great, they're highly configurable, whatever. Kinda so I actually haven't used Hermes and a and a buddy of mine was was like, when are you gonna give Hermes a try? And I literally responded with never. Not going to because
Vlad: Yeah.
Kevin Williams: There's there's really like, okay, I get it, and there there are there's a limit to the stuff that I can experiment with. but sort of waiting for the easiest technology, the most accessible technology, I think is gonna be the right bet for a lot of people. And we're right
Vlad: For sure.
Kevin Williams: at the cusp of it.
Vlad: Yeah, I mean, unless you have like an urge and or an urgent requirement to do things right now. So we just took the open claw and built one of the features that we had to have to facilitate certain like interactivity with our system. We just took what we we knew was available. And probably there were a couple of alternatives, including Hermes, that were already available at that point. But it did the job, it works fine, so no complaints. But in terms of which technology to invest in, I mean it's getting scary. As I said, the the the the development cycles between the next best thing and and the other next best thing are getting super short. I mean, like literally within a week. So you don't even know. So it's like, whoa, what's going on here? So should I even get started?
Kevin Williams: Well that was a comment from OpenAI Dev Day, which was they're six months ahead on their own feature production because they're using their own tools. And they themselves are amazed. They're like, Wow, that turned out to be really easy to do because we were just using Codecs with a billion tokens a minute. you know, using Codecs eight point or whatever the heck they're using internally to to to
Vlad: Whatever they're using, yeah, for sure.
Kevin Williams: to do the thing and just, you know, rampaging ahead. Yeah. Okay. So the
Vlad: You never know with these guys.
Kevin Williams: so the the the piece we we we alluded to a little while ago, but I d I bet that most of the audience has not really leaned into computer use. So computer use has been around for a while as as far as the ability to move files around, access your downloads file, access your desk desktop, et cetera. Totally fine. one of the big unlocks for for most of us with Astra in particular. And now Sol 6.1, et cetera,
Vlad: Mm-hmm.
Kevin Williams: is computer use as in the ability for the LLM to use your browser, to interact with your computer, et cetera. And the example I gave earlier was like interacting with Google Cloud, which anybody who's had to set up like APIs in Google Cloud knows that it is the most Google of Google environments, which means it's really finicky and it has a lot of like
Vlad: Yeah, for sure.
Kevin Williams: levers and dials and things that you have to deal with. And now at this point, I'm just like make it so, like straight out of Star
Vlad: Mm-hmm.
Kevin Williams: Trek. That's really my thing. I literally say make it so because I'm a dork. but it's like,
Vlad: Yeah.
Kevin Williams: you know, like talk to that a little bit because it is a little terrifying to like let it loose that way. But I've had a couple of weird experiences, but by and large, it just freaking works. Like I launched
Vlad: Yeah.
Kevin Williams: a A hundred ad meta campaign yesterday, like h completely hands off with a mixture between the MCP and you know, using computer control.
Vlad: Yeah, for sure. I mean, definitely I I think the most staggering thing about this whole development is like how far and how how much this has developed from from day one. Because as as you said, the computer use is one of the actually the the early features of AI. And I think this is the one feature that most people who are not necessarily doing this every day were expecting. They were thinking that AI is gonna come like literally at their computer, whatever they're using. take their software and just do this thing better, faster, no mm, making no mistakes and kind of like 10x the outcome without changing any of the foundational work. We have been into quite a few conversations with customers, and this literally what they wanted. So when we said AI consulting, they began by showing us their random software from 2004 that barely runs and crashes, that has its own local database that doesn't know what cloud is, and you know. And it provides some sort of like a weird CSV that is the basis of their business. and I'm not trying to kind of poo-poo on them. Those are people from the industries which are not like very tech enabled from the construction and whatnot. And they said, Okay, guys, you probably need a new app because what you need, what you're describing as a problem, is not necessarily an AI problem. That's just a software problem. So your software is
Kevin Williams: Mm-hmm.
Vlad: outdated and your your needs are completely different from what it does. And they're like, no, no, no. We actually love our software. We just want the AI to do this for us in that program from 2004. And like, okay, we're gonna have a problem here. So let's build a harness around it and whatever. So we kind of like were I was we were sidetracking that conversation to go to a different place. Let's start somewhere, and then slowly the plan was to convince them to actually redo the whole thing because did they did not it do AI. But if we would have that conversation today, I said. I would probably say sure, we can probably give it a shot because as you said, Astra and and the the new generation of the LMs is pretty damn good at this and they're not spending all that much tokens on the whole thing. So in in a lot of cases it's actually a better strategy to try and just, you know, give it your computer for a ride and w rather than building the whole like piece of software. Looking for MCPs, configuring CLIs and you know, dealing with expired keys and whatever. I mean, just let it let it take the wheel. And the last bit of it is that it's just I think maybe this is kind of the the first thing also about it, is because it's getting so much better, it's actually much more efficient. I mean previously I think Claude was a lot better and a lot more advanced in in the browser use and the computer use. But one problem that I continuously ran into is that it spent my whole credit limit like on a super simple operation, like literally within tens of minutes. Like I worked for 15 minutes, claw does the computer use for me, and then it expired. I'm like, dude, I'm I'm
Kevin Williams: Yep.
Vlad: not I'm jump what's going on here. So and now it actually can work for hours and apparently it stays still under limit, which tells me that it's probably just a lot more efficient. So faster, cheaper, just Gets the stuff done.
Kevin Williams: So I think there's a point here that's a little bit under the hood, which is under the faster and cheaper bid and the exigentic functions. so Muse. If you want to strip back what Muse is, you know, we're not supposed to talk about models on this show, but Muse is built on Meta's Spark 1.3 model, which is a thoroughly okay model. It is not a dumb model, but it's not a frontier model. Muse
Vlad: Mm-hmm.
Kevin Williams: is freaking brilliant. And when you interact with it, it's basically batting outside of its weight class. for its model capability, because the model has been tuned to this sort of activity in a way that it doesn't need the frontier intelligence. I think
Vlad: Mm-hmm.
Kevin Williams: Opus 5.5 or even Luna 5.5, not Luna Luna's on the I'm gonna confuse myself, Sonnet 5.5 was was yesterday. Like these are being tuned for efficiency and specific use. And you're going to start seeing tuning around computer use as well that's going to
Vlad: Yeah, definitely.
Kevin Williams: make them much more efficient. And I I'm fascinated by that because, particularly when you add like decision models, like a JEV in, and there's actually a new decision model that's clearly a JEV ripoff for for open AI, etc. We're in a really, really good place. And I I have been an optimist on pricing for a long time, sort of a dubious optimist. And because everybody's like, the subsidy era is gonna end and all of this stuff's gonna get heinously expensive in a hurry. And you know, we were talking before this about what happens when you go off of your plan and you use tokens and it's like crazy expensive. But
Vlad: Yeah. Not a good place. Yes.
Kevin Williams: There's a there's a balance for the model companies, for the frontier companies of adoption and practical use versus the ROI. And they
Vlad: Mm-hmm.
Kevin Williams: know they're not dumb. They know that if they increase these prices too fast, they're gonna chase people away from things like being able to use your dot for you know hours and hours and hours and enculturating
Vlad: Mm-hmm.
Kevin Williams: that into your business. And I I think that there's an argument to be made that. People are once it gets a little bit more advanced, people are going to be reasonably willing to pay rates of like $500 a month for a subscription. And there is a new $500 a month GPT subscription out there. And I could probably justify that, but if it was $5,000 a month, it's going to be very
Vlad: Mm-hmm. Yeah.
Kevin Williams: hard for small businesses or individuals to justify it. And I it's going to r sort of stunt the market. So these
Vlad: Mm-hmm.
Kevin Williams: efficiency efforts that are happening under the hood. That are orienting to like least common denominator models, I think is what's going to give us the staying power. Like it may not be quite as fast, it may not be quite as brilliant, but if you're not doing protein folding, like you don't need that brilliance so much. And that that's going to keep costs under reasonable control, particularly when combined with market pressures and you know their need to keep adoption high.
Vlad: Definitely. There's also this whole big thing going on with the the open source models. So there's a whole like war in that specific sector that it's apparently building up for that the frontier companies such as OpenAI Anthropic, obviously, Groc maybe to a certain extent, they d do not openly voice their concern about the specifically the open source models, but it's very obvious that this is what they're Trying to kind of throttle or at least diminish or like hold back a little.
Kevin Williams: Mm-hmm.
Vlad: Because one of the options of how this can
Kevin Williams: Yeah.
Vlad: unfold is that we all know, and this is not a secret, that it costs a lot more for OpenAI or Anthropic to run their models than what you're reasonably able to pay for it. So I I somewhere I heard that you can Theoretically, I mean not everybody does that obviously, but you can theoretically use up to twenty thousand dollars of company costs for the the cost of the two hundred dollar subscription. like I don't know to which extent that is true. That's probably hypothetical, but I'm I'm guessing it's like somewhere out there. and th that's probably one of the reasons why are they also like upping their amount. And I think Anthropic has also been playing around with the amount of credits and subscription costs a little also for that purpose. Obviously the They wanna grow, but you know, it's it it it stinks, so they burn a lot of money. So
Kevin Williams: So just super quick, because we're we're gonna run out of time here and I wanna I wanna get into MCPs for just a second, but what is your go-to open source model?
Vlad: I'm yet to pick one, honestly.
Kevin Williams: Okay.
Vlad: yeah. Because I think I'm kind of I'm too busy playing with the new things that are being launched by OpenAI and Anthropic every day. But to one of your points made previously, I also had a M1 up until last week and it was it was getting to me because it was getting a little slow. So I actually got myself a new Mac that is very much local LM ready. It has forty eight. gigs of RAM, the M4 Pro. So I'm bringing on. I can probably launch OpenAI here locally, if it would be available. But you know, whatever is the next thing, I'm ready for it and I'm definitely gonna take it for
Kevin Williams: So there's a there's a whole rabbit hole there with outsource processors as well
Vlad: For sure.
Kevin Williams: of that that people are doing. But I I think in sort of the last bit of technical bits I'd love to leave people with are the the increasing flexibility of MCPs. And MCPs are model context protocols. And these are basically flexible gateways to sources of data. And they're so easy to spin up now. And I don't think that people really realize that. That if you have a data store somewhere that's accessible, slapping an MCP on the front of it, it you can do this with greater or lesser efficiency. But you know, people need to know that they can do that. Like I think one that that is just anybody listening should do is anybody listening to this show is probably doing a lot of transcripts. And your transcript provider probably offers an MCP that you can connect to and you can connect it to Claude and then be like, hey, remind me of that last conversation I had with Vlad and It it will pop up. But I encourage people to retain their own transcripts because it's one of the pieces of a business that is just so valuable. And
Vlad: Mm-hmm.
Kevin Williams: being able to put an MCP on top of it and basically using MCPs to collect information. So, like something I was working on yesterday with Dot. was again the marketing dashboard and that my my internal marketing dashboard has probably eight or nine different APIs or MCPs that are connecting to it and it's pulling in all of the data from all of the sources. I can put an MCP on top of that entire app that has the collected data and the collected intelligence of what's going on with my ads and what's going on with SEO and AEO and whatever. And not have to rebuild all of those pieces individually on the like the chat bot side. So I guess sort of speak a little bit to how you are you are deploying MCPs. And there's this this idea of interrelation between MCPs that I don't think a lot of people are thinking about yet, that they think of them as a single point. I'm gonna go get my, I'm gonna use a connection to Google, it's probably an API connection to get SEO data from Search Console. I'm gonna pull that one piece in. But the inner relation between all of those pieces of data is like the next step of complexity.
Vlad: For sure. And I think that honestly, my take on this whole MCP thing is that we're getting on a track where most people will not even realize that they're using an MCP. And they probably shouldn't care at all about that. But you know, all these fancy terms like okay, spin up your open AI, the droplet, the whatever. That's
Kevin Williams: Yeah.
Vlad: most people are just gonna be saying, I I just say, you know. I want it so, and then it it it's off to the races and it's just getting things done. So I this is sort of how I work with the MCPs even now. So I obviously connected to whenever and whatever I'm using and interacting with the most, like super base for my transcript. I use Fireflies. I I'm you know, all day, every day connected to my MCP. This is how the dot reminded me about that guy not being invited to the meeting from today because it pulled my transcript from Fireflies and kind of connected the dots. No pun intended. So and but it's a a lot of times I'm actually not using MCP because of for example I'm you I'm working with multiple instances of the one thing like for example multiple instances of super base and the MCP like out of the box usually only lets you connect to one. So again my specific use case probably for most people it's not even going to be relevant but there are so much more ways to access certain data. And sometimes AI is just able to figure this out on your own. So you and I'm I'm sure that a lot of people may be using MCPs or CLIs or APIs without even realizing at what exactly did they use at this particular moment. As long as it gets things done, it probably doesn't even matter. So I guess that, you know, definitely it's a powerful tool. Definitely should be out there and you should know that this is a thing. And if not, you know, ask your agent, it will describe it to you. But I think very soon, may I dunno, very soon nowadays is maybe two weeks from now or a month from now, this
Kevin Williams: Yeah.
Vlad: whole thing may not be even be relevant. It will just, you know, work. So I don't know if it's a good thing or a bad thing, but we're definitely moving in that direction.
Kevin Williams: Well, Vlad, this has been great. I seriously being able to geek out a little bit more on this. I really enjoyed it. we should we should kick Matt off occasionally so you can you you and I can can can do this. let him roam. Let him roam. It's all right.
Vlad: Yeah. Let him roam. So yeah, the man likes to roam. Yeah.
Kevin Williams: but yeah, sir, if people wanted to I don't know, do you have socials or anywhere where people might be able to follow you? If you're not into that thing, don't worry about it.
Vlad: Yeah, I'm on Twitter. I'm on LinkedIn. Yeah, yeah. So yeah, hit me up. I'm I'm happy to geek out more about all this whole thing.
Kevin Williams: Awesome, man. All right. We'll talk soon.
Vlad: Thanks, Kevin. Talk to you later.
