The AI Daily Brief: Artificial Intelligence News and Analysis - RIP Vibe Coding. Feb 2025-Oct 2025.

Episode Date: November 3, 2025

With NLW currently on the road, he's joined in this conversation by Sean “Swyx” Wang — developer, writer, Latent Space host and newly joined member of Cognition. They explore how AI coding became 2025’s defining story, why “vibe coding” is ending (sort of), what comes next for developers, and how “Agent Labs” are reshaping the balance between model makers and product builders. Swyx also previews the upcoming AI Engineer Code Summit in New York and shares why “code AGI” could deliver 80% of AGI’s value long before full AGI arrives.Brought to you by:KPMG – Discover how AI is transforming possibility into reality. Tune into the new KPMG 'You Can with AI' podcast and unlock insights that will inform smarter decisions inside your enterprise. Listen now and start shaping your future with every episode. ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://www.kpmg.us/AIpodcasts⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠AssemblyAI - The best way to build Voice AI apps - ⁠⁠⁠⁠⁠⁠⁠⁠⁠https://www.assemblyai.com/brief⁠⁠⁠⁠⁠⁠⁠⁠⁠Blitzy.com - Go to ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://blitzy.com/⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ to build enterprise software in days, not months Robots & Pencils - Cloud-native AI solutions that power results ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://robotsandpencils.com/⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠The Agent Readiness Audit from Superintelligent - Go to ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://besuper.ai/ ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠to request your company's agent readiness score.The AI Daily Brief helps you understand the most important news and discussions in AI. Subscribe to the podcast version of The AI Daily Brief wherever you listen: https://pod.link/1680633614Interested in sponsoring the show? [e-mail]

This podcast is sponsored by Google. Hey folks, I'm Amar, product and design lead at Google DeepMind. We just launched a revamped vibe coding experience in AI Studio that lets you mix and match AI capabilities to turn your ideas into reality faster than ever. Just describe your app and Gemini will automatically wire up the right models and APIs for you. And if you need a spark, hit I'm Feeling Lucky and we'll help you get started. Head to AI.studio slash build to create your first app. Welcome back to the AI Daily Brief.

This week, as I am out traveling for my anniversary, we are going to have a combination of regular shows as well as some different formats that we don't normally get to do. And one of those is an interview with the man the myth, the legend, Sean Wang, better known as Swix. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. All right, friends, quick announcements before we dive in.

First of all, thank you to today's sponsors, Google Gemini, Blitzie, robots and pencils, and super-intelligent. To get an ad-free version of the show, go to patreon.com.com slash AI Daily Brief. And if you are interested in sponsoring the show, shoot us a note at sponsors at AIdailybrief.aI. Also, another quick reminder of our ROI benchmarking study. It's live at ROIurvey.com. Please take a minute to add a couple of use cases. I'd so appreciate it. And there's a bunch of goodies in there for people who help. Now, you might have heard me talk about SWIX on here, or maybe you've heard his podcast

late in space or his events, the AI Engineer Summit and AI Engineer World's Fair. and even though many of us who are creators or listeners of this show aren't technical or aren't developers ourselves outside of vibe coding, I think it's a really valuable thing to spend our time understanding what developers are talking about. As I discuss with Sean in this show, it's a little bit like previewing the future. And so what we do in this conversation is look at the big themes that he is thinking about and the big conversations shaping that sector of the industry and also how he's turning those into key themes for the AI Engineer Code Summit, which is coming up in New York. Now, for those of you who will be at the AI Engineer Code Summit, I will be speaking there and

I'm very excited. But without any further ado, let's get into this conversation and bring Swix once again back to the AI Daily Brief. All right, Sean slash Swix, better known as Swix, how are you doing, man? Welcome back to the show. I'm doing great. Thank you for having me again. Yeah, it's always great to check in with you. As I was just saying, I think the reason that I'm always pointing people to you and the set of content that you're around is, I think, especially for folks who are outside of the kind of AI engineering conversation, understanding what the builders are talking about is kind of like living in the future a little bit.

So what I wanted to do today is dig into maybe some of those conversations that are driving the AI engineering community. And the specific context that I think is interesting is you have a big event coming up in just about a little less than a month now, where obviously you have to think about and crystallize those things into content. So maybe let's kick off by just if you want to tell us a little bit about the Code Summit and how you think about this event relative to the others that you do. Yeah, and I should also flag that you're speaking, which I'm very excited about.

Yes, and I can't wait to be back with you. So I've been organizing AI Engineer Summits for three years, and usually they are kind of generalist. They focus on just whoever are the best speakers I can get and the general state of AI. And I think that now the meta is kind of shifting towards focus and concentration on a certain topic because when we have as many sort of applicants as we have, because it's like a, you know, you have to apply to get into this conference, we get to pick.

And like the best vibes are with everyone you run into is all concentrated on the same theme. Like you gather for a certain topic. And even like changing the name and like focusing on a certain theme, changes the entire vibe of the whole thing, which is very cool, very, very, very fun. It is something I realize as a meetup organization. So we're doing, this is our first ever summit entirely focused on AI coding. And we're doing enterprise and individual contributor days as well. But I think like the focus is on like why coding has emerged as something that has particular product market fit and especially emerged this year.

And it seems weird for me to say this as someone who's had a whole career in developer tools and kind of always focused on AI coding. We've never done this before. But I think this is the year, like, you know, most people don't even remember that Claude Code only emerged in March this year and is now larger than, you know, $600 million business. And it was like after our last summit in New York when you were emceeing. So like a lot has changed. Cognition and Cursor have emerged as like very large startups. So I can't even call them startups anymore. It's like what we've been calling is agent labs that are starting to rival the model labs in terms of,

of market pool, valuation, employees, what have you. And I think it's one of the most interesting stories of the year. Yeah, I mean, what's fascinating about this is it is, I don't think anyone would disagree that this is, if not the dominant or most important AI theme of the year. It's certainly got to be among the top two, you know? And it was not on the radar as the thing that was going to drive all conversations, you know, when everyone was doing their end of year content, you know, their end of 2024 into 2025 content predictions, no one, at least anyone that I saw was like, this is the year of coding, this is the year of AI coding, AI coding, AI coding agents. It was the year of AI agents broadly, right? That was sort of like the money's on bet for what

happened. I mean, vibe coding, only, Carpathy said that tweet in February, right? It's, it feels like a million years ago because of the inevitability, but it really, you know, we are kind of just catching up with ourselves in some ways. A little bit. And I actually also, you know, have a spicy thing because generally I agree of Andre and everything, and most people do. But I think the one thing that is happening right now is that the software engineers are feeling very uncomfortable with vibe coding. And I think, you know, you talk about how we're all six months ahead of the main street. Vibe coding, you know, I declare the end of vibe coding being cool this month. And I think a lot of what we're

we're meeting to discuss in, at AIE Code Summit, is like, what's after vibe coding? Like, how can we avoid the slop and, like, build software that we don't hate? Don't get, don't get stuck in rabbit holes that the agents might go down sometimes. And so it's going to take work from the model labs, which we, which we have represented. It's going to take work from the agents. And it's going to take work from the customers, which we also, you know, want to hear from. So I think it's interesting because, like, there's new terms. people are vibe coding super popular but I think it also might need to evolve in some way.

Yeah. Well, so so let's actually try to unpack this a little bit because this is this is sort of to me this was like okay declaration like Sean's now in in spicy mode for what's coming with this event right? I think the tweet was RIP vibe coding 2025 to 2025 or something with that like perfectly constructed tweet. But so let's talk about what where the where the discomfort is coming from and maybe sort of like what the difference between what someone who's sort of excited about this term still is thinking about when they see it versus what this group of engineers who are getting more uncomfortable with, when they see that term what they're kind of perceiving. Yeah. I think the issue comes with like every one of us, every software engineer is very happy

that people who are non-technical can get to somewhere productive without engineers. Engineers are expensive. They're hard to work with. They're divas. You know, whatever. Like just, you know, they don't need to help make your website, your personal website, when lovable and bold exist. And I think that nobody's, nobody's, that has any issue with that. I think it comes to a head when you start to say like, oh, I vibe coded this. Like, come on, it only took me like an hour. Now here, here take it.

And I expect the full thing by Friday. And like, well, you know, you haven't dealt with any of the hard stuff. you've only painted the sort of superficial picture and you confuse that for the full working app. That's one issue. That is the sort of non-technical to technical handoff that is not being discussed, negotiated. In fact, what is happening is the infra layers specializing for the non-technical people so that the sort of vibe coders, the non-technical people are basically building of a completely different stack than the technical ones.

And so when you hand it off, you have to completely rebuild. because it doesn't use any of the same tech. I mean, somewhat exaggerating. I think the best crossover tech right now is SuperBase, which is why SuperBase is doing so well. They've basically quadrupled valuation this year. But there's a lot of experimentation in just that front. Then there's also the inter-software engineer fights,

where software engineers are also vibrating, of course, but some of them are being a lot more sloppy than others. and the people who care about software, care about security, care about maintenance, care about honestly, just getting things right or understanding your code

so that you don't get into trouble. Because LMs just do run into rabbit holes and sometimes to really get them out, you have to understand the code. You can't just wash your hands off it or just flow based on vibes. So when that stuff happens and people are irresponsible,

then they also tend to leave PRs to other people have to clean up. So, you know, I think, like, people just want something better. A lot of people are talking about spectraven development as a way forward, which is something that Amazon is pushing a lot, as well as a number of other people. Like, my top speaker from World's Fair was Sean Grove from OpenEI, who was basically pitching spectrum and development and model alignment specs. So, like, I think there's a lot of action around this, the term that has to sort of replace a compliment vibe coding hasn't emerged yet, but I can definitely feel it in the air. It's literally present in every conversation I have. Everyone's sick and tired of vibe coding.

Yeah. So it's super interesting. So a couple things. One, there's this classic pattern with change, technology change, where we forget temporarily that the paradigm shift isn't going to be from a set of problems to an era of no problems. It's trading one set of problems for another, which hopefully are a, it's a good tradeoff. It's a sufficiently good tradeoff that that new set of problems we'd rather deal with because of the gains that come from the switch, right? And I think that that second part of the conversation that you were just mentioning, sort of the intra-engineer conversation, is a lot about that. It's like, okay, well, now we have to reconcile with, you know, all of the stuff that comes along with if we can do XYZ much faster or

automated or with background agents. It creates this new set of problems. And we are still going to have to deal with those. We're going to have to re-architect our systems and sort of, you know, the way that we work to accommodate that. And I think that that's a very natural process of like figuring that out and actually sort of rationalizing what it looks like to use these systems well, even as the technology is changing. And I want to come back and kind of talk about maybe the sink async async ascasing spectrums and a couple other things that you've talked about as it relates to kind of where these things are. The first one, you know, I was thinking about this. we really don't, we don't have a word for the difference between sort of professional and

amateur in the context of a democratizing technology, right? Like, you know, if you think about, like, I was trying to, trying to make the proxy of like content creation with social media, right? TikTok and Caput come along and everyone can make videos. There's clearly a difference between amateur videos and Christopher Nolan and no one would not acknowledge that. And in the middle, it gets blurry, of course. And there's some people who may not be as technically good. but the things that they produce, people like more. And, you know, but there's still like, you know, the terms that we have are all are all dumb, right? Creator, influencer. Like, they just kind of, they don't actually convey this gap. And I think it's actually, one, I think it's completely

unsurprising to me that coding is sort of figuring this out first in the context of AI, you know, AI becomes this mass democratization technology, but there is still a difference between, to your point, like my sick terminal-based, you know, AI Daily Brief website that I use Loveable to maintain and like an actual product that goes out and, you know, an enterprise is not going to freak out on because it's got, you know, kind of its security setup. You know, we just don't, we just don't have good terminology for that, which I think is a challenge. Because to your point, I don't think anyone is actually in disagreement that these things are different things. Yeah. I think to some extent, it is our job to figure it out. Like, this is not an unsolvable problem. And so I don't,

I want to put people at ease here in terms of like, you know, keep, you know, keep doing. what you're doing, keep up with the bolts and lovables and bipodding in general, I think it is the job of the engineers to try to figure out that transition path because we haven't worked it out yet. I'm gathering people and trying to focus people's energies on this because clearly, like, when a new technology emerges and it is somewhat disruptive to the old technology, people who are tied to the old technology complain, which is exactly what they're doing here, by the way. But also, the goal is not to reject the new technology, is to embrace it and figure out how to reshape everything else in order to accommodate it.

So I think like there's there's more synergy here than like people fear when they, when they first hear about this stuff. Yeah, I wonder if there's, I mean, you know, I don't know if it's an interim solution or not, but it feels like there is there's a role or at least a function around sort of translating. You know, if you've got all of, especially if you think inside an organization or a startup, you've got all these folks who are now able to speak with code, right? Instead of talking about features they want, they can just, you know, mock them up, which is, you know, what we do, whatever.

every company I know what this does, you're talking about sort of the challenge of translation. It feels like that's that's a thing that someone could get really good at, you know, both helping people sort of, you know, build things in the right way in the beginning. But anyways, there's lots of developments that I think are going to come on that front. Yeah. Okay. So the next thing I wanted to talk about, which is sort of, you know, builds off of this a little bit is what this landscape of AI and agentic coding platforms, the full breadth of it now, because part of the challenge and why sort of vibe coding are

I think is that like, if you go back six months ago, it's like, who's going to win, Bolt or Lovable? It's literally that. And then Claude Code comes and it's like, okay, now, you know, as opposed to now people, people with a passing glance see Lovable, Bolt, ClaudeCode, CLA, CLA, CLA, CLE, C. Cognition, Factory. And it is sort of this, you know, this broad spectrum. And you actually wrote about this a little bit when you sort of shared that you were joining Cognition. Huge congrats, by the way. I think that's, by the way, for my money, maybe the most, useful. I'm making a career switch blog post that I've ever seen. Usually, that's a very, very sort of self-indulgent thing. It's just like, here's my trajectory. That was like kind of

packed with interesting observations. And one of them that you talked about is the sync async spectrum. I would love, you know, without asking you to kind of boil the ocean, share kind of roughly how you see the topography of these, you know, of these categories of coding tools emerging right now. Yeah. You're making me think about. other conversations I've had since that publication. But yeah, so totally. I think there are a number of charts that people have made. Basically, coding agents are enormously popular.

Now we're just figuring out what the ideal interfaces for them are. So probably initially started with GitHub co-pilot, which is just spicy autocomplete, as they say. Devin launched like two years ago with sort of the web app, sort of interface code interpreter is also in the mix somewhere in there where you can chat and it starts to generate code and run and execute that code. I would say then cursor obviously with composer and all the other cursor agent stuff that they're launching. So I think like now the form factors are you have the IDE or VS code extension, you have the web app, you have Slack or

whatever your sort of team collaboration thing is. You might also want to put linear in there. And then finally, you have the terminal, which is obviously the newest that emerged on the scene this year with cloud code. Basically, you just need universal handoff among everything. And I think, like, that's the goal.

You know, everything I described is all the surface areas. All the companies pretty much have all of them now, I think, with, like, cloud code going to the web and code coming to the BS code extension. Everyone's got everything. And I think that the handoff is not worked out yet. So cloud code is the first one to work out the hackiest possible version, which is Cloud Code Teleport, where you can just sort of dump the JSON of the chat and continue

with it locally, because they're the same instance, same cloud code on both sides. But I think like there may be some more evolutions from there because that's not naturally how we transfer contexts between engineers working differently. And so, like, in my post, I started talking about the sick gazing spectrum and you kind of need to own that, which is why I was very impressed with cognition buying WinServe when WinServe was off for grabs. Because, well, you know, here's like the number two IDE. It's for cheap because, you know, like a month ago is worth $3 billion. Now it's worth like less. The rumor is the $300. I actually haven't even confirmed that number. But yeah, I mean, like, you know, at some point it's worth buying. And actually, you know, you start to have a leg up in that sort of. sync side of the spectrum while async is having extreme product market fit, right? Like I talked a little bit about the numbers in the cognition blog post as well. So, like, I think that's good. I think actually sync async might be a bad framing, which is really weird for me.

Because one thing that's happening now that you're going to see with cursor 2.0 today and also what cognition is launching is that the async side is moving faster rather than slower. Because I think there's been a perverse incentive to measure all the. these coding agents based on the number of hours worked. And like, you know, where else do we do that? Well, lawyers and like everyone, every and like, you know, everything that we hate because you're just incentivizing them to take more time, which is horrible. Like, no one actually wants that. We're just using that as a poor proxy for like how, you know, what difficulty of work and you're actually doing. So everyone's working on faster agents, I think which is good for users, ultimately,

because that's what we want in practice. The async side is becoming more sync. And then the sync side is changing in terms of like the mindset, right? Like, why do you want synchronous code? Well, the, the actual answer is because not everything can be vibe coded. Like, the anti-vib code is to turn your brain on instead of off and use AI to augment your skills and thinking rather than to replace it and with scrolling Twitter, right? So the sync mode is for the deepest focused and hardest problems where you need the centaur combination of human and AI. And so that's what I posted in the recent thing we shipped on SREGRA, where we have

the sort of async value of productivity, right? Like either you're super productive because you're in flow and you're focused and you're working on hard problems. If agents take longer, then you start to switch away and change context and lose context. And then later on, when you start to get more productive again because you're able to employ parallel async background agents on stuff that is like really commodity and like you can trust that all. them to nail it. This episode is brought to you by Blitzy, the Enterprise Autonomous Software Development Platform with infinite code context. Blitzy uses thousands of specialized AI agents that think for hours to understand enterprise-scale code bases with millions of lines of code. Enterprise engineering

leaders start every development sprint with the Blitzy platform, bringing in their development requirements. The Blitzy platform provides a plan, then generates and pre-compiles code for each task. Blitzy delivers 80% plus of the development work autonomously, while providing a guide for the final 20% of human development work required to complete the sprint. Public companies are achieving a 5x engineering velocity increase when incorporating Blitzie as their pre-IDE development tool, pairing it with their coding pilot of choice to bring an AI-native SDLC into their org. Visit Blitzy.com and press Get a demo to learn how Blitzy transforms your SDLC from

AI-assisted to AI Native. AI changes fast. You need a partner built for the long game. Robots and pencils work side by side with organizations to turn AI ambition into real human impact. As an AWS certified partner, they modernize infrastructure, design cloud native systems, and apply AI to create business value, and their partnerships don't end at launch. As AI changes, robots and pencils stays by your side, so you keep pace. The difference is close partnership that builds value and compounds over time. Plus, with delivery centers across the U.S., Canada,

Europe, and Latin America, clients get local expertise and global scale. For AI that delivers progress, not promises, visit robots and pencils.com slash AI Daily Brief. Today's episode is brought to you by my company, Super Intelligent. You've got 100 what-if ideas, but which one becomes an agent. Super Intelligent maps every AI use case across your company and helps you create an agent plan that you can actually execute. We match opportunities to your tech stack, your data profile, and your team. No more guesswork, just a clear path from pilot to production. If you want agents that deliver business outcomes, start with planning.

Go to B-Supert.ai and sign up for a demo. What this reflects to me is the richness of the, just the topic of AI coding. There's why you can do an entire summit about the variety of conversations going on here. What are some of the other conversations that you're trying to bring in, you know, that have been maybe part of previous summons that you've done? You know, evals, memory, context, like sort of what, you know, rag? What are some of the other big kind of big hitters that are going to be, you think, key parts of the conversation heading into this year's event? Yeah, I think memory and planning are always going to be huge.

Context engineering is obviously a huge theme this year. And we have the guy who, one of the three people that coined context engineering speaking. Dex is a fantastic speaker, one of the top speakers at World's Fair. And I think like then the other part is honestly just like organizational transitions, which actually uniquely as a podcast, you will cover, which. is rare, which is more of a leadership topic, right? Like, sure, like, the AI exists, but, like, how do you, like, move an existing large organization to take advantage of it, to upskill your team, and maybe potentially reorg in order to capture the opportunities, right? Like, I think, like, this is one of those things where, like, for the first time, I'm able to feature people from, like,

Goldman Sachs and McKinsey and, like, you know, like some of the top enterprises in the world. on northwestern mutual, you know, and like Bloomberg's coming back this year. There's just a lot of like very interesting, especially East Coast stories that I wanted to feature because a lot of tech is like very West Coast centric, but there's a lot of good stuff happening in enterprises too. Yeah. On the organizational change piece, one of the things that I think is really interesting about and I think to me was reflective of just how dominant the AI coding theme has been this year

is when we started, you know, when we were kind of first doing some of these agent audits around the beginning of the year, it was very often the case that the engineering departments were surprisingly some of the holdouts. They were the sort of most intransigent around wanting to adopt new systems. And while I don't, while my perception is not that that's gone away entirely, it does feel like there has been a major shift over the course of the year, perhaps as the tools have gotten better, as the models have gotten better, as maybe our understanding of how to integrate these systems has gotten better. Certainly not universal, but we see less and less, you know, just, you know, over my dead body kind of engineering departments

when it comes to some of these transitions. Yeah, totally. I think like there's a lot of knowledge sharing in this kind of stuff, but it's also like not fully well mapped out. And honestly, I'm waiting to hear from you and, you know, the rest of the speakers on the, on the leadership day to map out like the state of affairs and like what is working, what is not among the enterprises that you talk to. So speaking of that, one term, you know, basically going back to what you were saying about vibe coding, it almost feels like part of the challenge is that this same word or same phrase means different things to different people, right? I think that context engineering is going to be a term that has a similar bifurcation or potential bifurcation in the year. Because context engineering

is a very sort of like technical set of questions for engineering. who are thinking about how to design systems that better interact with context. But it is also now a leadership or sort of a change mindset as people like basically sort of akin to prompt engineering for individuals where organizations are thinking about context engineering as how do we sort of organize our data, broadly speaking, to be ready to be used by these systems. How do I think as I am prompting individually as a sort of, you know, a frontline worker in a company, how am I making sure that I'm giving it enough work context? And it's not that that's obviously a totally separate thing, but, you know, the one is not thinking about different ways for kind of

technical methods for the LLM to access different information. It's more of a mindset shift getting away from just strictly prompt engineering to making sure that your your quad skills are updated with all the things that they need. And I wouldn't be surprised if we see, again, there's sort of like the enterprise non-technical conversation around context engineering, which is going to be sort of like, a very broad use of the term context and a very broad use of the term engineering, as opposed to maybe the more technical conversation. Cool. I don't have a view on that yet.

There's something you're picking up better than me. So I'm curious to learn more. Yeah. It's a prediction, not a fait accompli. So the last couple of things I wanted to ask you about, move back to the blog post that I was mentioning, the Devons and the details. Two things that I think really stood out to me. one was your kind of very simple 80-20 sort of notion of code AGI.

I'd love to just sort of like hear about that a little bit. So the quote is, I'll quote yourself, you so you don't have to call yourself. But the line was the central realization I had was this. Code AGI will be achieved in 20% of the time, a full AGI and capture 80% of the value of AGI. So talk, talk, I would love to hear just a little bit about kind of how you think about that. I think it will resonate even with my. non-technical audience just based on how much coding has shaped what we've all done with AI this year despite not being coders. Yeah. Well, I mean, so I would say that there's a little bit of

self-cringe when I really boiled it down because obviously the world is never that simple, but you have to think about the highest order bit. And you have to think about concentrating your bets instead of spreading them out when it comes to power laws. And so 80-20 perito principle framing is the way that I do it. Okay. So and then the other irony is code AGI is a, I don't know what's what's the word for like self-contradiction. Because if it's general, it should be general. It shouldn't be.

Right, right, right. But like, you know, all that's all that aside, I think that the general sentiment is what I was trying to reflect, which is literally value catcher versus timelines. And I think those are the right two axes to really think about in terms of where to spend your time. and where to invest maybe, which are the same thing. You're investing your time and investing your money. And so I think, like, one, I think the obvious statements are all listed in there, which is how, like, code is, like, a verifiable domain. It's much faster.

The people working on the code are also the people, like, you know, consuming the models. So, like, there's just, like, a general virtuous cycle that is obvious in there, and, like, basically doesn't need any more elaboration. I think the interesting thing comparatively here is also the value side. instead of just the timelines, which is obviously happening now a little bit, but like you have to really, and for me to join a company that's worth $10 billion,

you know, like what's the upside? Like 20? Like no, like it has to be 100. And so I think like you have to really think through, like, is that even on the cards? And I think, yeah, probably. And that's mostly because of the flexibility of code. I think that the best way to communicate this

is with like how many people and how many times people have observed that you can use cloud code. to do non-coding tasks, right? Because it does generalize, it has the sandbox of pools. We used to, you know, in the chatbot era, only do, you know, embeddings retrieval, right? But now we have, like, agentic search, and that basically requires a document library, and there are all the things that people talk about in the, you know, the modernized LMOS stack.

For people who are interested in this, check out Jerry Luz talk from the 2025 World's Fair, and he talks a little bit about the emerging stack here. And so, like, I think, like, that is probably the case where, like, there's the things that we learn in coding agents basically generalize. And actually, the people who work on coding agents first will find it faster because they already have. Like, it's, like, super obvious to me that they already seen it. In some ways, like, Claude Code is a new foundation for Cloud itself. Like, when people talk about, like, the Cloud Platform or they talk about, like, Cloud for Finance, or Excel, which was launched this week, it's all based on a foundation that was built with QuakCode.

So it's like, it's funny because I'm not even really really putting my neck out on this thesis. It's just, I'm just calling it out as something that's already happening. Yeah. No, it's super interesting. Like I said, I think it's a fascinating way to look at things. And the last thing that I wanted to ask you about is, so I've said a number of times on the show, probably enough to start to annoy people, that I think two dominant themes heading into next year, at least for sort of like the business, the AI at work side of things.

One is, I actually think, is context engineering and just thinking broadly about what's the set of information that we need to provide, you know, whatever AI we're using for it to do better than just whatever it sort of can do out of the box. I think that's going to be a massive theme. And I think that part of why it's going to be a big theme is that by making it a theme, it gives organizations the license to do unfund, very difficult things like, you know, big data projects that were, you know, less set. sexy than like coming to this year is like what what demonstration agent can I build? I think going into next year it's going to be more like how do I get this entire house in order and there's going

to be you know, sort of wind at and wind of people's backs for that. So that's one. I think the other very obvious one is ROI and performance. I think it's easier said than done, but I think there's going to be a lot, a lot, a lot, a lot of discussion around, you know, how these AI and agentic systems are actually sort of impacting the world of work, you know, be it time savings, cost savings, new capabilities unlocked, I think that's going to be a major exploration. The third, which I'm just starting to sniff. And so I'm not ready to sort of call it on that same level is I think that I see this conversation starting around the product era of AI and the emphasis on products in which AI is situated being the things that people are releasing and focusing

on, as opposed to the conversation just purely being, you know, how does this model compared to the one that was 0.5 before it. And you had, it was not this, this wasn't the conversation, but one of the things that you talked about was the sort of difference between agent labs and model labs. And I love that just that if you want to share that framework, because I think it might have a stake in that larger conversation as well. Yeah. Okay, there's a lot in there.

So first of all, products era is a broader thesis than agent lab. Yeah. I think product era is basically, in VC terminology, is the application layer winning, right? Like, and definitely two years ago, application layer was very unsexy. People made fun of it. You're just writing GPT wrappers. Now they're like $30 billion companies and like, you know, 50X sales and Harvey and Cursor

and all these guys are doing super well. A bridge, you know, open evidence or what have you. So I think like, yeah, the product era has definitely happened, but the specific type of products that is doing super well is agents. So, like, that's how I make that transition. I think, like, as a product person, sometimes you can overthink it. And if you really just look at, like, what the heck people are actually having PMF with, it's just agents. Like, Replit spent, like, two years, like, working on AI products and, like, got nowhere.

And then they built an agent. And then suddenly, they're, like, at $300 million revenue. So it's, like, kind of obvious. And you just take it literally anywhere. Like, you know, like, notion, like getting serious. agents is very good for Notion, all that stuff. Okay. The agent lab is a thesis that isn't quite fully worked out yet, but it's really just the case for building AI companies in a different way than has been in the past. Obviously, I love two-word, coining things that are two words,

and I love, like, I love the way that like people start to organically adopt it, which is why I know this terminology is working because now people are saying it without even me being present in the room. The Asian lab thesis, I'm going to pull up this guy's coverage of my post, which is really helpful. It's basically like people shipping products first to their model first, right? A lot of AI companies in the past, they would just basically say they'll raise a bunch of money, announce it of a bunch of money, announced they have a bunch of cracked researchers, they buy a bunch of GPUs, and then you don't hear for them for six months or a year. And then they come out with like, oh, here's our model, you can't try it.

we hear some interesting updates from our model. That's exactly, by the way. I mean, I'll come right on and say it. Like, when we launched Street Grip and in Cognition, like, I was like, oh, this is why like magic with their 100 million token model never launched because they're a model lab. And Cognition is the agent lab, build the agent first, and then build the model. And I think that was like a back-to-front theme that I think really starts to play well.

It remains to be seen, obviously, because I think the bitter lesson applies and scales. in infrastructure and GPUs is king, how much of the relative value agent labs can capture with model labs. But I think that's really bifurcating. And it's so weird. Yesterday, opening I kind of proved my point. Like, did you watch the livestream from yesterday?

Basically, Sam was like, we're giving up on products. We're building, you know, infra. We have like chat GPT. We have SORA. But that's about it. Like everything else is third party. We're going to be a platform.

You should make more money than us on our model. right? He said all this. And like I think to to me, I've been watching open the eye as long as you have, that's never been that clear. Like they always wanted to be. Yeah, totally. I think it's probably been not clear to them. I think they've been debating it back and forth constantly. They hired a CEO of applications. That's curious because now they only have two. But like, you know, there's, there's going to be applications built on chat GPT. But like that's, that's a different thing. Anyway, so I think like now the swim lanes are very clear, right? You want to build AGI, go join a model lab.

You want to build products that serve users and vertical domains, build an agent lab. And I think like that's really what I'm seeing with the agent lab thesis. I think there's probably like more to flesh out here on like what a good agent lab looks like versus a bad one. But like I'm pretty curious. And I think like that explains the entire differences between the vibes that you get from agent labs versus model labs. I think one of the interesting implications, maybe we'll explore this in the talk in a couple of weeks, is it might force enterprise buyers to think a little bit differently. I think that it has felt for a while like you could effectively avoid pretty much all that's happening in the long tail and just deal with, you know, the foundation model companies or maybe the one sort of like leading vertical. player in your interest, like if you're legal, like maybe you deal with Harvey or, you know, if you're

in medical, you do it. But like, but not, you know, one of the reasons that I don't have a ton of space on the show to cover as many of the cool new products as I'd like is so much of the audience is like, well, I can, I just, if I use it in my personal life, great, but there's no universe in which that's coming in. And if it really is the case that the model companies decide that they really are going to be platforms and let, uh, let the sort of, you know, the agent labs build the next set. think you will have to see an expansion in just the procurement process, which is a very, very discreet part of the conversation, but an interesting one. Yeah, no different take on that. I think maybe the one whole in this thesis is maybe anthropic because they're really building a clock code to be

an agent lab within the model lab. And every model lab can easily build an agent lab, for sure. It is just a matter of resources and a matter of honestly social pecking order. To be an applied AI engineer inside a model lab is like low status. You're paid half what the researchers get paid, probably less if you're working in meta. I think it's interesting how seriously the lab's taken. And obviously there's a very, very wide variance. But typically, like typically, and I speak to plenty of people, you know, in those roles, they are more like the four deployed engineers, but they're not involved in research and the company clearly values research more. And it just, that's just how it is.

Well, Sean, awesome conversation. Kentucky for hours, but excited for the event coming up in a few weeks. Thank you for hanging out. And keep telling us where the future is. Yeah, I'm excited for your talk. Do you want to preview? What are you going to talk about? I don't know yet.

But what I do know is that I'd like it to be substantive as possible. So I don't know if you've seen, but I've got this thing live right now, R.O.I. survey. Like I said, I think I think that next year there's going to be so much conversation of ROI. And this is like the kindergarten version of ROI. It's literally like, tell us your top use case, which of these eight areas is sort of like the biggest area of benefit, time save, cost save, whatever. And then give us your subjective rating, you know, of it. Like how many hours per week? It's, it is so generic. But I still, you know, it's been live at the time of recording for like 36 hours and we have, you know, 250 plus use cases that people have logged in and said,

here's how it's benefiting me. And already that's such a different amount of information that we don't really have access to. So I'm hoping that there's something that's interesting there, maybe combined with some of the other other readouts and learnings that we've had from superintelligence. So it's not just me rambling. It's a little bit more data driven. But we'll see. We'll see what's ready by November 20th. Good. Yeah. The ROI of AI is a perennial topic, just like every other leadership thing. It's weird because I can just have the same schedule every year. Yeah, it's totally different.

Yeah. I mean, I hopefully solve some things. We'll see. But human problems will always make new ones, you know, to replace the old ones. But yeah, absolutely. I'm really trying to wrap up. Yeah. Thanks, John.

Let's see you soon. See you soon.