Anthropic cuts off Windsurf's Claude access overnight — AI Engineer World's Fair organizer Shawn Wang explains what it means for the OpenAI acquisition rumors

Thanks so much for hopping on. I wish I could be there. I have the biggest FOMO I've ever had with any tech conference because the lineup seemed absolutely fantastic. Unfortunately, I have a large family and and a studio that is covered in junk and we're moving in.

So, I appreciate you taking this remotely and we'll have to do the next one in person, but thanks so much for joining today. How are you doing? Hey guys. Uh glad to be back. Um I It's funny cuz like your setup looks so great on camera. I'm just imagining the mess that's off off camera. Oh, yeah. Yeah. Yeah. Yeah.

It's definitely a work in progress. Uh but part of the brand is of course just showing the nice parts. Um but but but congratulations. Uh this is a many years in the making. Can you give us a little bit of the the history, the plan, and then I want to go into some of the hot topics that you've been discussing today.

Yeah. Um I guess this is the fourth conference we've done.

uh we've done um and I started this basically uh right after the sort of chat GBT moment and talking with enough developers and understanding that the people who can wield LM APIs are going to be way more powerful than people who just chat with products and this is actually going to widen a lot because they can basically wield serverless intelligence.

Um and so I coined you know I sort of popularized the term AI engineer. Andre Karpathy sort of endorsed it.

He said that um he does believe that you can get very far without ever training anything which is a big thing for him to say because I think the status quo at the time was you know it's high status to train models research scientists but now it's actually consensus now that you want to work on rappers rather than models y and uh there are many many multi-billion dollar companies that um have spoken at AI engineer that reflect that fact and yeah so like I guess and there was a big shift there where where there was there was a moment where uh fun like the vibe had shifted a little bit to like the application layer, but there was still the idea that if you were going to build an AI legal startup, you were going to train an AI legal foundation model and now it's moved all to post- training, all to RL and how you're how you're prompting and how you're integrating and ultimately like the user interface.

And so, yeah, it makes a ton of sense and I'm sure there's a ton of companies that are beneficiaries of that. Boom. Yeah. Um, and it's not even companies. It's more just like customers. Sure. Like because the foundation model labs are never going to work directly with like your healthare system.

Like it's going to be like a bridge that um that comes up and and does that. They're not the foundation model labs are not going to work with directly with lawyers. It's going to be Harvey, you know. I think so that that directly led into the rise of vertical AI enabled SAS. And I mean that's consensus now.

But yeah, I'm sorry. Right. That's a Slack thing that by the way the conference is still going on. We just had the morning keynotes. Uh I have my chat with Greg Brockman later today. Fantastic. And he doesn't know this, but I can I might as well just share this.

But like we have a nice little cameo from uh Jensen Huang coming by. Um no way. That's great. So like we're trying to level up. Uh the c the keynotes this morning were fantastic. Like the Microsoft, you know, they they like really went on. They're they're going so hard after the AI audience.

We were just reading about that in the information that platform platform platform platform platform says Satia. Yeah. Yeah. Um I think they see this as their chance to overtake AWS. That's great. But like AWS is also, you know, also sponsor and like and having a very strong presence with us.

So like uh we just want to be the vendor neutral uh place, right? Like all the big clouds, all the big labs work with us. We have the the MCP team here with the entire steering committee presenting as well.

Um, and we just wanted to be for developers like this is where you this like kind of the trade show you come to hire people, learn about what's new and upscale. I have a couple more questions. How long do you have exactly? Five minutes or I can go till like uh 1:30. So Okay, great. Yeah.

So we had we had a guest and we we I moved into a different day. Okay, first uh perfect. So um I want to know about the mix of attendees.

How many uh how many folks are are trying to start venturebackable application layer AI companies versus is there a new trend of someone who's building more of a vibecoded almost lifestyle type uh AIdriven business?

Uh are there folks from either companies that have established themselves and are now trying to bolt on AI or are there lots of folks that are uh working for large companies and just want to stay ahead and become AI engineers? What's kind of the shape of the audience if you can characterize it at all? Yeah.

So, uh we we do surveys but we don't we don't know specifics to to to that high level granularity.

I would say about 50% are are people working at um medium to largesiz companies and trying to upskill and then the others are uh smaller companies and and you know our most popular title is like founder or CXO of like a a smaller startup and those are those are venture backable and I think mostly that's just a function of us being in San Francisco because obviously we we we will have that startup bias.

Um I do think that um one thing I I kind of don't really care about this whole lifestyle versus ventureback thing because uh for example I have a tiny teams track that speaking this year. Tiny teams is something that I'm trying to push as an idea um of u companies that have more millions in ARR than employees. Mh.

Right. So your your revenue efficiency is so high because obviously if you pay each employee less than a million dollars, you're probably profitable and therefore you don't actually need the venture money except to pour into marketing. Uh, and it's your choice. You can be profitable.

I have a sixperson team uh making more than more than $40 million. Um, and uh, yeah, I mean it's it's absurd like the amount of leverage you can get with with uh, agents and also building AI products for other people to use.

You're where you're just kind of passing through or slapping a margin on top of uh, the tokens that you resell from the big model labs. I think that really makes a lot of sense. Totally.

uh has the narrative of like, oh, if you're building an AI, like you're going to become uh obsolete by Google or or or or open AI is going to steamroll you in their next Dev Day keynote. Has that narrative dissipated? And what if so, what's driving it?

Is it kind of the pre-training scaling law wall that we're kind of seeing with GPT 4. 5? Are you guys still I I don't know if people still talking about that. I mean, everyone, you know, everyone's moved on. Yeah, we've moved on to different time, right?

Like, yeah, I think I think most people have agreed like there are still new pre-trains happening, especially with the open source model that OpenAI is working on. Sure. But yeah, I mean, I think we've just like we've seen it come and go enough times.

So, for example, we have OpenAI launching chatp codeex, which is just head on a direct Devon competitor. Yep.

But um you know Devon's not worried because they you know have been doing this for two years and they're much more polished in terms of the integrations and they have different things that their customers already like and so I think it's just like everyone is going to need their version of a thing and so this is the sort of house storeband version of what CHG codeex could be and it's it's not competitive just because the the ocean is so huge for software engineering and Devon has at least established the category by being first there.

Um and I think you can see similar versions of that across across the the domain. Um yeah, I haven't I haven't really seen anything there yet. Um although it's not to say that it doesn't happen. It does happen.

For example, with the first wave of GPT3 startups like Jasper um uh but it I mean yeah so far there's no fear that in fact people very excited to meet the foundation model labs like this is where the engineers meet the lab people um and lab people train them on how to use their LMS.

I think that's that's a perfectly harmonious relationship to be honest. I I haven't seen What was your reaction to the the news around anthropic and wind surf yesterday? Uh a little bit of drama on the timeline, but so yeah. Yeah. Yeah.

Uh give us uh give us a walk through for those who might not be familiar of what happened and then I'd love your analysis. Yeah. So the history of this is that um Windsurf is an independent company that uh basically kind of followed Kurser's footsteps and launched an AI agentic.

They've done very very well for themselves in a very short amount of time. I think they launched we're the first podcast that they launched with in October or November last year. Uh and and then they there's been rumors that they got acquired by OpenAI for $3 billion.

Uh those are rumors that are unconfirmed by both sides. I've talked to both sides. Yeah, it's it's I've been fascinated how everybody just takes it as fact. I take it as fact. He wore two polos. Nobody Yeah, he wore two polos.

But again, at no point every every news outlet, legacy news outlets have been reporting it as fact, even though no side has verified it and nobody said it closed or anything along those lines. Yeah, the the only thing I can say is um I Windsurf is speaking tonight right at right before Greg Brockman goes on.

Um, and there's a reason. Um, exciting. We're not like we're not dropping a ton of alpha, but I'm just saying like I'm not I that's all I can say. That's all I'm allowed to say. Yes, of course.

And I I think that so there's there's there's a there's a relationship there just as there is uh a fact that both cursor and windsurf have benefited a lot from their relationship with Enthropic because cloud 3. 5 and 3. 7 have for whatever it's worth been regarded as the best coding models. Yep. Open will disagree.

Gman will disagree but whatever like the community has voted. Yeah. Um so overnight I think like two days ago one day ago uh anthropic cut off the first party access to claude to uh for cloud to windfur this is their top model and now they just don't have access as a as a first party tool.

You can, for example, still bring your API key and use your accounts on Windsurf, but it's right. You can't just like like Windserf, I'm gonna pay you 20 bucks. Use your account, whatever. I, you know, I just don't want to worry about the the the rate limits and stuff. Um, that's gone.

Windserve, they just woke up overnight and and that's gone. So, a lot of people are very upset. Uh, this is a very big no no if you're like aiming to be any sort of like credible LM API provider to just cut off access.

um Google hasn't done that even though you know if you take the the sort of rumors of the acquisition on face value um but I think this does lead credence that I think uh anthropic at least thinks that Windsurface is just a competitor product now so I think like in so far as you're putting odds on whether the acquisition actually closes the odds have moved up interesting yeah the other the other factor that's interesting is how much does this benefit cursor if you're a developer who loves various cloud models and Now maybe you have a reason to go spin up.

Do you know market share of cursor versus windsurf? Have you seen like estim estimations? What is it? I I think the numbers I've seen again I I really I saw it on on my timeline on Twitter. I didn't save it and press but someone please look this up. It's something like Windsurf is 5% of cursor. Wow.

Because based on the based on the market caps I would assume that windsurf 30% I know of cursor.

So the thing that you don't see there is cursor has entirely developed business on the IDE where Windsurf used to be codium which used to be like this GPU serverless GPU company that has significant LLM inference uh for enterprises like they they spent the last four years doing that. Yeah. Yeah.

Um so they're buying that team that revenue that product as well as Windsorf. Oh interesting. Yeah that makes a lot of sense. Um what uh how do you think that the uh agentic coding market is shaping up?

we've been kind of talking about maybe it's three different markets like I can go to GPT uh like I can go to 03 and I can ask it for a question and it will just write code and write Python and execute it and I don't even have to tell it to write code it's just if I ask a question that requires code it will write some code then of course they openAI now has codec and then they also potentially have windsurf or or you know you can think about the idees uh as a different um as a different uh entry point into that market uh are we seeing like a true permanent bifurcation between uh synchronous and asynchronous uh AI coding usage or do you think these all blend together at some point?

Yeah, I think that actually you're just catching up on something that has already been a thing. Um that was the divergence and actually they're converging.

So they they already like it already started that you had the synchronous ones like cursor windsurf and then you asynchronous ones like Devon and Factory AI for example. Yeah.

And they were all sort of like inner develop what we call the developer inner loop which is hands on keyboard coding which is developer outer loop which is uh PR review and all that kind of stuff or like file an issue make a PR that kind of stuff.

Um so those are merging the explicit goal of codeex is something I'm going to talk to Greg about uh later today is that they will merge those paradigms just that they're merging the reasoning models and the the sort of instant thinking models uh with GPT5. Um, and I I that totally makes sense.

It's technically really hard to do because once in one's on your machine, one's one's cloud. And also they're just different set of user experience paradigms. Is my uh video freezing by the way? It is. You look great. You look great. I see you guys. Great.

Uh, how how seriously do you take Google's uh coding agent jewels? Do you think it's just an experiment that they're putting out there or do you think they'll actually invest in it and try to get real adoption? You're trying to get me in trouble. You're conflicted. You're conflicted.

I'm sure I'm sure Google has has is a part of Google's great product wise. They have had hits like no LM that then have failed to continue the momentum for whatever that I and I think that is I think that most people are extremely unfair to notebook. I think they have really well.

It's just that you will never repeat the initial wave of excitement that they had. Totally. Just never will. Like Yeah. And I mean Notebook LM seems interesting because that feels like a case study in if you're a startup and you're fast following an innovation that actually came out of a hyperscaler.

That's where you turn into a bullet point on the next dev day or something like that. Because Notebook LM it felt like there there was this amazing tremendous momentum online. Everyone was excited and then it was kind of faltering and everyone was like, "Oh, there's no app.

" Uh, so maybe I'll build the app and I'm sure some people built notebook LM apps or notebook LM spin-offs and then it just took it just took six months and then there now there's an app and now and I'm sure they will continue to iterate on this and so yeah. Yeah, I I I know to some degree.

Uh but but but I I think the lesson is probably like yeah there's probably some narrow window where you can cash grab as kind of like a cynical like ripoff startup but like realistically if you if you found out about the concept of turning uh a deep research report into an audio product from notebook LM like you might be just behind the ball and so like like like you don't you don't have the right to own that market long term as opposed to you know I think dev and cognition like they have more of a right and and and factory they have more of a to uh continue to fight with Google's uh Jules product because they were really there earlier.

They can they issued the original hype and and and brokered a whole bunch of enterprise contracts and like kind of got a couple years down the road before. So they're not like a fast follower to big tech. They're actually uh big tech is maybe fast following them. Uh yes.

I I think I think it just boils down to execution and everything, right? Yeah, of course. And I think like uh that's that's one of the reasons why we've actually somewhat broadened out from just engineering to we've added AI product management and AI design. Okay.

And I think that really good this is this is product management. This is just straight up can you keep up the m momentum? Can you listen to your users? Can you come up with creative new stuff that keeps the momentum going and some some teams can and some teams cannot.

Um the only thing I can say with regards to Google and like the fate of jewels is um the initial the founding PM of of Nom left and she's speaking here with her own startup.

Um it's like the the it's really hard to keep employees of people who like start interesting products for Epic Lab because they will get any number any amount of money thrown at them. Yeah. Open it up. $20 million seed round, 10 million of secondary if you quit your job right now. Uh probably not that extreme.

Um I I wanted to ask you about uh Yeah, cheap. Uh uh I wanted to ask you about GitHub. Uh how much do you think Satia cares about, you know, GitHub? Uh do you think he's, you know, really pushing? Yeah. Is that an important wedge into the AI coding market? because obviously they were first with GitHub Copilot.

Um, and you could imagine that that it's just a phenomenal distribution channel. And when you look at the the value of and the way Microsoft executed around Teams rolling into you you're already using Outlook now you're using Teams instead of Slack. You could imagine really great adoption.

Maybe not at the most cutting edge companies that are hyper online, but we could see one of those charts where we're like, "Oh, wow. " like Microsoft's really crushing it in in developer tooling. Uh I mean uh absolutely I I think it's extremely core.

Uh what would you put GitHub co-pilot revenue at right now if you want to if you had to hazard a guess? 500 500 million. Yeah. It's right there. You nailed it. Let's go. Standalone that is a publicly listed company. Wow. Yeah. Yeah. Yeah. That's like the perfect Here's the thing here.

I mean here's here's the challenge, right? Here's the challenge is that no one is talking about GitHub copilot. Is that a challenge? It's inside of Microsoft. So yeah, I mean no one's talking about No, I know.

But you know, nobody's saying like look what you know there nobody's sitting there being like I'm blown away by Copilot, right? I I think I think uh I think that people who live in the Microsoft stack um are because that is what they have access to and like their company agreements are.

I think on Twitter there's a novelty bias, right? People people always want the new thing. They want to they want to support the little guy. It is I support the current thing. You want to support the current thing.

But okay, here's the mental framework I want to leave you guys with on on that stuff is um data gravity and the mo is a thing like data has attracts the compute attracts the money attracts stickiness like once once my is one place everything else moves towards that data and um if my code my code is the the most valuable form of data I have that's that's the most uh expensive uh data to acquire um so if my code already lives with you then it's much easier for my coding agents to also with you and uh and so so they they have a home turf advantage.

I think the GitHub acquisition back in the day was one of the smartest and most enterprised acquisitions uh in in developer tools history. Um so like I think I think you should you should take that point of view.

So like that that moat is so strong like any like try try to be like a no-name random dev tool startup that that comes up and says hey give me access to your codebase like like I need to run agents on it. They're like, "No, like I I'm already doing Yeah. What hap what happened with codeex yesterday?

There there was one user that was reporting that uh private repo information was leaking from one to another. Was that user error? Did you did you track that at all? " Now, you were probably getting busy. No, I I was I was running the conference. Uh yes, and individual things are going to happen.

Um I think GitHub's reputation is is going to last through that. I think totally it has to be really egregious for for that kind of stuff to slip through. Yeah.

I I mean I think like more broadly I think that there's a standard stack of what we are calling what Andre Karpathy is calling the LM OS right that I think you guys are kind of going and learning about but I think it's it's pretty well established by anyone who's who works in software agents coding agents which is you want a sandbox you want a browsing environment you want branching you want u fine-tuning on your codebase like there there's just a standard stack of things that cloud codeex jewels cognition um um factory and anyone else in this field, they're all converging on the same thing and it's just who does it best and who reaches who who solves it for their customers best.

Um I think that's that's going to be the the name of the game for now. Makes a ton of sense. Uh do we have to roll over to the next person?

Uh no because I I heard I mean I guess my question on data gravity is uh it it seems like uh if we compare the evolution of video generation models Google has a really amazing cornered resource in YouTube because that data is so big. It's so many tokens.

It hasn't been it's growing and it also hasn't been like exfiltrated to the public web. Um, but it feels like GitHub, although it is a massive data set, it's not nearly as big. I think it was I I think somebody clocked it at like a hundred million tokens or something. It's it's just not it's just No. How big is it?

Way bigger. Well, public repos because I I I don't think that they can train on run private repos, right? Or can they? Uh, who? GitHub GitHub. Yeah. GitHub. GitHub. I don't think they would let themselves. Yeah, they wouldn't. Yeah. Um, I don't know. 100 million just sounds way too low.

And that's my that's my gut reaction. Sure, sure. Keep going. Keep going.

Uh so so so the question is like how durable is the data moat at encoding versus uh in video generation purely I'm just thinking about the the raw data set size of YouTube has to be orders of magnitude larger than GitHub and so I would imagine that it's it's Microsoft probably has a less durable advantage versus Google's like kind of march down the video generation pathway if truly uh restricted by data.

Not all data is comparable like that. So you're comparing verifiable data to unverifiable data. So code can compile and if it runs it if it runs you can do more code like that. Um and that creates the RL loop that lets you generate synthetic data for for more code.

Uh with videos you have what you have and you can train on that and uh so you're technically not even allowed supposed to train on that but who knows what Google's doing behind the scenes. Well, you think Google can train on public YouTube videos, right? I don't know.

I I I imagine that like I I've posted videos on YouTube. I imagine that I've opted in at some point. I think it there was like a big fuss with MKBHD and all the other guys Yeah. making up a fuss about this. You remember? Sure. So, I'm just I don't know.

I'm not a lawyer, but like I'm sure is a completely untested clause that just has to go to court. Yeah. I just want to say if uh to Google if you remove John's rate limits on VO3 you can train on our back catalog full permission please. Yeah, I I think I think there is that diversity of of opinion, right?

Like do you want to just like give yourself to the machine or do you want to keep your keep your data to yourself and it's it's like there's no in between like you're either one or the other. It's it's a very strange dichotomy like it's not a spe like many things in life are a spectrum.

This one is not like you you run into someone they're either Maxi privacy maxi or they're like an AI maxi. Yeah. Yeah. Yeah. No, that makes sense. What what else is on the cutting edge of of of debates that are kind of raging at the conference uh this year versus prior years?

We've we've lived through like the P Doom debate. We've we've we've we lived through like the Leopold and Brener era. We've we've shifted to the geopolitics debates, but what's kind of on the frontier of like hottest topics to discuss?

Uh first of all, Leopold was right and uh I don't know if people know that he was on Ilia's team when the the the sort of board drama happened and like he was like directly connected to it. Anyway, he we haven't lived through it. We are living through it. It is happening.

It is directly leading to the geopolitics because he was foreseeing that and he goes exactly right. Yeah.

The way I would the way I would uh I would kind of characterize at least my takeaway was that his piece uh was a uh was a little bit of a pivot from the paper clipping doom and a shift to geopolitical competition in AI. Is that is that a mischaracterization? No. Nailed it. Okay. But like he was right. Yeah. No. No.

I I I completely agree. I completely agree. I'm just wondering if if if it's like if the consensus is that he was right, then the book is closed. We're not debating that anymore. What are we debating? What's more are the debates? Yes. Yeah. How to do great AIPMing. How to run a tiny team.

We have a robotics track for the first time. That is the Tesla Optimus is is speaking. Physical intelligence. Whimo. Whimo just overtook lift. I saw that that already. Um voice is the hottest thing um in in terms of multi modalities.

Um like everyone's sort of building with voice because I think it's like finally good enough. Yep. And um I think maybe the last thing I will highlight to you is we are also emphasizing security for the first time. Um security is like kind of a boring topic.

It's nobody really wants to talk about like how to secure your system but like they actually do now because they they have real money running through their their product. So there's all that and then that is like roughly important equal in size to the excitement about MCP.

And so we have an entire MCP track with the anthropic team here. Very cool. because it's uh because they're nice enough to to to come by and that fills up the the whole ballroom that we have. So, are we going to get payments in MCP? Is there a subtrack for putting stable coins in there or something like that?

Uh there there's nothing in the official spec, but we have a number of people and we have a speaker talking about the sort of MCP economy that's being enabled. Um I do think that remote MCPS and authorization give you the foundation for basically just remote agents that you can buy and hire.

Um, and that that's that's that's effectively what it is. It's just that we're still in that stage where there's still these things are still trivial enough that you can actually just write your own. So, you have to overcome that build versus buy for this to actually kick off. Okay.

Um, and uh I would just say like I'm not at all a crypto person, but I would say that the crypto people have been ahead of us here and they haven't found that much yet, you know. No, makes sense. They're but they're actively hunting and if anyone finds it first, it'll be them. It's it's circles presenting here.

Solana presented at my previous conference and uh they're all they're all on it. Stripe is also on it with with their stable coin thing. Um something's going to happen here. Yeah. Uh in in robotics, are we getting to a point where we're starting to see an ecosystem of companies pop up like we've seen?

No, it's just it's what what do you Yeah. So they all vertically integrated. Sure. They all building every their stack. I don't see like any horizontalness if that's what you're going for. That's exactly what I was asking. Yeah. Yeah. It's so weird.

I think um you know there's just so much custom needs that you you have to sort of reinvent the universe every time. The one uh thing that that does reuse bots is a cloud chef uh which which is um you know a very young company but I I ran across them when I tried their food. So they're they're a kitchen robot.

They do demonstration learning from a single shot uh from let's say like a Michelin chef and you just give them you give it the ingredients and and you will uh it will just flawlessly just kind of repeat that that cooking for you and you can hire it for $12 an hour.

So it's meant to directly replace uh human labor uh which is very expensive and very labor intensive and unreliable. Let's let's put it that way. Sure.

Um and uh they they use they can they can live on top of any other robot arm or Tesla Optimus substrate um which okay um because it's more about the the robotics sort of uh framework than it is about the individual uh hardware. And I think like that is the first time I've seen that happen.

I'm like relatively new to this, but like this is the first time I'm like I'm like, "Oh, okay. " Like he's actually pretty confident he transfers across any any system as long as they have the minimum required set of device drivers basically. Um, and I'm like, "Yeah, that's cool. " Like, yeah. Yeah.

So, he's he's speaking tomorrow. What about other pieces of the robotics data stack? Are we thinking about like data brokerage or or kind of like a scale AI or maybe scale AI actually working to generate more robotics data? Anything on like the sim tore gap?

I saw some semi analysis summary of a paper that was pretty cool. So I would say that as an as a sort of industry practical conference those are just in the domain of research right now. It's in research. Yeah. Yeah. It's it like there's there's a lot of papers out there.

Um Jim Fan had a fantastic talk at Sequoia Ascent that I I recommend everyone watch if you haven't seen it about the physical touring test. Yeah. How you need to uh just do a lot more simulation. And there's a there's a lot of work on being done on this.

Um, I think Google Genie is the other one that I would re recommend people do. There's like a really interesting tiein between like the generative video world and the robotics world that you wouldn't necessarily expect until you like spend some time in that.

Um, but we just haven't focused on it because people cannot get jobs in it. Like I I want people to get jobs at my conference. That's like almost the whole reason of bringing the companies, I bring the engineers, they meet, they they fall in love with each each other.

That's like to me that's like the most fulfilling thing. Can you give a a highle update on the hiring market today? What what fact? What's fact, what's fiction?

Uh, you know, it's one of those things people love to, you know, talk about how bad things are and yet every company that I know is like can't find people and maybe there's a talent kind of gap there. But, but what's your read? Yeah, I mean uh I I think both are right, which is weird.

Uh this is actually the topic of my conversation with uh Greg Brockman later today because I I half the people I I meet at my own conference are worried for their own jobs, right? They do ask about this.

It's not like they're not they're blind to it, but they're just like I'm here to see what else I I need to do or how how how do how does my skill set need to change for the future?

And nobody knows, you know, everyone says the trit thing, which is like, oh, I'm going from I write the code to I manage things that write the code. So I become an engineering manager, but like what does that mean in practice? How much knowledge do you need to supervise an agent?

You actually kind of need a lot uh when it goes when it goes off track. So like it that's super unknown. Um I do think that the juniors that the like fresh out of college students, they're a little bit cooked unless they're unless they're unless they're good. So it really amplifies skill issue.

It really really my skill issue. Yeah. Yeah, it's a hard thing to to to you know try to uh tell somebody you just need to be better five times smarter than you are and then the job and then you'll have like a bunch of offers.

But I also think I also think that the the solution here and and we we just hired uh an intern who already shipped a new product. He he started uh Monday. He just shipped uh something relatively simple but it's cool to guest directory. He shipped he just sent us a V1 of that product that he built last Thursday.

We saw it and we were like, "This is super cool. Why don't you start on Monday?

" And it's that one kind of interaction that can get your foot in the door that allows you to develop skills and build relationships that, you know, hopefully we, you know, we work with him for a long time, but even, you know, I'm sure people are already seeing him at TBPN and and we'll probably try to poach him, uh, before the end of the summer.

So all it is it is like a changing shape of the software engineering employment because like I feel like like the vibe coders can come into organizations that might not have a full engineering department and then have an impact because of the way the tools work but uh it's certainly it's certainly a lot of change very very quickly.

I would I I would say that's not the consensus yet that hire a a like chief v code officer. Yeah. Because like humans are still very much needed to to patch the gaps that the models are not good at and it's really painful when they do go wrong and they do go wrong.

You know, I think like all the hypes that you see on Twitter uh fails to admit like what happens like one month, two months, three months after. And that's honestly what you get paid to do as a software engineer. Make maintainable software, not Yeah. green field all the time.

I've even seen people repurpose old old projects and say, "I I vibe coded this in 24 hours and" and everyone's like, "Wait, no, this was published in 2019. " Like, "You spent months on this. " Uh, and you're just hyping this up. That's misinformation.

Um, I I I can't I don't pay attention to what people say there, but I do care what people report for themselves that they they've done at serious companies. So one thing I would highlight there is Booking.

com did about 10 um company years worth of migrations uh in 3 months with uh with their sort of automated um uh code migration that they that they did with source graph. So um we did a talk with them at my conference in New York uh in February which if you want to look that up.

Um so like those that company has been around for like 14 15 years. That's that's a serious code base, right? Totally. They're just reporting their success. They got nothing to to gain from selling you on it. They're just they're just happy about it and they want to teach others. So you want to look for those.

They're not trying to sell you anything. They're just looking at reporting like their progress at a real company. Um and it's that's really what I try to optimize for for the for the conference. It's it's hard because not everyone's incentivized to do it.

You just have to create an environment where they get something out of it by meeting their peers at at a thing like this. Well, give us the plug. Where can people watch your talk with Greg Brockman later today? Uh yeah, it's on YouTube. YouTube. engineer doot. And yeah, I'm What would you ask him?

I mean, you're having him on eventually. So, you know, what's like the the way that you would open the conversation? Oh, that's a good question.

Um, for Greg, there's a lot, but um, I mean, the question that I feel like is at the top of a top of mind in like Ben Thompson world this week is, uh, the shift from training to inference and how and how workloads are are are shifting.

This was in the context of an NVIDIA earnings, but Greg obviously has intimate knowledge and so um in uh OpenAI is obviously going to do larger and larger training runs, but they're also doing tons and tons of inference. How is that shifting?

And then what I want to know, and I don't know how much he can speak to this, is is is like where how will if we get to a world where we're in 80% inference, 99% inference, something really extreme, does that change the type of data centers we're building, does that change the chips that we're demanding?

Are we moving to A6? Are we moving to FPGAAS? Something like that to like speed up the actual inference workloads? At what point do we actually bake these models down if they're deemed like good enough and then we're like orchestrating them? If we hit a plateau, that might make sense.

Um, at the same time, if there's a lot of promise on algorithmic progress and we're expecting like, oh yeah, we're going to leave the transformer behind eventually. Uh, well then, yeah, you don't want to bake all that down.

So, I think that that's like it's a little bit more of like a semi question, but it it is an interesting question for him as well as he's like seeing the workloads. It is very relevant. Um, he he does have a role to play in Stargate that I'm not super clear about.

I'm going to ask him about and uh yeah, I mean I I I I I would say that nobody's betting on the end transformers except for very small very small number of people that are just experimental. Um and uh it's it's really about just scaling the RL. Um yeah, I mean there is this interesting uh thing that's happening.

I believe images in chat GPT seems to be using a number of different algorithms combined. Uh so there's a little bit of diffusion in there potentially. There's some uh there's some transformer based stuff.

I don't know if you have a if you're pushing back on that, but that's I'm not there there's a there's a strong hypothesis based on the hints that they have dropped from the people that worked on it. Yeah. That there that there's multiple algorithms at work, right?

It's it's a it's a diffusion head with the transformer backbone. Yep. The rest of it is just regular 40 and that's so good at chat and has has decent world knowledge. Yeah. There's not like a ton of complexity there beyond that that we know of. Yeah.

And so what I'm interested in is like we saw the demo from uh from Google on diffusion, textbased diffusion, 900 tokens a second. Uh I've heard good reviews about that. I've heard kind of mixed reviews on that.

Maybe it's not a path, but uh what does what does the future of an LLM look like if we're if we're applying the same path that we've seen in images where we're seeing an ensemble of models come together to create a better like even more multimodal multi multi- uh like multiarchchitectures uh because it feels like we're moving it we're moving further and further away from the single big transformer being the answer to everything and and so what does that mean?

Are we going closer or farther away from the single big transformer that is like the the god in the weights? Yeah, this is where I can offer a little bit of coaching for uh for your audience and maybe you you guys have talked to a lot of people. I don't know.

This is where people usually say the term mixture of experts and they're wrong to use that. Just and sometimes they say mixture of ages, mixture of experts. They're they're they're incorrect. Um anyway, uh we have talked with a lot of uh the the frontier lab researchers. Yeah.

No, none of them believe that that is the way forward. Um that is that is an optimization for the current thing which is which which one mixture of or mixture of architectures. That's mix of architectures.

Um let uh okay um yeah so experiments for example with Jamba with from 1821 in Israel where they mix for example like a mamba uh layer with with transformers. Um that seems to be promising but even Shazer a correct character was just use scaling out transformers.

Um I I I asked a variant of the question you just stated to Nome Brown in an upcoming podcast that we did. We recorded it. We haven't released it. He's the same way. He's like, "Anything you're trying to do fancy around mixing up weird architectures is just not going to scale. Just scale out the the the basic thing.

" Um, and he's pretty strongly convicted in that. I I know people I can't name a Gemini who is also pretty strongly convicted in that. Um, I have reason to suspect otherwise. Obviously, like you can't shut this shut down anything because like it might be true.

Um, I'm just telling you what the people working on this would say because I've asked. Okay. Well, I mean, if that's AI, the last thing I wanted to kind of cover, I I would want to get a sense of, you know, prediction markets are pricing in that uh Google continues to dominate on benchmarks.

How much is OpenAI even going to how much do they even care about being at the top of benchmarks? Is it is it purely an ego thing or is it just about uh you know, usability and and value for users and code quality and things like that? It's usability and value.

Um uh you I think you need to notice that the benchmarks that Google talks about at Google IO are no longer the benchmarks that OpenAI measures itself on and um you know I want I want to stay friends but like there are some that are more sus than others. Sure.

And um I think like at the end of the day the the the customer will win. Um, the only one that that has not that's clearly been caught out kind of lying, kind of like being whatever like underperforming, I got to say it, it's Apple.

Um, when they launched Apple intelligence, we were all very excited about the Apple intelligence paper, which is a which is a beautiful document that had no EVLs apart from their own internal eval. So, they were not accountable to the standard set of evals that everyone else has, right?

So, um, step one is you you should try to hold yourself to some public benchmark.

I flawed just do it you know but and then uh so so so Apple did not do that um and and then like now there's a question of like are you holding yourself to a benchmark that can be gamed um and there's been accusations of L Marina being gamed and unfortunately that's what Google has put all it ships on they may do that for Gemini 3 but for 2.

5 that's what they did u and probably because they spent the last year doing it it's all revenue from here on out like that is the revenue is the benchmark GitHub copilot 500 million size gone for GitHub revenue is also tricky.

Yeah, you guys know how like Gemini I don't know if you know Gemini gives a billion tokens a day per human. What? Your personal API token. Okay, I'm kidding you. You can you can you can you can set up a camera right now, run Gemini 2.

5 flash on it, uh run like a frame a second and ask you to do whatever you want for free. Wow. Um it's absurd. So for Google, thank you. Intelligence too cheap to meter. We love also. Okay. Like that chart that you saw at Google IO where there are tokens going like this. Now you know why it's free. Yeah. Yeah. Yeah.

Yeah. Yeah. Makes sense. I mean it's they also have a very very good model. Don't get me wrong. But I'm just like revenue is a short-term play. Like you're trying to maximize revenue now.

You might cut off cut off yourself from getting the largest data set in the world which is the uh the sort of early adopter humans that are going to give you their data because they're like train on me daddy. I I'm you're giving to me for free like whatever train on me.

Um and and that's a that's a deal that lots of people I think a theme that we we talked about is lots of people are going to make. Yeah. Is free why they they can eat your data. Codeex is free. Why? Yep.

Um so I I talked about this in in one of my recent posts like the the the marginal cost of software has is is has gone down to zero and like uh can it go negative? Will I start paying you for uh for me for you to use me? Absolutely. Because then you become my labeler. Yeah. You're you're a labeler I can never pay for.

You know, usually have to like hire someone in like the Philippines or something. We've maxed them out. So now I got I got to I got to pay someone in Silicon Valley to label my my software. So I might as well just give you my coding agent for free or like even honestly just pay you for good feedback on my coding agents.

And why should that why should it stop at like $20 a month when like there are people in New York who I've heard by the way get they're like for they hire former bankers they hire former hedge fund guys pay them 500k a year to label. It's wild. Well, thank you so much for hopping on.

Uh good luck with the rest of the conference. Let's make this a regular thing. Fantastic conversation. Thank you for taking the time to your event and we'll be there next year whether you like it or not. Fantastic. You're invited for sure. We'll talk to you soon. Thanks so much for Thanks, Sean. Bye. Cheers.

Uh, fantastic. Well, uh, we got to tell you about Wander. Find your happy place. Find your happy place. Book a wander with inspiring views, hotel, great amenities, dreamy beds, top tier cleaning, and 24/7 concier service. Is a vacation home, but better, folks.

I want the TBPN army to clear out every single wand of summer. Mark clearing order inbound. Um, this is an interesting one. Oscar has joined the Fortune 500 for the first time. I know what you're thinking and I agree. What? Only now?

And Josh Kosner says, "I'm deeply proud of the tenacity and pers persistence of the Oscar team. We've been humbled many times. From down 95% from our IPO to now entering the Fortune 500. I am excited for all that is ahead of us. " Wow. Congratulations to Josh Kushner. That is a phenomenal run. Amazing.

And somebody else was quote tweeting this saying that um there's only two people who have incubated a Fortune 500 company who are actively investing and it's Palunteer with Peter Teal and founders fund and Josh Kushner with Thrive and and Oscar. So uh what a phenomenal run.

It is rare but the incubations are starting to work and the incubations will continue until morale improves will continue until the global economy is worth is growing at 10% a year.

Um anyway, uh every person who works in tech needs at least non five non- tech friends to interact with closely on a weekly basis so they can understand how the general public actually thinks says Kit Volulta. You posted this from a different account. It was a copy pasta. Somebody copied the whole thing.

So I found the original shared you shared the fake one. The ripoff. I shared a ripoff. Yeah, you shared a ripoff and it got community noted and I found the original one right here. Wait, so somebody just stole it? Yeah. Yeah. They didn't even screenshot it and put it in banger archive.

They actually but 20k on this post. The repost has 15k likes and so it got but it got community noted. This is a copy paste. I you feel like that could be uh that could be like automated in axe where it's like this is the exact same text as has been posted elsewhere.

Uh you should just know that you should maybe go to the source on this if somebody's posting the exact same thing. Except you know Josh Diamond would get that every day because he's constantly boasting good morning we are going to win. And so he would have the note every single day.

Um, JT uh, Jira tickets says, "Web design, you mean dig digital physioamy. " Very funny. Oh, you put another one in here from JT. Guy who is burnt out from his six figures, five hours a week big tech job. You got to just hit the hit the hit the roof and just uh get some sun, you know. Anyway, Pavle has a good one.

I don't think Whimo can work in New York City because during rush hour, you need to break the law to get anywhere. And I'm not sure how you could get away with embedding that in software. And IB says, "Travis would have gotten it done. Never too late. " Somebody says, "Full self-driving gladly speeds for me. " Yeah.

Uh, I thought this was good from Will too thick scoops. Uh, okay. Sounds good. Thank you. And it's Jersey Mike's order pickup at 12:48. Dutch government collapses over migration dispute. Why is this? This is a Gmail. Oh, it's just rolling up multiple emails or something. This is Apple AI summary. Oh, yeah.

It's an Apple AIM. intelligence is the summaries are so good. It's genius. 75,000 likes. That's a lot of attention on that. You got to upgrade, folks. You're missing out on entertainment. Pay Apple to help me produce Apple intelligent viral bangers. Yeah.

I want to wrap up, but we got to say congratulations to Jesse Michaels. He got on Joe Rogan. He said it was the honor of a lifetime to sit down with a goat himself, Joe Rogan. Many people have called him a young Rogan. I've called him that. He is fantastic. He's been on a fantastic run with American Alchemy.

his podcast. He's climbing the charts. If you're into aliens, AI apocalypses, Bob Lazar, nuclear weapons, secret anti-gravity research, go check out uh the Jesse Michaels uh Joe Rogan Experience episode. It's episode 2,31. Rogan's putting up big numbers. He's giving us a run, but we're catching up. We're catching up.

That's wild. Um in a couple years, lastly, just wanted to cover this quickly. AMD has acquired Bream. Okay. Reported today. Interesting. Uh to help reduce Nvidia's market dominance when it comes to AI hardware. Talking to George Hotz and and we're going to have him on the show, but uh he's been he's been back in AMD.

We got Dylan Patel coming on the show Friday. We'll ask him more about AMD and their their plan to unseat CUDA as the dominant AI platform. Yep. And we have other news. We are officially we have doubled the size of the show once again. Jordy, hit that gong. Hit the reel. Hit the real gong for me. Grab this.

We're going to gong cam. Uh, we got 64,000 followers on X. Thank you for everyone who's been around. Oh, there we go. That's great. Thank you for watching. Thank you for supporting us. It's been a fantastic journey. It is a pleasure. We got the whole crew in the studio. We got a whole bunch more cameras here.

We're growing every day. Damn, we got the studio cam here. The studio cam's going. Thanks for cooking. Thank you for watching. Thanks you for all the support review on Apple Podcast and Spotify on Apple Podcast. Yeah, I feel like we're doing low there. Yeah, it doesn't really make sense.

But if you're on Apple and you're listening and you haven't left us a review, we'd love a review. So, thank you so much. And we'll see you tomorrow. We have an awesome day. We have a we have a bit of an AI day coming together with folks from OpenAI and Anthropic coming on. Uh it should be a great show. So, stay tuned.

We'll see you tomorrow and have a great rest of your Wednesday. Goodbye.