ジェンスン退任？ Fable 5.2とAnthropicのIPO、Gemini 4 CarbonとHydrogen――AI Engineerからライブ

and we are live welcome everyone to thursday i for what day is it today there's been a lot of news I'm a little confused. November 12th. Yes. This is Alex Volkov of A-Avengers with CoreWeave and the host of Thursday Eye. Joining me in studio live from AI engineer, Ray Fernando. Ray, welcome to the show. Welcome, welcome. I'm really excited. A lot of exciting stuff going on here in the building, and I'm excited to be here with you. And, of course, while we're here, IRL and Ray's joining us, we have friends joining from our regular co-host. Please welcome Peter Gosteve, Wolf and Raven, Wolf. Jan Peleg is also here. Folks, I don't know if you had the time to check the news this morning If you did, this could have been the most consequential week in AI's history

And we've been doing the show for a long, long time And we do not say this lightly This could have been the most consequential week in AI's history Why? because this morning when I woke up, I got a text from Ray, actually, and I went to check my timelines, and Jensen Huang, the CEO of NVIDIA, is now the ex-CEO of NVIDIA, also resigned from the board. Wow. After a week of bombastic allegations and him admitting infidelity and an affair with a supposed Chinese spy, Alice Huang, and then another bombshell came about his another affair with somebody at his company everybody was trying to figure out what's going to happen with Jensen what's going on and as of this morning the legendary 40 year CEO of NVIDIA the guy who 60 direct reports are reporting

directly to him the guy who's moving open source directly forward the guy who's bringing metrics with differentiation to the world. He's no longer the CEO of NVIDIA. And as of now, Colette Kress has agreed to serve as the interim CEO until the permanent successor is named. And so they're starting to search for the next CEO and will consider internal and external candidates. This is a fast-moving situation. We have notes from last night. The notes say the board is thinking about Austin Jensen. But now we're here. We also have an incredible week Besides this, besides the fact that we're here live at the AI Engineer Code Summit in San Francisco, which we will talk to you about, we have five guests lined up. Really an incredible show lined up for you. Besides this, we also had Fable 5.2 this week. We had Muse Ember this week. Gemini Carbon came out at full availability. That's a big week. News Research came out with a new model. It's a great week, generally.

but I think this news takes the absolute just highlight. So with us as a first reaction, let's get Wolfram, you texted me immediately. And then Peter, I would love to hear from both of you. Immediate thoughts on this? It's a sad day, a really sad day. I mean, this feels like a psyop actually. That's my conviction basically. It is an alleged spy. There is no proof of this. I haven't seen anyone show the proof, but it was enough to start this. Then afterwards, they came up with another experimental affair of Jensen. So that is his personal thing between his wife and himself. And I don't want to speak about this, though that's one thing. But the other thing is he has been a stout proponent of open source AI, open rates AI. He has been against what Antropic tried to do by controlling the GPUs that they can't

run open-led models and so on. Though I think he was one of the biggest proponents of the open and free availability of AI, being able to run it on your own machines. And now with him gone, and now that the Democrats took over, there is also this increased push for AI regulation. So one of the proponents that could fight this is now out of the race, maybe, hopefully Hopefully not. But yeah, it's all strange timing that this comes up now. And if extramarital affairs alone were a reason not to be able to lead a company, I mean, there are people leading countries who have extramarital affairs. So I wouldn't go after this, but yeah, looks like a setup and a fire after the scary potter incident. Now we have this media just slashing out at him. I will say, Wolfram, we're on conference Wi-Fi, so if I talk over you, I apologize because we have some delays. But I will say, I don't think the affairs are the thing that moved Jensen out.

There's a whole thing about this being a Chinese operation. And the extramarital affairs could be the thing that the media talks about because it's salacious. And, oh, you know, she looks like that. She looks like whatever. that's not interesting to us but if you know if Justin got honey potted and because of this did some stuff that's something that definitely should be discussed Peter Gosteva I would love to hear from you as well yeah that thing it's hard I really don't know what's going to come out of my mouth right now because I have so many thoughts in my heart dude yesterday I was going around at the speakers dinner I was going around with people they're like oh what about the show tomorrow I have no idea how to cover this. Like, I have no idea what to say. And then the situation changed. And now he's not the CEO anymore. Just go with, Peter, people like watching. Just go with whatever you ask, man. What's your reaction? There's so many directions, right? There's so many angles. And I think the reality, we also need to realize,

we probably don't know everything. And I agree with Wolfram as well that there seems to be some kind of, feels like there's some bits of information that are missing. and for me the big question is what was going on what is the craziest super villain name of this company in Singapore I'm going to look it up yeah so I was listening to Bill Bishop today because really very good report on China and he knows China, he lived in China and DC and he was talking about a lot of rumblings and unhappiness with Nvidia and US government and not Trump necessarily but people behind in his whole apparatus. And there was, so there's a whole, that dimension, what's going on there. Then there's a whole, I don't know, just good old fashioned sex scandal, which you know,

I also don't particularly have any interest in. But people seem to have picked up on it. clearly if that's happened that's it he's gone uh which is that's a whole other dimension there's also dimension of just he's such a powerful man who's been coordinating so many things and it's this kind of weirdness of like well you go down because of this like how how done with that like why why did you do that so there's this competition as well so yeah many many directions Ray, we talked with many people yesterday about this here at San Francisco, like where AI happens. Right. Anything? I like to zoom out a little bit more, and I'm just trying to think about, you know, what is it that NVIDIA does, and why is it, like, everywhere, right? We're using everything, you know, and I think it also parallels to Apple, right? Like, what is Apple doing? And, you know, when Steve Jobs passed away, I always try to think about those moments. Like, did anything change internally in the company? And when I was working there, it's just like, you know, there's still that same North Star for everything.

And, you know, Apple doesn't manufacture, you know, like specifically, you know, they contract that stuff out and they market it really well and they sell a lot of devices. And NVIDIA is kind of in a similar boat in that way, if you look in parallels, right? Like everyone's making products, everyone's using their stuff. You know, Microsoft doesn't manufacture computers and, you know, all the chips are kind of coming together and they're doing these partnerships. And, you know, Jensen is like the guy who's signing all the deals and trying to make everything happen. And Jensen is the chief allocator of compute which is which is insane like he's the guy basically Obviously there's other people obviously folks or 56,000 people in video something like this somebody please fact check me There's a lot of people like and video is not done today just because Jensen is out, but it is a turmoil and And yeah young public super quick on you as well would love to hear Man, I fear for open source, man. Yes, let's talk about this effects on open source. Absolutely. Jensen was the outspoken first tweet on X.com is a manifesto about supporting open source,

how important that is, and in light of everything going on in governments all around the world. walls and we really lost an important figure today in this this debate at least at least just showing the other side of what everyone sees on the news uh put everything else aside it looks way too coordinated but let's just put it aside yeah we tried not to yeah we try not to go deep into conspiracy stuff here. But there is definitely multiple things once things are starting. Ray, you want to show this? Let me send you a link so you'll be able to also show stuff. Yeah, the first tweet, like you're saying, was July 24th from Jensen Huang when he first joined X. And he says, for my first post, I'm sharing a letter at NVIDIA signed on why open models matter. I will transform every industry,

power every company, and be built by every country. And open models strengthen safety, cybersecurity, accelerate innovation and diffusion, and enable sovereignty. The world needs both frontier closed models and frontier open models. So it just kind of speaks to their lifeblood of, you know, then like really just trying to make sure that the whole ecosystem thrives closed and open. It just didn't really matter. And I think that's still really important. And, you know, if you think about this person, they had lots of hands and everything, but there's a lot of really intelligent engineers, as we've all worked with and talked with and you've had them here on the show. They do incredible stuff and they're off on these different ventures. I think it's a sad day, but I also think that I'm still kind of bullish with a lot of people there. It's not like everything just stops. We'll talk more about Jensen, but let me just tell you where we are. Followers of the show know what AI Engineer is. For those of you who are tuning in who are new, AI Engineer is the premier conference for folks who work with AI.

The category started by SWIX three and a half years ago, pretty much when Thursday I started as well. AI Engineer is a conference that brings together all of the top folks who are building with AI. We don't have them back here. We just have the AI E-code Summit. But we came live to you from every AI engineer like previously. Peter and I talked to you guys from London. Wolfram and I, we've been to San Francisco And Europe, London as well Everywhere Yam and I were from San Francisco At some point, Yam were running around When Logan Kilpatrick joined us So we are always covering this AI engineer So if you are a fan of the show for a while You know what's going on Many of you are at the conference right now Not listening to us But you will listen to us on the pod And the very interesting thing That I like to do in AI engineer Is give you the zeitgeist So, I don't know if you had a chance to, but yesterday we had a live stream. Me and Ray were running around like crazy folks in the hallway track of AI Engineer.

This was really fun. Yes. It was really fun. So, if you didn't have a chance to, you'll see a bunch of clips, hopefully. But besides this, we chatted with folks like Lauren Ten, who's a potato on Twitter. And you guys know Potato from Pistec and from the Grogba team. We chatted with Lydia Halley from Anthropic and Sunil Pai, who is now also Anthropic. Sunil, a friend from the pod, is also Anthropic. So we did already have a bunch of interviews. And what we're trying to do with the engineer is to bring you the Zeitgeist, to bring you to San Francisco, bring you the Zeitgeist. And for that reason, we're going to have one, two, three, four. We're going to have five guests at the second part of the show. Right? So we're going to cover the news. And then we're going to have five guests. We're going to have Kent C. Dodds, who is a lecturer, a long-time developer relations person. He's working on his own Cody thing. Many of you know him. Kent is a great dude. Then Michelle Chan is going to join us. Michelle is a group program manager at Cloudflare. She had a talk that specifically focused on Cloudflare being a Neolab, which is great.

And they started training models. They started serving models at their scale. I would love to chat with Michelle. Then we're going to have Ivan Leo from Gemini, Google DeepMind. That's going to be cool. And he's the MTS over there. We had multiple folks from that team on the show already. At times, you guys know. And Ivan is going to talk about something that's going to drop. But, yeah, we're going to have Ivan from the Google DeepMind team talk to us about their drops that already dropped. And there was two that already dropped. And then there's more to come. We're sitting across the hall From the AIE Code Summit And the hall There's going to have keynotes in a second Zach Lloyd The founder of Warp And the CEO of Warp Used to be Terminal And he's going to talk to us About what he's going to talk On stage about He's literally going to come off the stage Walk over and sit down

Like the schedule is this tight So I think it's very incredible. You guys are going to get basically what the folks here are getting, but kind of the behind-the-scenes thing. And then to close it all down, we're going to have Swix, the man himself, the guy who created the engineer, everyone's friend, the person who curates and decides who goes on stage and who is the sponsor. Swix is going to join us at the end of the show at 10.45 a.m. Pacific. So a very full show. And in order to get it started, let's jump into the TLDR. Let's just run through everything that happened in the eye this week And then let's talk about Jensen some more because folks this is like I think that this is what people want to know Like everybody wants to hear about Jensen Let's go to the deal there Alrighty, this is the deal. Yeah, this is all of the major news So, TLDR is basically all the major news items that we have that happened this week.

A lot has happened. A lot of it is, how should I say, undercut by the Jensen News, overshadowed by the Jensen News. There's some word there that says, hey, none of it really matters compared to the biggest news of today. So, let's just start with the folks who are here. With you, Alex Volkov, Evangelist at CoreWeave. And sitting right by inside me, Ray Fernando from RayFernando1337. Check out his live streams as well. And online joining us today, Nistan Tahira, Jan Peleg, Wolfram Ravenwolf. Nistan, just joined. And Peter Gostov from Arena, Wolfram Ravenwolf from CoreWeave, Jan Peleg, and Nistan from... At Prism. At Prism, thank you. Okay, and at the top of the news, we already talked to you about Jensen Gates. Jensen Huang is no longer CEO of NVIDIA. He was ousted as of this morning from the CEO position and from the board

after allegations and then admittance of extramarital affairs. Two of them showed up. One of them is to basically a potential Chinese spy, Alice Huang, with a web of Chinese companies that supposedly got GPUs from NVIDIA. So we don't know anything. All of this is reported. And then there's the second affair, and now there's a leadership transition, and now the CEO is... Colette Kress. Colette Kress, who serves as the CFO of NVIDIA, is now the interim CEO of NVIDIA. This is a huge, huge scandal. This is a huge effect on the markets as well. NVIDIA down, all of the new labs that we also work at, the market seems to be reacting very, very badly to this news, which is obvious. We are concerned about the open source effects of this. Meanwhile, somebody is having a grand old time.

Anthropic has released Fable 5.2. Right. Antropic has started their IPO roadshow, and Antropic is going to go IPO, I think, Wednesday next week. Wow. Antropic IPO, after dropping Fable Lifeline 2, which we will talk about here. Peter, I'm sure that you've already tested Fable. I've seen you talk about Fable. We'd love to hear from you and from you guys also about Fable. So, Anthropic launches the IPO Roadshow, 825 million shares of around $70 to $78 at 1.3, I think, valuation is the last one. Somebody please confirm if this didn't change after the NVIDIA news, but I think it was like 1.2, 1.3 trillion valuation. The IPO Roadshow started. Also, Cloud got work, whiteboards. Those two things are not nearly complete, but this is what was written. then we have Google Gemini being so back it is incredible literally everybody who I talk to here

with are getting excited about Gemini Carbon Gemini 4 Carbon is now generally available everywhere you can use it in your subscriptions we chatted with Phil Schmidt yesterday we're going to chat with Ivan Leo about Gemini 4 Carbon and it's stuff But also, on my timelines, the vibes are good. Very, very good. And that's not all the Gemini launched this week. They also launched the Decision API. That was cool. Gemini Decision API basically follows the trend of Type 1 models started by TypeSafe, Jev, the CEO of which was on stage here at the AI engineer yesterday talking about this. Gemini Decision API, besides being very cheap and fast running on top of their models, is also multi-model, including images and video and audio, and we're going to talk about this very soon with Ivan, so stick around. On Meta front, Meta also had some news. Meta also released another model,

although kind of got lost in the fray of releases. Meta released Muse Ember. It's the smallest Muse model, so Meta is basically haiku. And a pretty cool model as well, showing up very well on benchmarks. In open source news, anything else in the big companies and APIs and LLMs that folks need to know that I missed folks? It's been hectic quick. Anything else besides those models? Peter, anything else huge that came out that maybe we missed? No, I think, yeah, sounds about right. I feel like there's some image model we're missing. We're going to get to image models, yeah, there is. Minimax H3 came out, yes. But, okay, let's check this. Minimax H3 follows the Minimax M3, which is their video model, Minimax H3 image. Yes. H3 is a video model, and it's H3 image now that uses the video.

Oh, H3 image. Yeah, you're right, you're right. Thank you. Let me try again for the recording. Folks, Minimax releases H3 image, which is the image version of H3, their video model. Thank you Wolfram 4K images and open weights Shout out to open weights We don't have the applause button here But yay We need to talk about license But license But it looks great I think it's the state of the art Open weight image model right now And open source LLMs Our friends from News Research Who moved towards building personal agents Are now back Still back With News Research releasing are at 27b it's a quen post trained towards western values and i will go on the limb here and say that this is going to be more and more and more important that we're going to be able to westernize uh eastern models i think it's a cool uh cool approach it's apache to license on top of quen and shout out to news research for this effect any any big releases in open source this

week that we should mention? And this is going to look at you. What's on top of Hug and Face? Interesting. Wolfram, looking at you guys. I think everyone was at home playing Zelda, so there were a lot of... GTA 6 coming up. Oh, I forgot, did you mention Ernie 5.5 Flash? Ernie? Dude, honestly, we don't talk about Ernie. It's been like that for two or three years, Ernie always comes out No one's trying it No one's using Ernie It's very interesting But yeah, let's just mention Ernie It also comes out HY4 I think HY4 Pull it up Looks like it's 286 billion parameters It is at a practical Size and 14B active

trained from scratch on 33 trillion tokens. Looks like they put quite a bit of stuff into it. Yeah, I don't know. I think we'll wait for people to actually run their own benchmarks. After the new one. So I can talk about that later. Okay, cool. So in 2.5, LongCat 2.5, weights are finally available now, which is May 2.1's native multimodulation model with 1.6 trillion parameters and 48b active in tools and agents folks I don't know if you knew this I didn't know this until yesterday that we talked with Matt Palmer and and Lauren Tan aka potato or actually potato aka Lauren she goes by potato a Grogbot was beta all this time you guys know I didn't know Grogbot apparently was under beta and yesterday Grogbot released 1.0 which launched where Grogbot can join your team in Slack, and it can join your Zoom and Teams meetings, and Grogbot can sit there and transcribe,

and we chatted with the team here about Grogbot. I know it's beta, so I think it's big that this new category of thing is now 1.0 with a bunch of features. We tried to get them on the pod, but I think some more time is needed to get them on the pod to get approval from the SpaceX AI folks. All right. In vision and video. Today, in addition to the five human guests that are joining the show, one AI guest will join the show. Emma from Nuance Labs is – folks, I don't know if you've seen this. A lot has happened on the timelines, but this is by far – and shout-out to our friend of the pod, Quindle Kramer, and congratulations to daily folks about their acquisition as well. So, Quindle and I, we talk about, like, we brought Quindle here all the time to basically tell us about the state-of-the-art live agentic AI. This release from Nuance Labs broke my mind.

You guys will not believe how crazy this is. Nuance Labs released this insane, they call it Nuplex, Full Duplex Audiovisual AI Persona. And we're going to show this. We're not sponsored by them. My mind was blown away. And I've been doing this for three and a half years And my mind continuously gets blown away Which is incredible Because we're not yet domestically adapted to progress Which I think is very, very cool And from the forge, powered by CoreWeave I will just mention CoreWeave had Finally, we can talk about this Wolfram We had our earnings report out And the CEO and the CFO and all the folks Told you about what actually happens within CoreWeave As a reminder, CoreWeave is sponsoring this podcast Both me and Wolfram, we work there We are not persona, whatever, that has related information All we are saying to you is public And on that public thing, CoreWeave had an incredible quarter Despite whatever happened to the market, despite the fear, etc. CoreWeave had an incredible quarter We're going to tell you a bit about this Besides the quarter stuff

CoreWeave also powers the Wavy AI driver training Right? So, I don't know how many non-Europeans know Wavy but Wavey is like the leading autonomous car company in Europe, I believe. And now Wavey has signed a multi-year deal with CoreWeave to train self-driving on top of CoreWeave. So not only are, you know, we're providing GPUs, now self-driving cars in Europe, hopefully, thanks to CoreWeave. And then we're going to have guests. I think that's pretty much all of the main, main news from AI, but I will open the floor to folks who are tuning in in comments to the co-hosts. What is... Should we start with Jensen, or is there anything else that I missed that's very, very important from this week? It's a big topic on everybody's minds, so go ahead. What is there for us to say that hasn't been said? But let's do this as a full segment, folks.

NVIDIA CEO, Jensen Huang. He's no longer NVIDIA CEO, Jensen. He's just ex-CEO Jensen Huang is out As Nvidia's CEO After a terminus week For him and for his family I honestly don't care about The affair allegations Like that's between him And like Wolfman I think you said Between him, his wife, his mistress number one His mistress number two and whatever The memes are just Just crazy Okay I kind of don't want to participate in that Because obviously a lot is happening to that family So I will not then I will just say This is a person that is very important to the ecosystem of AI You know After 40 years at the company The Helmholtz company With 60 direct reports And extremely technical person Extremely technical If you ever tuned into a GTC keynote Jensen goes down

And he barely rehearses this. He just, like, talks about the stuff, about the different deployments and GPUs, etc. And now we are entering into a world post-Jensen, which no one of us knows how this world looks like because Jensen has been there for 40 years. He's like a staple of the community. And then when AI came up, and then we noticed that, oh, we actually can train these models on top of these gaming rigs, Jensen very quickly rose and the company rose to being the top company in the world, right? So, I don't know, folks who are watching, is it going to be interesting for us to go blow by blow about the first affair and the spy allegation? I think the spy allegation, let's talk about this. This is the most important thing that I think is what's happening here. The first allegation that came out, apparently for a while now, many, many folks in government and some research firms by the way on next i don't know if you guys follow some folks have been talking about this since may uh there's been rumors that jensen huang has been either honey potted or in an affair

with a rumored chinese uh intelligence asset or whatever people immediately jump to the word spy uh alice huang who shares the same last name as jensen but is not related and uh apparently rose to fame very quick and then has a web of companies there in Singapore and China and other places. Peter, I see you have something to add here. What is your take on the situation? We'd love to hear from you. I'm talking more as a casual observer of the stuff trying to read up on it. so you know don't that's all that's all our viewers want from us yes i don't want to project any authority but just like reading around about it i think there are a few things one is that it doesn't look like the it doesn't look like she was a spy as such so it's not like she was trained

in whatever mi6 equivalent in china and like i it looks like she was an influential person and then And maybe she had some, as a result of which she had some links with the Chinese authorities. That seems to be more likely. And then they probably connected with her, which I must say is probably an assumption you should make anyway. Like if you're dealing with an influential person from a government, they probably have some connections. So I think that that feels like maybe an unwise thing to have done in the first place. And I was a little bit surprised by the reaction of that. There were a lot of people who were like, well, maybe it's not too bad. And then for the second affair, it was like, oh, he's gone. But I thought the first one was so much worse. It's so much worse. Yeah, and there's probably a lot more going on that we don't know about. So it feels kind of odd for it to be not enough for him to resign for the first time,

but enough for the second. I just don't quite see the consistency but I was talking to my wife about it this morning and she was saying that I'm an idiot and I don't understand anything so she's probably right I think everyone has a different take on this and I think all we can offer here, I honestly folks I have no idea how to cover this, I have no idea there's the whole thing where you know we're sponsored by CallWeave also and CallWeave gets GPUs from NVIDIA and before jensen was completely out it was like should we even talk about this because then you know um but then also we don't know no one knows like i don't gonna sit here and pretend because we're in a nice studio with a nice microphone that we have any idea what's going on like the news media does we have no idea what we can talk about is our fears and i think like for a lot of engineers here the the fear of what's going to happen is drifting down specifically from the open source angle because Jensen has been a huge proponent of open source AI and dissemination and he has been in President Trump ear talking about open source and its importance and with the hug and face acquisition Jensen has solidified like

hey open source is important as also the open source alliance that Ray you just talked about right the first week that he did yeah so let's talk about what we think is going to happen to open source and whether or not this is now in the company culture or are we expecting a closing down or you know taking over by whoever is going to IPO very soon or some such companies. Wolfram, what's your take? So I think he's not being pro-open source out of the good of his heart. He's a businessman and he's selling GPUs after all which are required to run models and he was very successful with this. So it's not like he's a superhero in that regard, but he was or is a very influential person. Like you said, he's talking to the president and getting caught by the president on some interviews. So his influence is not to be underestimated and he was a persona with the leather jacket and he was a leader in that way.

But a person as powerful as him, he was a creator of the company as well, the founder, to take this away. Within days without a trial or anything, there has nothing been proven, just allegations, and he has been hit that hard. I mean, that is a huge blow to the open source community and to freedom in general, I think. Big Danison, I would love to hear your thoughts as well. Pro-community, you're saying the community is pro-Jensen. What are you seeing? Yeah. When you see accounts like on Twitter, like, for example, Autism Capital or others that post the exact same meme 150 times in a row to the point where sometimes it's not even funny anymore, although the early ones were pretty funny, it looks like this is a pretty – it does feel like quite a coordinated campaign. and you don't post that many means unless you're specifically paid to do it.

I think NVIDIA's investors should ask for Jensen back. I don't think in today's world when there's kids getting bombed, anyone seriously cares about a consensual extramarital affair. That's for them to decide and divorce and figure it out in their own family. Nobody actually cares about that. It feels like there's something very shady going on. And, again, we've gone through this with the Sam Altman shakeup, and there's a lot to speculate here, but it really doesn't look like he did anything terrible. I mean, for his wife he did, but then she would get a very big divorce out of that. I don't know if you guys told me to speculate. They probably just coerced him on the table and they said, hi, we're either going to take you out for having an affair or we're going to take you out for selling seven and a half billion dollars of GPUs to China.

So he just picked the affair. That's kind of what it looks like. And honestly, I'm going to go out of my way here. But if they're going to have a war with Taiwan and stuff, probably the old heads and military heads are not going to want an Asian guy leading the most important company. that's also a big speculation from the community and yeah people are not happy they want Jensen back and as both as like the community side for supporting Hockingbase and open source but also the money side as an investor are you going to make less money or more money with him gone you're probably going to make less NVIDIA stock is down 27% this month so yeah this This looks really shady. This is not good. There's some really shady thing going on here. A lot of stuff is happening lately in the world of AI and markets. And I will just say, two weeks ago, I brought to you guys, we haven't covered this in depth,

the Commanding Heights essay by Wilma Nivis, the financier. The Commanding Heights, if you didn't have a chance to read this, talks about sovereign takeover of AI. It talks about AI being so big that the two countries basically that are going to own this, which is the United States and China, are the only ones. As much as folks try to pretend there's a third one, there isn't. And it's so big now, and it's such a big part of definitely the stock market, as we're seeing right now, based on every new stock market jumps 10%, 15%, etc. But also it's such a big part of the economy that there's no way to avoid this. AI is going to get taken over by government at some capacity. The sovereign is going to take over because we're talking about, you know, technology that's more powerful than, you know, nuclear, et cetera, because of all the hacks, et cetera. And some folks are speculating that this has started with this announcement. And they went directly to the source, both of them,

the Chinese via the Chinese spy asset, et cetera, not clear honeypotting or what, and the Americans with the reveal and the affairs, and, you know, in the U.S., how this goes with the media. So, Ray, thoughts? I'm just trying to be seriously as optimistic as possible. I think it's very confusing, and out of all the different things I could focus on, it's been really rewarding. It's kind of the reason why we're also here at this conference is that, you know, this company, amongst others, amongst all of us as engineers coming together, are really trying to figure out what this frontier looks like. And I think that, you know, while there's currents and stuff that are currently going, and it's a little crazy to, you know, do this type of stuff, I think the knowledge that we've already learned, it's like the toothpaste has come out. We've already understood that, you know, using model weights in this way to make advanced applications and solve advanced problems for humanity is still like the North Star literally for everyone here that we talk to, right, from, you know, genomic research, cancer, and all this other stuff.

open or closed source all around the world. And I don't know, I just kind of want to maybe instill like a more positive message. I hope that those things will get worked out. And I'm optimistic that all the smart engineers and people who are still working there are still going to keep pushing to drive that, no matter the organizational changes, because I've been through organizational changes a lot. And at the end of the day, you know, the smart people who are at the ground, like just like he was, he's just more hands on, just like these other big CEO people, you know, they may disappear or may come back. But, you know, they're still driving all the... You know, Sam Altman can't have all the attention and focus to focus on every single problem. You have to have really intelligent people and bring that culture together. And I'm hoping that they can rally or something. Let me join this positivity just for a second, okay? Because we are here, like, the fight to move on positivity is the thing we do. We do need to talk about a bunch of other stuff. I'll just land on this, okay? The company is bigger than one man. and you need a lot of people building the most ever in the history of the world like requested resource that's going and growing and growing in popularity people the vcs the thing they know

stuff they go and say hey you know the company cannot survive with the founder they're citing jobs they're citing travis kalenic from uber well let me tell you what apple is still very very important and apple is everywhere we're sitting here with a bunch of apple devices despite steve jobs being gone for a while uh travis kalanik i don't know where he's at but i'm taking ubers everywhere ubers are still around uh and nvidia sells the most requested thing in hardware and it's going to grow in popularity as we've been telling you for a while like like influence is going to keep happening and whatever turmoil nvidia is going to go through i think they will keep selling gpus surprise surprise i think the demand is going to grow surprise surprise so whatever terminal they're going to go through, there will still be AI happening. Open source, we'll see. But like Wolfram, you said, he wasn't just like pro open source from the goodness of his heart. He is selling the hardware that runs it and requires it. And so the machine will continue. The company is more than one man. Wolfram, I think we keep talking about the company is a super intelligence on its own.

It's not one person. There's like a whole web of relationships now happening there. Besides contracts, there's a bunch of contracts as well. So they will be fine. despite what the VC class of folks will tell you online and send you doom and gloom for likes and clicks. Folks, we have to move on. Because Fable 5.2... Yes, one last one. I want to know what you guys think. What do you think he's going to do next? I will just... Did you see that? Somebody said, you must hire Jensen. He said it was an honor. I doubt it because I don't think those two personalities will work, but... Maybe. Saruji apparently may be poached from Apple. I don't know. There's these Mac rumors that are pumping up. They're bringing him back, guys. The stock would go up if they brought him back. They're going to bring him back. You think he's going to be back like Sam Altman? I think he's going to be back. Like Jesus style? Like we're direct after three days and come back? We have a bunch of predictions. Not financial advice. Not financial advice.

Don't underestimate his main strength, at least more recently. I think it's right that the company is more than one person and they probably didn't need his engineering expertise now to build better chips, but he's an amazing deal maker and just a great personality. So you can imagine him doing like a massive fund to fund the whole AI thing. I hope she does something like that and just start afresh. And I think he's big enough to be able to do that. That would be awesome. All right. Good luck, Jensen. Whatever you're going to do If you're going to come back Please come back If you're not If you're going to do something else Good luck Jensen Huang folks We have to move on Because the next hottest topic That we have taken over the airwaves Without the Jensen news Is Entropic IPO And Entropic releasing Their latest Fable 5.2 model Entropic is actually sponsoring The engineer here Lydia Halley Grabbed the stage And started talking about this What do we think about Fable?

folks what i think about fable let's talk about fable 5.2 uh 10 cheaper i believe than than the previous fable and mythos we don't get access to mythos but like when they release they both say both models are releasing it's basically the same model with different constraints etc uh peter as always i know that you've played around you've posted some stuff would you like to show us or tell us a little bit about this model is it the best in the world or is opus 5.5 good enough that we don't care anymore like i would love to to hear from you about this Yeah, look, it's 100% better. And I'll show you our leaderboard. And it's a question of, always a question, right? Is it worth it? And I think, yeah, you might want to debate that. But this is our agent leaderboard. And the way we derive it is that we let people use it for free. And then we don't tell them what the model is. And then we can pick up from the signals from real-world tasks, whether they like it or not. or some objective signals and look at this difference.

We study whether... Wait, hold on. Let me zoom in on this one second, please. So, for example, here, praise and complaint. This is one signal that we pick up and look, it's nearly double versus any other model below it. So people really... Can you tell me a bit about this praise versus complaint? Is that what you pick up on, what people react to and say when they prompted it? Yeah, so, you know, when you use agents generally, You would provide them verbal feedback or correct them, so that's steerability, for example, or basically interact with it. And we can analyze those logs and extract that kind of data from it and see how people use it. So we can collate all of these bits of feedback, and we construct this leaderboard from that. So, and it's not, we don't, there's an explicit, did you like that? But the rest are more signals around Just generally how you use it And yeah, for example It's way better on our metrics than Opus

It is more expensive So you need to bear this in mind And it uses the same number of tokens So it is pricier I would also say I have been using it For the release And I did find that it is quite both models obviously are competent, but once you start looking at the details, you can see that Fable has much broader diversity. So if you want to, you test it on different, I don't know, more creative prompts, you can see how much more you can do. Then for example, first precision and world understanding and just picking up the details of or extrapolating what else you might have wanted to do. That Fable is noticeably strong at that. Whether you care about this I think is another question but certainly yeah so certainly in terms of just that kind of quality of the model

it is higher but obviously opposite is also amazing so yeah whether it's really worth it is another question but it's also from my personal experience, I don't feel like it's enormous shift and what's interesting is Rumov and they're clearly named Opus 5.5, but why is Fable 5.2? It's not 5.5. So that's a little odd. So maybe they were hoping for a bit of a bigger jump as well, possible. Could we order leaving themselves space before IPO to get to that one more checkpoint? as we know they have the checkpoints like they always keep training the models they have the checkpoints they're always evaluating them and maybe this is a you know uh even trying to curtail the release of gemini uh and carbon and kind of like how how much this is getting uh affection uh folks anybody else try fable wants to comment on this i did it go ahead it's absolutely incredible

um from a couple different perspectives i used fable 5.2 inside of a thing called bitrig And so these guys are the co-creators of Swift. So they left there. I'm not sponsored by them. I used to work with them at Apple. But it's really fascinating when you give an extremely intelligent model this type of tooling to go in. And so it has all the best practices for Swift and laying out code. So I just said, I gave it this big challenge. I wanted to make this multi-platform app. And it figured out all the attention to detail and just kept going and going and going and going. It even wrote its own metal shaders and all the little attention to detail that you would need for, like, animations, taptic feedback, and it just really kept diving deeper. And so my thinking here with this model, and I have not seen this with other models in that same type of test, is that if you start to create the agentic environments that will allow models like this to thrive, I think people have to rethink and remove all their previous assumptions for skills and approaches to things and try them in these new ways.

And so I never described any technical details. I described more intent in my prompts and the level of detail that it was able to just completely continue the verifications and also check back on its work over and over again to get the stuff done. It was extremely impressive. So that's Fable 5.2 on extra high because I just wanted to test that out. I do find Opus 5.5 Medium, like you say, extremely fast, really good for churning out lots of great work. But there are these little holes that I just don't know exactly where it just doesn't really finish the rest of the work. But for the most part, if you're maybe a person who's somewhat technical, maybe you're a doctor, you're not an engineer or something, but if you can describe what you want in your language, I think Fable 5.2 with the right tooling can help build the tools that you want for the future. that is just outstanding. I'm ready to push ship on the App Store right now. That was great. So we actually had a chance to catch up with the Anthropic team yesterday during the live stream. I don't know if we have a chance to show you this, but basically the new meta is Fable 5.2 is an incredible orchestrator

for long content stuff. You can ask it for your most challenging tasks, and it will grow. And basically what they're suggesting now is to use whatever projects feature you have in cloud code or other places. Projects feature will be run by Fable and then the actual threads will be executed by Opus for that, you know, you don't want to burn all your tokens. It's very easy to get through your quota with Fable 5.2. It's $8 on 1 million input tokens and $40 per 1 million output tokens. It's down a little bit, down 30% from $10 and $50. So $8 and $40. It's number one on Agent Arena, right, Peter, with like 30% like better. It's number one on artificial analysis. This is like, it feels like the top model in the world right now. So shout out to folks at Anthropic. And again, the new meta is try it on long horizon tasks as the orchestrator and as the executor, you try Opus 5.5 if you're in that ecosystem.

Google will try to tell you that, you know, there's an alternative for you there. All right. immediately after Fable, Entropic launched their IPO Roadshow, which is where the company goes into places where money is and asks for money, or basically says, hey, you know, in Entropic's case, not ask for money, it's more, we're going to offer you a chance to buy a piece of us if you want it. And apparently it's like super, super oversubscribed. This is something that banks do all the time in IPOs. They oversubscribe so that, like, there's a lot of demand. Antropic launched an IPO roadshow 825 million class A shares Around price at like 70 to 78 A share This is not a financial show Nothing of this is financial advice Do whatever you want to do The ticket is going to be And we talked about this They want to raise 64 billion dollars Alright Let's move on I think one more topic Before we move on From the forge

I do want to talk about the news new model have you seen this, the news Aret thing shout out to news research for releasing this I do want to chat about this with Wolfram and you Areti Areti Areti 27V I haven't seen, okay their work is already going into other government models and stuff behind the scenes Let me announce this, and then we'll discuss, folks. Our friends in news research released Arete 27B. They took QEN 3.8 27B, which was just released a month ago or something. They post-trained it and supervised SFT, supervised fine-tuning, toward what they call Western values. On their 803 prompt eval, the propaganda rate went from 22% down to 0.9%. So they made some brain surgery on the Eastern models, and now the capability didn't move.

On Hermes Bench, it's 65% and 65% for the base model, so like barely, barely moved capability. Apache 2, and there is the evals, there is the data. And this was the most reply to open source post of this week, a huge debate in the Japanese and Chinese communities about whose values a model should have. and I will go on a limb and say Western here you know this is the values that Merck thank you I'm fully okay with this, shout out to New Swizzards for this post many people have started giving them flack on the timeline but many people are also like hey it's going to come down to a very serious debate between the two cultures very very soon there's only one that allows free speech where do we want to land basically is the way that I see this so shout out to news research for this effort I think this effort is going to be more and more important as more and more open source models

are coming out from the folks at China and we don't know what's in these models, what kind of trigger words Ray, thoughts on this? I think it's a pretty smart thing to do to just start to establish some of these baselines because it's open source too. Is it also open weights so that we can take the weights and retrain them and do whatever you want? I think that's the whole point of open source, right? Yeah. Wolfram, love to hear from you all the way from Germany, friends of news. What do you think about news? I mean, that is a big thing about open weights and that you can retrain it and make a new model out of it with your alignment, with your values and your capabilities. and being able to do that is the big thing about freedom. Otherwise, you only get an API and you can access a model. And they may be more intelligent, but the thing is you have no idea what they are trained on, even what secret prompts they may be injecting and stuff. So, yeah, it's super important for freedom.

And the ultimate freedom is to train your AI the way you want it. And that's what is possible that way. And that news is doing this with Western values is a great thing. Yeah. All righty. Time to move on, I think. Based on the schedule that I have, there's an AI guest that I want to bring you, but I think we're running a bit low. Let's do From the Forge. I just want to tell you about Wavy. Let's do From the Forge super quick, and then we're going to have an AI guest here with us. From the Forge. And how about it? All right. As you guys know, Thursday Eye is proudly sponsored by CoreWeave, where I am employed and Wolfram is employed also. And it's been a hell of a week for CoreWeave. Hell of a month as well. The Jensen News is obviously related to us as well,

but we're not going to care about this. But there's stuff here that's not relevant. So I'm just going to say, Okay, there is, we did the earnings call, and you guys can go and do your own research and read about this. Corey had a banger quarter. Everything I'm about to say is from the public release. There's no previous information here. Corey reported Q3 on Tuesday. Revenue was up $3.6 billion, up 168% year over year. Adjusted operating income was $379 million. against guidance, like, significantly more than guidance. I'm not going to read you the whole thing. Go and see if that's what you're in. But, like, Corey locked in and delivered. And there's a lot of stuff about that. Corey talked about whether or not, like, this is a thing anymore. And you guys do your own research there. What we should say, though, is the stock closed up.

And, obviously, after NVIDIA, everything is changing. But it had the best day since, I think, June of last year. And the whole AI infrastructure was pulled together with it. The debt question. Mike and Trader, the CEO, addressed this because everybody was thinking about, like, whether or not CPUs can be bought by debt. Mike and Trader basically said on Bloomberg. Corvive won't use the unsecured debt market at current yields. And the next day, Blunga reported a loan of at least, like, for stuff backed by GPUs and customer contracts. customers will be preparing for those GPUs ahead of time and about half of the yield of existing ones the guidance raised and we have the last one is basically what we keep telling you here on Thursday all the time we have 10x as much qualified demand as available capacity through the end of 27 I will say this again slowly you guys are rich core with we'll see the stock market is a crazy thing I have no idea Revenue per employee is like 3.8 million.

And to put that in perspective, NVIDIA is about 5.4 million to 7.21 million. He's not an employer, by the way. He's cleaning off his stuff. I'm a propaganda guy. Anthropic is 6.6 million revenue per employee. And you're kind of measuring that. So where does Apple stand? That's 2.87 million to 3 million revenue per employee. So, I mean, whatever you're all doing over there, you guys are cooking us some magic. Fox are executing, putting up the best infrastructure in the world. We have 10x as much qualified demand as available capacity to the end of 27. We keep telling you this on Thursday. There's no demand shortage for AI. The folks have not yet realized what's going on when the norm is going to get proactive agents. They didn't. None of this is financial advice. Please don't take anything from us. We don't know anything. It's all public information that we're reading to you. But still, this is now the CEO of the company that sells you to use to other companies that sell the models is saying, 10x more as much qualified demand. Wolfram, you want to jump in here super quick? I just wanted to add, and this is my personal opinion,

like the whole podcast, we are talking from our own opinions, we are sponsored, but we don't get told what to say. And so, COVID is more than just GPUs, it's more than just the data centers, the agent loop and the full forge and everything. But I think AI can clone software very easily and can do so much, but it can't just create GPUs out of thin air. And the data centers are where the real mode is. That is my opinion, my personal opinion. Because, like you said, Alex, when the world wakes up and realizes how important AI is, and everybody wants to use it and for everything, it is like electricity. And then everybody needs access to this, and you can't just create it out of thin air. That's about the data centers as well. They are critical for the economical growth of any country in the future. I want to say also, in addition to, you know, Core is more than just a GPU company. We're, like, we're basically a hyperscaler at this point, as far as I am concerned.

What is important there is there's a lot of services that we provide for these companies that run GPUs, and also to you, recently. The reason why this segment is called from the Forge is because Forge is our, like, SaaS layer basically to allow you guys to do whatever you need with AI including inference going sandboxes including like a bunch of stuff a lot of it went like on GA this week and huge companies they're not only using GPUs but using the whole stack is coming out with with great announcements and one of the main ones that I've seen this week is Wavy it's a self-driving company out from the UK they signed a multi-year agreement to train its AI self-driving foundation models on CoreWeave. Very soon, across the world, starting with Europe, I think Wave is the biggest self-driving European company. They basically will start training self-driving on Europe. And also, there's a number that makes me very happy. Wolfram, I'll just mention this and we'll move on.

70% of the new cloud customers CoreWeave added were already Forge customers. Forge is the outcome of what Weights and Biases and OpenPipe and a bunch of other acquisition the core we did very smartly to build up a cloud that's not only about gpus and a lot of the stuff is coming to fruition here so very happy about that shout out to the folks at core for a banger quarter a lot of really really good work went into this and now let's move on to another guest on the show amazing um let me choose if you can't see dots who i don't think needs any introduction from folks who've been following the web, I don't know, web developer roles, etc. Kent, I've been following you for such a long time, and it's been incredible to kind of like see our paths kind of like start crossing more and more in the AI trade. How about you tell folks who you are, what's your history, what's your background? Yeah, yeah. So for well over a decade, I've been teaching people how to build excellent

user experiences on the web, starting with testing, testingjavascript.com, going into React and epicreact.dev and epicweb.dev, epicai.pro, because the .dev was taken, and epicproduct.engineer is my most recent thing. But at the beginning of the year, everything changed. Well, the end of last year, about a year ago. And so all of that stuff is less relevant now. And so I put a bunch of irons in the fire, and so I'm doing YouTube, and I'm doing still educating with Mega.dev, and I'm working on a product, too. All right. Mega.dev is something that you and Joel Hooks and Angie Jones and T.O. and our friend John Lindquist. Did I get all of you guys right? Awesome. Mega.dev is like this super cool thing where you just basically teach folks how to dev in this new era of AI. I want to hear from you first of all two things one what did you talk about because you're one of

the folks who are like moving the Zygos forward and two what are you hearing here that's new that you didn't hear like before like where where are we yeah yeah sure so I am working on a product and developing that product whenever I teach somebody something I have to like the way that I do that is I build something real and then I teach people how I built that so that's kind of where Cody came from at the start was just like, let me build something ambitious and real. And, um, and then I'll teach people how I did that. And so that's what I did here was, uh, teach people how I'm building Cody. And, and the interesting thing, this is getting to your Zeitgeist thing is, um, the way that we build stuff now is by building a machine that builds a machine or a factory. So software factory is like what I'm hearing constantly. Uh, and so I just talked about my flavor of that. Um, when I, when they first asked me to come to speak, this is kind of what I was going to talk about, but it was not really called a software factory back then. Yes. So, like, this term has just really taken off. Kent, what is a software factory?

I don't love the term, but, like, when you say it, like, I think most people come to a pretty clear understanding of what that means. Yeah. Software factory is not a new concept. Like, 10 years ago, we had software factories. It was just that the workers on the factory floor were humans. Yeah. And now we have software factories where we are the managers of the software factory and the workers on the floor are the agents. And it's our jobs to walk the factory floor and grease the wheels and try to help the agents work more efficiently. And what's interesting is if we're walking the factory floor, how do we figure out how to get agents to walk the factory floor so we don't even have to go down to that level? So we're constantly working on raising our level of abstraction so we can take more leverage on the agents. And so we're constantly working on raising our own level of abstraction so that the work continues without us. I think this is like a great definition of a software factory. And how are we going about this How are you doing this Yeah The practical application of Software Factory I think for me there are three pillars for delivering a great product

And that's the ultimate goal. Software Factory doesn't exist for itself. Yeah. We're trying to deliver a great product for a user and deliver good value. And so there are three pillars. The first pillar is making sure that the work that is done is quality. So, like, the PR that's opened or whatever, the commit has quality code in it. And then the second is if a mistake manages to make it into that commit, how do you catch that mistake? So through verification and review. And then third, if a mistake manages to get through that review process or that verification process, how do you limit the blast radius? And again, none of this is new. This was always the concern of technical leadership at companies since forever ago. So it's just that the people who are actually doing the work are agents. And so there's a lot more power and opportunity there, and it's very exciting and cool. But those are the three pillars. Make good stuff in the first place, catch any mistakes, and limit blast radius. Limit blast radius. I think that's wonderful.

I think also... Oh, I'm sorry. Yeah, please go ahead. Those three pillars. Then you start plugging signal into that system. and one of the things that I have found super, super valuable is a friction log. And so the agents will complain. You've got to get the agents to complain because they can very easily just work around all the problems and they're happy to do that. But you tell them, like, what stunk about working in this code base? They log that and then you've got an agent to go grease the wheels and it just moves faster. That's interesting. So you're saying, is it like a system prompt that you have? Is it like a skill? How do you get to the point where agents start complaining back to you? Because they're not trained to do that. Right. No, they won't. So the way that I do this, and this is kind of getting into my own factory here, I use Kodi. So Kodi codes, let me, if I can, briefly explain. So every one of these agents and software factories, they're building this ecosystem that is a walled garden, where the deeper you go, the more helpful it can be, but also the harder it is to switch.

So Kodi is a separate place where you can keep your credentials and your integrations, your automations, all of that, and all of the agents can use that. And so that's what Kodi is. And so I use a Kodi package or automation. It's just like a repo full code that Kodi can execute. And what that does is it is basically like managing the process of taking a PR to production. And so the agent has a very short skill that says run this CLI command and it will just get you the next thing. So progressive disclosure. Okay, run the command. Okay, great. I just marked it as ready for you. Now you need to wait for CI to run and for the agents to review your code. Okay, next tick. Okay, now there's an error in CI. You've got to fix that. Or here's this error of you. And so you just, like, once it's finished with its work, it hits that endpoint again or that CLI again, and it gets the instructions. And so one of the instructions is, okay, everything's green. Merge it. So it merges. And then, okay, now it's into production.

Great. Next tick is, okay, now what was so bad about this? And so that's the process. It's just like it's going through this tick to get it in production, but it's not done once it's in production. It sounds like a factory to me. I do want to hear from you. What excites you right now in the world? Because you are tracking the latest developments. You are advancing yourself and showing other people, this is the way the factory should be. What I'm most excited about is what I call invisible software. Everybody is building software factories to make building software easier. Well, guess what? If it gets so easy, then why can't everybody do it? Yeah. And with proactive agents, like the thing is people will say, well, your neighbor doesn't want to build an app. Like they're not going to even think that they should build an app. Well, they don't have to because you have proactive agents that say, I noticed you have this problem, so I built this app for you to solve it. And so we're going from hundreds of millions of developers who build software to billions of people whose agents build software for them. And that's invisible software. They don't know it's software. They don't care. It doesn't matter to them. But it's software that's built for them. And I'm building Kodi so that I can be there to support that future.

That is incredible. Ken Dodds, thank you so much for joining. Thank you. It's always a pleasure to see you at one of these, but you never were on the podcast before. I'm very happy that you came. Yeah, I'm thrilled that you invited me. Thank you, Alex. Thank you, man. Thank you so much for joining. Thank you, buddy. All right, folks. Next up, we have another guest, a friend who's also joining for the first time, Michelle Chen. Michelle. How's it going? Welcome to the podcast. Folks, we have Michelle Chen. You are a product manager at CloudFare. I am. This is your first time on Thursday. I welcome. I'm so glad you're here. You had a talk, and you're one of the first three or four talks, which is incredible, and you talked about CloudFare as being a neo-lab. And I am sitting over here and thinking CloudFare is a firewall. Not really. I've been building on workers for a long time, but folks who watch the show know my love for CloudFare. But I would love to hear from you. How do you talk about that? What makes Cloudflare a Neolab? Yeah, for sure. Yeah, Cloudflare, I think most people know us as a CDN company.

That's like 16 years ago we started as a CDN. We've actually been a developer platform for like 10 years already. Yes, workers. Shout out to workers. We love workers. Everybody who built this can. Observable objects. Yeah. And everyone that builds on us. But it's funny because I feel like even six years later, I still have to tell that story and people get shocked when we say we're not just a CDN. And then again, extending that now, we're not just even a developer platform. We're also a Neolab in the world of AI. And so, like, I would say I would give ourselves that title now. We released some new models. Yes. We released, like, four models in three weeks. Wow. At some point, we were doing a weekly model release, and I didn't sleep for, like, three weeks. Yes. There might be more coming, like, tomorrow. So, spoiler alert. Oh, okay. We love embargo and spoilers on Thursday night. But we are live, so, you know, just, like, watch. What's the stuff that you can talk about? We do love a good spoiler. I know. Thursday gets the spoilers. Yes. Michelle, I want to ask you, what do you define as a Neolab? What makes a company a Neolab? Yeah, I think you're trying – I think things that we do here is we try to push the frontier of research.

We actually – not only do we train models and release them, we actually open source them as well, so other people can be building on it. And we've seen so many good people building forks of class or fine-tuning it to fit their use cases. And I think that's what makes it a Neolab. It's not even just about hosting and serving models and selling tokens, but actually like are you contributing to research? Are you helping push the frontier of intelligence? And I would say like we're doing a pretty good job there. So I feel like that's a good qualifier for what a neural lab is. Are you contributing back to the community that everyone is trying to build on? Yeah. And then AI comes. And it's kind of strange-ish to Cloudflare because you talked about like the global network that Cloudflare has built. That network, I remember from Matthew and Michelle some stuff about like there are some GPUs there but not really GPU cloud, et cetera. with small worker containers. How do you get from that to actually serving instruments on Twitter? I would love to hear how this culture turned into this behemoth of serving ag. Yeah, it's been a long journey. I'm so fortunate because I think I was there from the beginning of it.

I remember when we made the decision to have GPUs on our network. We made it in three weeks, and people were flying around the world with GPUs in their suitcases and installing them in the data centers that we have. But like I said in the talk yesterday, one of our hypotheses was that the models were going to get smaller. And so we purchased like a smaller chip that would, because we thought the models were going to get smaller, so network latency was going to be more important. And so we actually installed those, unfortunately, like fortunately and unfortunately. I will say that there's no good or bad prescription here, but the models got larger. And so we actually had to pivot. And then we ended up having to get like larger chips in order to serve these like trillion parameter models. So the evolution kind of just followed the industry as well. But I would say, like, after we started surveying these large models, our infrastructure and our engineering, machine learning, engineering teams, et cetera, like, really stepped up because we were able to do so much more. It's actually crazy how much a hardware generation of chips will affect your serving, right? There's so much more you could do.

And so from that, I think that's only March 2026. So it was really like just a few. Really? Yeah, that we started serving these really large. March this year? March this year. Wow. Yeah, and so, you know, we first launched Kimi March 2026, and now we've been able to do so much more with it. I feel like we've been really, really reactive, and, like, we had this hypothesis. It was wrong. Let's pivot. Let's serve these large models. Let's get really good at serving them as well. That's kind of how we evolved, and then now we're, like, you know, figuring out what else could we do, and, you know, training models and pushing that frontier as well. Some of the models that you trained are different. They're different, and I think you talked about this on stage a little bit. I would love to chat with you about the type 1 and the revolution that Diogo Elmeida and the folks at TypeSafe brought to the world. I think you mentioned this on stage at your talk as the winds of the east came and suddenly things clicked into place. Tell me about that story. I would love to hear how does Cloudflare see type 1 and decision models? What makes them different than regular LLMs as well? And how does this fit the Lego bricks at Cloudflare?

Yeah, for sure. I think the decision models are just, like, such a good complement to what agents are doing today with LLMs because these LLMs, we're just sending everything to them. You know, like, I will ask an LLM, like, Opus 5.5 to send her a div. And that's completely unnecessary, right? But, like, we just do it because it's the only tool we have. And I think decision models just added another tool to our repertoire where you can now do, you know, decisions really quickly, classifications, categorizations, et cetera. And you can do it really, really fast and really, really cheap. And so I think with TypeSafe, when they launched Jev, it was just amazing because I think it was the right place and time. I was reading, I was just researching this recently, but like in Innovator's Dilemma, people said like when you're a disruptor, a lot of it is actually just marketing and not technical innovation. Not to say that, not to knock on their technology, but what I'm saying is that like it came at such a right time and place for the AI industry. because everyone was just so tired of spending so much money on LLMs. It was time-consuming. People started getting fed up with it.

It was just the ripe birthing ground for a model, like a System 1 model. I think TypeSafe made it so that you have a very general classifier because in history, I think the classical machine learning models, if you wanted to do it on any domain, you'd have to hire an MLE team, get data sets, label it, train your own model, and then, yeah, you have a classifier only for that domain. So they cracked the fact that, like, it's a very general case, like, Zero Shot Classifier. How it fits with Cloudflare is, like, when we built Workers' AI, we had the hypothesis that models were going to get smaller. They got larger. But now they're smaller again. Like, system one model, they're small and they're fast and they're purpose-fit. And it fits perfectly with our infrastructure pattern where we want the model to be close to the user for the best network latency. And so I think that we decided to capitalize on that. We're like, oh, this is a hypothesis in action. Like, we can't just sit here with our hands tied because the model is closed source. Yes. And so we were like, all right, let's go build our own. And so we built our own models. We trained them. We released them for everyone as well. So let's name this. It's Clef. Yes, it's Clef models. Clef models, yeah.

We have three flavors of Clef. There's Clef, Clef Flash, and Clef Omni as well. So Clef is, like, the most high-quality one. Clef Flash is, like, fast. And then Clef Omni takes all the different modalities, like image, audio, and video as well, which is new. I would say, like, when we first launched these models, we were probably one of the first on the market to do, like, image and, like, audio and video. And then we also trained the models up to 256,000 context window as well. And Jev, a few weeks later, you know, they released a new version of Jev and they raised the context window and added image support as well, which is cool. But I'm glad that, like, we not only tried to, you know, emulate what Jev was doing and the success there, We actually try to make it innovative and different and, you know, have a larger context window and multimodal support as well. I don't think so. I haven't met Yoko, but apparently he gave like a great talk on stage. I don't think they are sad about this. I think they're so, so happy that the industry is adopting this new pattern and paradigm. And in the spirit of asking about adoption, are you guys matching the same kind of primitives on the API level?

So they innovated with like three new, like, Nul is there, but basically Bernoulli, yes or no. but with the distribution, how easy it is to folks to try this out and switch between them? It's a total drop-in replacement. We actually just copied there. It's completely compatible. I think throughout the industry, people set standards. Chat completions was a standard for so long, and so we kind of saw this as a standard. And so it's actually a drop-in replacement. The API shape is actually exactly the same, so people can try out Clef super easily. If you're using AI Gateway on Cloudflare, it's really easy because all the Jev and Clefs are actually all available. on AI Gateway, so it's actually really easy to use. And I think, like, what you said before about, you know, them being happy about more decision models, I think it's true. Like, we released another model after called Cleric, and it was the first of its kind. I love when guests on my show are going to where I was about to ask the question. Tell us about Cleric, Michelle. Yeah, so Cleric is an extraction model, but we, you know, we looked at the system one paradigm where, like,

there have to be other things here that are outside decisions that, you know, are also, LLMs are not the best at or not suitable for. And so we decided to launch a model called Cleric, an extraction model. And, you know, other companies started doing this recently, too. So I was like, I was so happy. What makes an extraction different as a task shape than a classification? Because you can just classify the stuff that you want and then kind of get them. Tell me about Cleric specifically. I think it's cool. like the thing that helped me understand it was when you like read an email you're not sure what the email is going to say you can't exactly pre-write categories for something that you've never read right yes and so with an extraction model you don't have to do that you can just say look for you know the email address oh so like so like a pii is like one tiny like class of extraction models i would like extract everything that looks like personality and like mark it down but this is like a generalized wow that's so cool how do I miss this the other cool thing is it actually returns like a confidence score

and a character span so you can actually reference it really easily in the input like oh I found this here go look for it and we'll give you a confidence score and our calibration is actually pretty good and so if it's unconfident you can just do an LLM review or you can do like a manual review as well and so I think it's really really helpful for that but it extracts and that's kind of like our flagship use case it can also compute which means like you know if you want the due date but then you only have the invoice date but you know it's net 30 you can actually compute what the due date is and then it can also estimate and so like if you are like oh I think you know I broke this my iPhone what do you think the estimation of the estimated repair cost is I can estimate that because we built it from an LLM it has like pre-trained world's knowledge and then we also can forecast And we, like, tested it on forecasting, like, Bitcoin values, and we actually traded a lot better than, like, some other models. I'm going to drop this only on Thursday. Okay. All right. But then I talked to my friend, and he was like, never, never trust a back day that's, like, you know.

Forecasting, looking backwards. Especially not in the current macroeconomical climate. Yeah. This is not investment. No. Nothing on the show is investment. Exactly. People know. Michelle, I do want to ask you some. Last one, Clef is open rates. Apache 2, shout out. We do have like a clap sound from my home studio about Apache 2. Thank you. Cleric will follow Apache 2. Can we ask? Yes, yes, it will. We have two variants of Cleric coming out. And we actually have another large decision model coming up. You guys are getting all the tea really early. This one we'd like to hear. See you next week Thursday for spoilers. But, yeah, we're big proponents of open source. I think we've benefited so much from open source. Like, the inference engines we run, the models that we run, like, it makes sense for all of us to be contributing back to the community here. And we want to really push the frontier of, like, System 1 research right now has been our focus. So I feel like that has been the cool part about it. And we're excited to make it open source. And, of course, it's hosted if you don't want to, like, you know, mess around with GPUs and try things out.

But, you know, for researchers or companies that, like, need specific needs and have a team to go try it out, it's all open source. We'll land on this. We're in San Francisco for AI Engineers Code. And the benefit of being here for many folks who are not watching us is kind of the bygist of what people are talking. Besides the type 1 models and decision models, we'd love to hear from you what you heard in the halls, what's going on. What did you hear that maybe have changed your mind or like something new that you learned around here that people will maybe, you know, make a decision whether or not to come to the next one of these. based on what you tell them. I was worried. I think when I was giving my talk, I was like, oh, this is too philosophical and hand-wavy. You have three philosophical quotes in there. I know I did. I think, again, people don't know the Cloudflare story. That's our step one is saying that we have a developer platform. But I think the other talks have been so cool

because we've been so technically deep and focused and just resonates a lot with engineers and developers. and I think people are trying to take away, like, how can they incorporate this into their workflows and their products that they're building? So I felt like that was really cool. You know, it's not just, except for mine, I'm so sorry, but it's not just, like, a typical talk. It's actually, like, practical applications that you can, you know, take home the next day and try it out. And so I think for AI engineer, it's just, like, the right aggregation of people here that can actually help boost your workflow and have an impact on what you're doing. I understand. And the design gets moving, And the IG is like a great point in time to catch up and see like what's going on. I remember a year ago, agents were not a thing. Agents were not working. And now agents are working. Michelle, thank you so much for coming. Folks, next up, we're going to move to a new friend with some breaking news. But for now, Michelle, thank you so much for coming up. Thank you so much. Have a good one. All righty, folks. Next up, we have some breaking news. AI breaking news coming at you only on Thursday night.

I just hit the breaking news button on the show. Ivan, Leo, thank you so much for joining Thursday Eye. Yeah, thanks for having me. You are working at DeepMind at Google, and the breaking news that we just hit is that embargo lifted, and you guys just shipped another model? What? Yeah, we just launched our new Gemini for Hydrogen model. It's a new flashlight flash model that is essentially competitive with the original Argon model that we shipped just a few weeks ago. If you look at the artificial intelligence kind of index scores, Gemini for hydrogen is comparable. And so the score is actually higher when you have the maximum thinking mode. We know that a lot of people want different models with different kinds of use cases. So even in anti-gravity, we have Slash Swarm, where you can use all of these agents running natively to get better results. Yeah, like this. Okay, and we also have like an orchestrator mode now natively that uses all of these like different models together. So it's pretty exciting.

I think like Hydrogen has been my daily driver for quite a while. Really? Okay, so let's backtrack just a little bit. Four-ish weeks ago, Gemini 4 Argon released. We talked about this. Genevieve was here on the show to talk a lot about the cool new 3D modeling stuff. Just two weeks after Argon, Gemini 4 Carbon was announced. but this week was the first time people could actually play with it and was released in GA. What are you hearing from Gemini Carbon? What's exciting? What are you hearing from future folks here at the engineering that are building with it? I think it's been really exciting seeing people like Carbon because we've been using it internally for all these different ambitious projects. For example, I think OpenCode wrote an entire new native application where you can build applications across different systems with Carbon. So they just ran it overnight and Carbon was able to propose a new system from scratch. Wow. Yeah, so the new system Rune was basically built on carbon. And so I think it's really amazing because carbon is a really strong and creative model. We spend a lot of time trying to make it better. And today I think we'll have Melvin at AI Engineer talk a lot about how we're able to ship these models very fast and try to improve them quickly.

And I sensed on my timelines that the vibes are shifting and Google is so back. People are saying they prefer carbon on top of Fable at many points, specifically around the price point. but also within the ecosystem as well. So maybe tell us, what's the best way to use carbon before we get to hydrogen? Where do people go to get those tokens? Yeah, I mean, the easiest way to try our models via the API is on AI Studio, which is, you know, you can just go to AI.studio. Other than that, if you're using anti-gravity, everything is pretty straightforward, out of the box. We've rolled it out generally available today. And we've been trying to use carbon in all these different areas, whether it's like clothing, scientific discovery. But I think to me the craziest part about Carbon and Argon was that you have a 1 million decode length And so what that means is that for all these really complex tasks If you just install the latest version of the SDK Carbon can just keep decoding all the way onto itself the problem And get to the ultimate outcome for you Tell me more about this in simple language 1 million decode length

Compare this to me to other models out there What are we talking about? How is the improvement and what it affects? Yeah, I think for most models you would see a maximum output token of maybe like 64,000 64k is usually, yeah. 128k, I think it's like a lot of the models today. Yeah. But I think carbon and argon and also hydrogen are probably one of the first few models where, for the first time, you have a model that can output a million tokens in a single pass. Wow. And so, you know. Just like keep spitting tokens. Yeah, we just keep spitting tokens out. You let the token. With no max length. We're not like end of thoughts. Wow. I mean, the max length is one million tokens, right? So it's one million input, one million output. Yeah. Which is pretty unprecedented in my opinion. We use it all across Google today. Yeah. We've been refactoring systems, trying to make them more efficient. And you just need to – sometimes if you have 128K contacts, like output tokens, you're just not able to get the right kind of output quality that you want. And so I think it's just we're trying different things. And in this case, 1 million output tokens is – I think something that's really exciting, especially for a better model performance, more stronger, more performing model. Yeah. I do want to ask you. So obviously, folks who work at Big Labs, they have, as part of their job, unlimited tokens to use their models.

but then because of that there's certain patterns that like come up with using them internally that then folks after the models like pricing goes down people start to adopt some you want to share some of these patterns of like how you said you like hydrogen is your daily driver and congrats on the releasing on hydrogen as well but but carbon is like really good and like long output tokens as well how do you mix how do you use them when do you use them when like pricing is not let's say concern and if pricing is a concern where where folks where would you recommend folks sending attention? Yeah, so I think for me, I use Hydrogen as a daily driver because a lot of times I want to be able to quickly prototype and iterate, and Hydrogen has that speed. If you look at Hydrogen versus like Gemini 3.8 Flash, it's like almost 2.5 times faster. And it's very efficient in the way that it calls multiple tools at once. So when you're doing daily tasks, you're prototyping things, you're trying to draft email replies, I think it's a great use case. What I really like Carbon for is that I normally send it on very long, obscure journeys throughout the company's code base. to try to figure out problems.

So when I, for example, I had a, I work on Kaggle right now where we onboard a lot of different benchmarks. And so one time we had an issue with our model proxy. And so what I, we were getting 400 errors with a specific model. What I got Carbon to do was say, hey, I'm getting this error. So Carbon tried to reproduce it locally. Then it dispatched a bunch of tasks on our background infrastructure. And then it realized that someone had merged into the PR that changed the permissions. And so it found the exact person that emerged the PR, drafted the message for me and said, hey, you can send this now. and so basically i think like it's stuff like this that you wouldn't expect a model to be able to do which is like find autonomously the person responsible for the change yeah debunk the change show proof of it and i think this is just a case of how like you know these models are just meant to be used differently i want to prototype with hydrogen but with these really complex gnarly tasks i really want like carbon to be just there like cranking in the background hydrogen is cheaper as well uh do you have the price off the top of your head it's roughly the same price as germany 3.8 Oh wow, that is awesome. And we're showing the stats here on stage. We got the sweet sweep is 13% versus the cloud summit is like a 3 and GPT 6.1 Tera. So just based on the benchmarks alone, I'm going to make this observation, you don't have to.

you know, carbon is kind of the fable astral level looks like and like many folks are comparing that carbon to like fable 5.1 before carbon took it out, hydrogen based on the benchmark that you guys posted is like sonnet and terra level this is like that level of balance between capability and speed in addition to releasing Gemini 4 carbon to general availability this week and now Gemini for Hydrogen. Yesterday was another release. Yeah, we launched our new Decisions API. So basically with this new Decisions model, for the first time, you get a truly multimodal Decisions model, which can take an audio, video, PDFs, text, and at the same time, you can have almost 1,000 questions that you can basically get answers to. I think what this really proves is that, you know, you have a model that you can use for a lot of lightweight classification at a very cheap price. If you were to take the same workflows that you have with Gemini 3.8 Flash and put them on the Decisions model, which is only like 10 cents per million outputs. I think it's like very cheap and you get, basically you don't really need to change much before.

And I think for a lot of different use cases, like we've seen, like this is something that's been rolled out in all of our like 1P products. Who will search uses it right now, just search for like malicious behavior. It's on the article. And then I think Stitch uses it quickly to kind of prototype and find stuff and we have classifiers running. But it's all possible because now we have models that run inference like very cheaply with this specific new business API response. I think the highlight for me when we talked to Philip was multimodality fully. Tell us about Diffusion API being that multimodal and kind of what are the benefits of using this. Yeah, I think it's really cool to have something that's like multimodal. So Thor previously, he had like this demo where he could talk in real time to a website that was using the Diffusion API. And he could actually classify what language he was speaking in. So he would speak in German, now it's English. and even because it's natively multimodal and it understands like your audio, like your intention, your voice, because it's that native modality, it was able to classify, oh like this is a frustrated like a voice, this is like someone that's like angry. Really? Yeah and so like Toro is basically

showing like oh now I'm speaking, now I'm coughing, now I'm like happy, now I'm like sad and with all these different modalities and even changing the language because it's multimodal, you don't have to have all these other systems to hold it on. So no longer transcription and static language analysis on top of just raw text, right? You just take the raw audio, you pipe it in, and the decisions API will be able to help you kind of get that classifier if that's what you're building. Yeah, and video as well. Video, image, audio, PDF, text. Swarm in anti-gravity. Yeah. What? This is so cool. Tell us about this. Swarm. Yeah, sure. I think the part I like the most about Swarm is actually it's super cute. So when you start a Swarm, you have this little bee that sends past all these other little bees. And you can see it visually in the actual anti-gravity UI. Nice. So I think like swarms are just another form of like inference scaling where you know you can just map out and just get all these different models to be collaborating on different things. Yeah. We've seen like these crazy results over swarms. You can create these like whole games that look like rollercoaster tycoons, alpine games. And I think this would take forever or might not have been possible with like earlier models.

Yeah. So swarm is just a really interesting way to do it. So in Anti-Gravity, folks can now invoke Asian swarms to do tasks like the Asian swarms basically did themselves at OpenAI when they hugged a hugging face, right? Yeah. and that they have a way to also collaborate between them as well? Yeah, they have a message board that they can post messages for each other. So they don't have to go and hack a message board online? Yeah. Hopefully not, you know. That would not be great. But, you know, at the same time, we also have a native way for you to orchestrate different models together. So now Carbon is orchestrating all these mini, like, hydrogen models. Yeah. So we have Swarm and we have orchestrate. And I think it's just we want to let people use our models better. So, hydrogen as the executors in... Yeah, I think that kind of is how we see the thing going. You want a strong model to sketch out what are the unknowns, map it out, and then you want a smart model like hydrogen flow and it's fast and speed. You know, batch all the parallel two calls and get things done. All right. So congratulations on the releases, including one today.

I'm very, very excited when folks who break the news actually come and talk about the breaking news. Ivan, this is your first time on the podcast. You're always welcome back. Thank you so much, man. Let's go. Thank you so much, folks. Ivan Leo from the Gemini DeepMind team. Thank you. And now we have another guest coming on the show. And then another guest there who is trying to wave down. A person I've been trying to get on the show for the longest time. And as he sits down, we'll show you who this is with us. Welcome to Thursday AI. a person who knows a thing or two about headphones, Zach Lloyd, the CEO of Warp. Thanks for having me. Warp Factories now is the new thing. We'll get there. Zach, you just came from the main stage. I did. Thank you so much for doing this back-to-back-to-back. I really appreciate it. No, I'm all warmed up. How was it?

How was your talk? I think the talk was great. I did one of these... I think it was awesome. Folks are getting what you're selling, which is great. Yeah, I asked the audience, I was like, how many folks are currently trying to build their own internal software factory? Yes. And it was almost everyone, which I thought was pretty cool. And if I'd asked that six months ago, I think people would have been like, what's a software factory? Zach, what is a software factory? Okay, I just find this on stage. So a software factory to me is something very specific. It's a code-defined closed-loop system for measuring and improving your cloud coding agents over time. Just to unpack that a little bit, it's like you want a system where you can define across your entire organization how all of the coding agents run, and you can declare that in code.

and then you can basically describe what are the workflows that they should do, and then you need a whole bunch of tooling and measurement for improving how that works over time. And that tooling part and that measurement part and that closed-loop system are really key aspects to it. I would love to hear more about the measuring part. Obviously, evaluation comes to mind. Is there anything more than this into this whole world? Yeah, so evaluation is the basis. So when you have a working software factory, the key thing is, like, you should be recording what all of your agents are doing. So all of their conversations, all of the screenshots, all of the videos they produce, whatever they produce needs to get recorded. And then you want a system where you can run evals on that. So it's like, did those coding agents do a good job across different things that you might care about? So did they produce high-quality code? And then you can go two layers above it So one if you define your factory as code you can then set up agents that will look at your evals and try to find areas to improve the factory We call this self Don we call this RSI

Yeah. It's all one big concept of the loop where agents are observing their own work and trying to improve it. And so your factory should have that. And then the other thing you should have the ability to do from an improvement standpoint that's really important is be able to replay past work that you've done on different configurations. So different models, potentially different coding agent harnesses, different skills. And that way you can measure if you're actually using these coding agents efficiently. And so the mindset that I think folks are getting into is what I call factory engineering, where you're just spending your day building your product. you are spending at least part of your day thinking about, like, how do I use these agents more effectively to build my product? Yeah. This is definitely the Zeitgeist hallway conversation in the hallway track, whatever Swix calls the hallway track, right? Some people come to see folks like you on the main stage. Some people, after the main stage, they hear some concepts that suddenly make sense to them

based on what they've been, like, feeling for the past, like, six months, year, et cetera. And then they really want to talk with people who also get this. And you all now have, like, this, you put this, like, metaphor in their minds for them to all, like, share the same thoughts. This is why AIE is so important, I think, for us, like, pushing the frontier. Would love to hear from you what else you're hearing besides, like, software factories that's happening in the hallway of AIE. What's new from the last time? This is not your first time talking. You've been here. No, it's not a few, yeah. A thing that I'm hearing a bunch of is how do you get better at verifying the agent's work? So this idea of, like, verification either through computer use or verification through formal methods. Like, I listened to a talk from Potato yesterday. Yeah. Lauren. Lauren 10, yes. And she talks about this. And I think, you know, there's traditionally been this big problem of you have agents that write a lot of code. That code gets stuck in code review. You don't know if it's any good. You don't know if it's high quality. And so the two approaches to fixing this, one is, like, you can do more upfront work before an agent works.

and spec out what it's going to do, and I think that matters. But then there's this interesting question of, like, how do you verify, rather than, like, having a human look at the code, but, like, have some sort of formal verification or computer use verification to make sure that the agent's work is good so you don't have to spend as much time reviewing the code. So that's something I'm hearing a lot about. I'm very excited about this topic. Last week on the show we had Max Kanata-Alexander. I don't know if you know Max. He's also, like, an NAE speaker. Okay. He's working on a new language based on this, and he's basically, His thesis is our tooling is not fitting the next era. We need to rebuild not only the whole infrastructure. I totally agree that there's a need for innovation at the programming language level. Maybe something that's more agent-native, where the assumption isn't that a human is going to do all this semantic and syntactic review, and you can basically specify up front how should the system work and then prove that the thing that's been written does that to the extent that it's possible. I don't know if it's fully possible. So tell me, what's your daily driver? What changed in the past, like, two, three weeks, let's say?

Yeah, so we're using a combination of models. We do a model route solution. And part of it, for the simpler things, we're using open-weight, so, like, GLM, 5.3. And for more complicated stuff these days, we're using, I would say, primarily the open-weight models, so, like, Sol or sometimes Astra. The way that we got there, though, I think is the more interesting thing, which is we just tested everything. And so we can, on our own tasks, be like, okay, how would it have done with Opus 5.5 versus, like, you know, the latest Thul versus GLM? We can test how would Pi Harness do versus using the Codex Harness. And so for us, it's like a really empirical question, and I expect that it's going to continue to change. You know, the anthropic models are awesome, but we do find they're pretty expensive. and so you know we're you know as a company spending a lot of money on these coding agents and we want to be efficient with it and so you know we're really trying to be smart around like

when do we use what let's speak about the company itself yep i am a fan of what you built thanks and i just want to tell you that the work that you guys did affected my life in a very positive way for folks who don't know what warp is uh or was i guess still is we have two products give us give What's the lay of the land of the Warp ecosystem? Yes. So Warp, for the first part of its existence, we built an agentic terminal. And so that's open source, 65,000 GitHub stars, almost a million developers used it. It's the best one. Thank you. I still think it's the best one. I appreciate that. It's incredible. Yeah. And the cool things about it, it really changes the user experience of using the terminal. It makes it feel like a modern app. And then we were very early to build in an agent-driven experience, So you could always just kind of talk to Warp in English or in commands, and it will do what you want. It'll fix all the annoying, like, terminal-related DevOps issues that you have for you. It's very, very good for that. And then we have a second product, which is what we're, you know, candidly more focused on right now

because I think it's more of the future of the business, which is our factories platform. And so, yeah, the way to think about that is it's infrastructure for companies that want to build their own software factories. And so it's a little distinct in that we're really trying to, you know, provide something that a platform team or an engineering leader can be like, okay, we want to invest. We're going to build our own factory, but we don't want to spend all this time building all the building blocks for it. Like, why are we building sandbox and computer use in a way of defining workflows and a way of defining factories as code and benchmarking and all this stuff? So we're trying to provide all those primitives as an infrastructure stack. And that's going super well. There's just a ton of demand right now from companies that are trying to figure out how do we move from a world where every engineer does their own interactive prompting to a software factory. Zach, I think Slicks is going to come on and disrupt us at any point now. But before this, I do want to ask you, where do you think we're going from factories forward? I keep hearing the word engine.

Did this pass cross your radar already? Are we like that fast? Are we moving that fast? I already changed the noun. No, my thought is software factories will expand to be product factories. So we're starting with more of just the software lifecycle specifically. Yeah. And then it's going to be like, can you take product signals from your users and feed those into the factory, have different user personas, use something like a factory. And then eventually it'll be like, how much of the business can you architect agents with a factory approach? Like HR operations? We're not going to the HR business yet. I think the general idea that you're engineering all the agents that do things at your company is going to be lasting. I don't know about an engine. It could be next. Maybe we'll do warp engines. Yes. I think that what you guys are building is very fascinating. And you're talking about building blocks. You were mentioning them very quickly. I would love to hear more about what are those building blocks of a factory that you guys are providing. Yeah, great question. So the way that we think of it is as an infrastructure stack. And so at the base level, we provide a code format in which, think of it as like Terraform for all of your company's coding agents.

The layer on top of that is we provide the ability to store all of the agentic data output. So like traces and logs? The traces, the logs, the artifacts, all of the records of the interactions you might have had with the agent and things like GitHub or Linear. So you store all of that. So not only the direct interactions, but also like the... Anything that can be used as a signal for improving the agent. And, again, when I say we store it, we also work with companies so they can store this on their own infrastructure. Like, we're very happy to do a self-hosted solution. We provide all the compute layer. So it's like we can host that. And that means, like, how do you define where these agents run? How do you define what their environment is? What secrets? What MCPs they have access to? And, again, that can be we host it in partnership with a sandbox provider or we can do it on. And there's another layer on top of that, which is the pluggable inference layer. So we provide the ability to run a factory with any coding agent harness that's ACP compatible. So you can use – we have our own harness, which is very good, but you can also run it with Pi or Cloud Code or Codex.

You can run it with multiple models. You can run it on a model gateway. On top of that, we provide this layer that we think of as measurement and improvement, so all the stuff that gives you the metrics, the ability to do self-improvement, benchmarking. And on top of that, we provide an orchestration layer. So think of it as like when work comes into your factory, what happens to it? Does it get triaged? Do you loop in a human to write a spec? Do you move it immediately to implementation? So all of that engine we provide. And then finally, everything is API-driven and MCP-driven. And so we provide all these different integration points into your existing tools. So think of, like, if you want to run your factory primarily through Slack, we support that, or through Jira, or through GitHub, or Azure DevOps. And so the sort of list of integration points is actually one of the big areas where you provide a bunch of value because companies, when you go into bigger companies, they have, like, very complex systems with lots of different integration points and requirements. And so we really help with that aspect.

All right. This is an incredible breakdown that I really need to re-watch again. Yeah. Because you just gave the playbook to companies to basically build factories. That was the talk that I just gave. I went through the whole stack. For anyone who's building it, I think those are the right layers. I think, from now on, I'll tell you what comes in immediately after you. I want speakers immediately after the keynote, so there's still one. I got the whole thing. So fresh. Zach Lloyd, thank you so much for joining us. It's great to be here. Thank you, man. All right, folks. Next up, we have, we're going to flag them down. Next up, we have, he knows the way of the land. Next up, we have Swix joining us while he sits down and grabs the microphone. I will just say, this is AIE Code Summit. Number 10. This is number 10 AIE? The 10th AIE. All right. I did the count yesterday. I didn't even know how many. It's been a lot. Mr. Swix. Hello. Welcome back to Thursday.

Thank you for having me. Welcome back to AIE. Folks, can we get wide? Welcome back to AIE. We're in this beautiful podcast setting. Yeah, it's getting up in production value. Yes, significantly more so. Look at this nice monitor we have, but that's not what I want to talk to you about. Can you show the daily context? Folks who don't know what this is, Swix is holding the daily newspaper that this conference is printing. There's a person over there by the... Sorry? Oh, sorry. Yes, we have people who are distributing and writing. Yes. Sponsored by the Major League Hacking folks. Yes. Shout out to them for this. Who are also now our exclusive hackathon partners for all AIEs around the world. Oh, that's incredible. So basically when I started AIE, I was like, one thing we'll never do is hackathons because I am in the content business. I'm in the speakers business. hackathons is like a whole like judging and whatever and so we just partner with hackathon partners mlh has been in this business longer than uh anyone i know yeah they're they do hundreds

of events around the world they they do this really well so i'd much rather partner with them than compete with them we we also i think simulcast on the dev2 uh page by mlh as well so shout out huge shout out to them and thank you so much for supporting uh thursday as well So, SWIX, day two of AI Engineer Code Summit, I've been getting speakers. Zach Lloyd was just here, Michelle before, et cetera. The thing that I asked them is Zeitgeist update. I think the first thing for me about it. Can I suggest a speaker? You've got to get Brandon. It's impossible to get him right now because he just did a talk. Everybody standing there talking to Brandon? Yeah, yeah. So, Brandon is from Artcraft, which is rewriting all of Adobe. in open source oh that's the guy that's the guy oh you missed this talk because you're doing 30 yeah yeah this is why it's a bad idea to do this yeah but you know obviously you have a tradition yes but uh that guy so originally i was like okay i'll put him in like a small track he's unknown

name and all these things and i talked to him and that guy is a cult leader so really he's a cult leader he's uh he's gonna do super well so um while you have him here you might as well all right i I need to grab Brendan. Maybe like a separate, you know, after the live stream. Yeah, folks, I can't promise Brendan on live stream. It looks like 7,000 people over there want to talk to him. But this is the clean room rewrite of all of Adobe Suite, right? That's the one? And Microsoft now. And Microsoft. He's doing docs and slides. Wow. All in last, all like feature parity, et cetera. He has a good software factory. Then you can do it. Yeah. then it's not like he's going crazy with managing 20 different projects. It's the same factory. The same factory is doing all this. Yeah. So this is kind of what the software factory gets you. And at AIE, the Zyga is his software factory. We talked about this with Zach. We talked about this. Yeah. What are the other ones? So with Zach, obviously, you're going to talk about factories. Yes.

With Diogo from TypeSafe, you're going to talk about System 1 models and decision models. Yeah. And then the last one, which is today's theme, is increasing ambition, including rewrites. So Jared Sumner is going to come and talk about Bun and how, yeah, okay, he rewrote Bun from Zig to Rust. And that was just a weekend casual project. But now he's also pursuing Bun AOT, which is kind of the most ambitious version of Bun that he always wanted to pursue, never had the opportunity until he worked at Anthropics. So increasing ambition. Yeah. And I think that was the opening keynote theme that I gave to the keynote speakers, which is now we're at the level of not really worrying about implementation detail and not really about token maxing. It's about how to be more ambitious with the projects that we do and we have the models to do it. My friend Swix, how has your ambition risen in the software things that you're building? And I would love some examples directly that folks may not know that are happening in Swixco and around AI. Yeah, Swix Inc. Swix Inc. is a real company.

It's a real company. It's in the world. Yeah, so there's basically AIE, Latent Space, and Small AI. I haven't talked as much about Small AI, but it is in there. And my ambitions have increased, right? Like, you know, I've rewritten my own GitHub, my own Google Drive, my own Calendly. Slack. My own Slack, yes. These are not three things. They are one thing, and it is Small AI. But also, I think even for AIE, we've increased in ambitions in terms of what is possible for us. We've crossed a million subscribers on YouTube. Shout out. That's incredible. I actually don't care about subscribers anymore. After you hit a million, you're like, okay. But one million is like a very, very nice milestone. You need to be congratulated on that. This is incredible. There's so many people who get a lot of benefit of putting this online also and repeat it as well. And that's incredible. So shout out. So what's funny is that we've actually had 2 million monthly unique viewers a month on the channel for most of this year. So actually, it was just mostly converting them from viewers to subscribers.

Just like, hey, like and subscribe, you know, they're very boring, simple things. But yeah, I mean, I think that's a good milestone. Like, does increasing ambition mean, oh, the next milestone is 10 million, right? No. Actually, it is how can we level up from being a conference that produces this kind of content to something even more substantial. So next year, my goal is to start AIEOU, the nonprofit education and certification effort. I want to do leadership retreats. I want to do more continents. We've just announced AIE India and we're working on AIE Africa. and I'm going to be going to AAU Sydney for NeurIPS as well. So, you know, those are the kinds of different things where I think we can sort of add to people's lives rather than increase quantity just for the sake of quantity. I think that is true in this event as well, right? Like next year's World's Fair is going to be 10,000 people. Wow. But here we intentionally go between, for the summits,

between 800 to 1,000. It feels like a good enough intimate thing where you can kind of meet everybody, meet every studio, You go, watch what you want to watch. Talk to every expo person and not feel overwhelmed. Yeah. Wolf's Fair is overwhelming, I get it. But we need one of them. Yeah. Okay, this is everybody. Just walking the hall, seeing the same people from day to day, I think it decreases entropy. And you can have a continued conversation from the hallway track a day after. And what we did yesterday with Ray Fernando, who's helping out, we did a hallway track online. We really tried to bring people to see what's going on in these hallways. And we talked to everyone. Everyone was huge, just like standing out. Potato from SpaceX AI and Lance from Anthropic and Lydia after her talk. And Sunil, our friend, who's now also in the... Did you see Lydia's talk overnight? Not yet. So Lydia had tech issues because Anthropic's IT security is AGI. And so we couldn't see her talk. Like Claude shut down her computer or something in the middle when she tried to share the screen from Anthropic.

So she went home and recorded it, and then we posted it at the speaker dinner. And in 13 hours, you got 200,000 views, which is the highest. From yesterday that she pressed the button herself, 200,000 views? Yep. Wow. So, I mean, obviously, people are very interested in Anthropik. Yes. She is a very good speaker. She's incredible. And also, I think AIE Code, we featured it as, like, a feature keynote, right? So, obviously, it was going to get more exposure. But 200,000 is probably 100K more than any day one view we've ever gotten ever. That is incredible. She's also great. We chatted with her during the live. She was in the hallway. Folks, you can come here and meet the person who gets 200,000 views. Just ask them a few questions about their talk. This is the type of stuff the whole talk gets. Okay, so if you don't know who she is, I know who she is because I've been in this field for a while, and I've seen her talk. So for me, I have the good fortune of just inviting people that I know are going to do well as future keynotes, and that's part of what AIE is versus our competitors.

So, three zeitgeists. We're talking about software factories, talking about new type of intelligence that's faster, system one models. We had Michelle Chen on here just before, a couple of stuff before you. She talked about not only decision, she talked about extraction, which is like also type one, but in a different way, which is very, very cool. So, I wrote it in the article here, The Case for Smart Software. Everyone go read it on Dev2. yeah I mean I think there's a lot of potential for AI engineers not to only just talk about oh is it Pi or OpenCode or CloudCode or Codex like no like actually there's a lot more engineering that we can do with LLMs and things that are not LLMs like the Clefs and the what's the other word? Cleric. So I mean like I think it's a very inspiring time because I think AI engineering is a little bit stuck in a rut of like well you can build bigger and bigger harnesses and then you can mess around with the prompts and like, okay, this year is NCP, next year is CLI, like all this like really minor changes. And then these things come along and a completely different category of AI engineering is opened up.

And I think that was what I was always hoping for. Are you partial to the word unhubbling that Leopold used in situational awareness to call these like new things as unhubblings? I don't know if Leopold coined that term. I think it's just a common term. He grabbed it and like popularized it. Actually, Anthropic uses it more. Tarek used it in his keynote at World's Fair. Yeah. And I think it's just a term that people inside the Frontier Labs use, basically just saying, like, don't put too much restrictions on the big models. Yeah. And I think that is the framing if you are in sort of the big model cult. And that's what System 1 models are trying to push against. It is not unhumbling to add software on top of models. It is actually making them more reliable and doing things in high speed and high quantity. Interesting. So, I have to ask you about this morning's news. Yeah. How do you see the world of AI where Jensen Huang is no longer CEO of NVIDIA?

It's very shocking. I don't agree with the decision. I think that he should have owned it. He should have said, what do you mean? I have 50 women, you know, and, like, all of them are consenting adults. get out of my private personal business. I agree that it is a very serious matter. She's a Chinese spy, obviously. And obviously, NVIDIA is a company of national interest. And yes, I think at some point that does matter. But the why now is the most curious part. Somehow, very curiously, conveniently, it's right after the elections. Yeah. And that's interesting. I am not qualified to analyze it beyond that. I think that it is very negative for the future of NVIDIA if such a visionary leader, a strong leader. By all accounts, everyone who's reported to him that we know personally, he runs the company so well, it's very hard to replace that guy. Very technical. Deeply technical.

And I was also talking to a lot of the attendees here, and a lot of people think it's bad for open source AI. Yeah. Because, you know, he bought Hugging Face and he echo hired Poolside and he was building up to this roadmap of American open source AI. Yeah. And what now? His CFO is running the company. So we'll see who takes over eventually. It might be the case that he comes back. I don't think he will. I think that this if this move was strong enough to dislodge him now it's quite hard to make the case to reinstate him unless the CIA clears him and the CIA is not going to do that this does feel like a sovereign this is a state level issue this is not private company internal politics anymore it is not even at the level of opening eye firing Sam Altman yeah this is the cia says you're in trouble yeah uh this is not about the affairs yeah it's about

who the affairs was with like this is the second affair it's all like mass media like bullshit i believe like okay yeah people have affairs and like you know it's it's not uh ethical but like it happens right and people survive um and like they would not have made him resign based on that but um yeah so we're in the world where nvidia uh is no longer run by jensen it's crazy Jensen has spoken remotely at AIE for like two minutes before, and we were working to book him to come to AIE for next year. And I'm very disappointed. And I think he's such an inspiration to all of us. Maybe now he will have time to come. If he wants to, obviously. I think we are not his priority. but like and even as you know someone that is Asian American like trying to show like learn about like what people like us can do in tech and the lead in tech

it's like we lost a great leader we lost a senpai yes he's still alive though we lost him as the leader of a company and as a proponent of open source and as a person who really really strongly was like for western values and open source values and like America on screen, et cetera, but with some reservations about how much to open up to China. But it's all state level. I agree with you. So the one that's like on the books right now is London next year. London is in February 17 and 19. There's tickets already on sale. CFP speaker application is already on sale. Nice. We will be next to Pokemon World Congress. So either catch AGI or catch them all. Nice. And I'm working on like Pokemon cards that you can kind of unwrap and give out and trade speaker cards. What kind of folks should apply to speak or you're done? Yeah, it's very interesting. Basically apply to speak if you're working on something cool and think it needs to be shared. But make sure it's actually cool

and I think there's a lot of people that are just doing it to check a box. We have more speakers than we need, right? Like we have roughly, we started out maybe with like a 16% acceptance rate for our speaker applications. And then last year, it was kind of like 12, 10 to 12. And now we're between 4% to 5%. Of acceptance rate of... Of applications to people actually on stage. Yeah. And I think... I don't like turning away people just to boost numbers, right? Like, it's not the goal of ours. So, like, just... We're trying to put a signal out for quality. I will always try to invite people like Brandon from Artcraft, who never applied, but obviously, you know, is one of the most interesting software stories of the year. So we tried to do that curation model. where it's like kind of 50-50 apply versus invited. And that really keeps the content really fresh and interesting. The new thing about Europe is also we're starting to delegate away from me as the single point of failure towards a speaker selection council,

particularly with Europe where I'm not based in Europe. So we're actually starting to appoint former speakers to a committee that appoints future speakers. and so that is to me how I'm going to help grow in every continent the AI engineering community there for people who feel strongly about that community and that region Anything else that folks need to know about AI that they don't know while we have you here what still excites you to bring this thing to people now that AI engineers completely changed We are hiring engineers for the first time ever Normally we hire event staff So my job is I take traditional events industry people and I make them, I hire the best that I can produce AI events. But we're starting to build up more of a software platform. We're hiring engineers to try to roll things out. So we're building our own internal tooling. I ran this Kill My SaaS program, so we pay a lot of money for traditional event software, for ticketing and all this stuff.

We want to vibrate all of it, right? So I pay $200,000 a year for this software that we're using here. and I want to pay, I want to open source, I want to own it, I want to customize it, right? And so we're hiring engineers for that, but also I think we have the ability to offer services to our engineers. I cannot offer this right now, but one dream of mine would be, like, let's say you pay $2,000 for a ticket, you get $2,000 in inference through the AI engineer gateway where you can try out all different kinds of services. And the same thing with a token millionaire card, like, I want an actual Visa MasterCard thing that you can swipe and get points around the world. And so, to me, AI Engineer is moving. I would love to move it towards a more membership model for both our sponsors and our attendees, where you're just a member. You get access throughout the year to a bunch of things, including the conferences. One last shout-out is Leo, the 15-year-old kid. Let's walk around this house. He's 15. Isn't that illegal in something? No, he's not drinking alcohol in the halls here.

I really want to shout-out Leo, because we we chatted with him here running around which is like talk to this guy he's like hey i dm swix i'm 15 when i was 14 i dm swix and i told him like hey i i'm gonna pay for my flights can you give me a ticket and then and then you got him a ticket yeah you know sometimes these people i don't know if they're gonna be rock stars or whatever but he has high enough agency to uh to ship interesting things first of all second of all ask me nicely yeah and like yeah i'm gonna give a personal scholarship to come into these things, and if it affects their trajectory in some way, then all the better, right? Bro, he's 15. I don't know what I was doing at 15. This is so incredible to me. I think this is a great point for us to conclude the conversation on the positive note. There's a lot of folks who are still learning. I mean, this guy is 60, 70, and he said that his talk at World's Fair changed his career. Name, please? Frank Boyle, Professor Frank Boyle. I think normally teaches at Berkeley. And nobody really cared about his lectures at Berkeley.

But over here, people love him. He's reintroducing people to ontologies. And more power for it. A lot of us are reinventing things that have been studied for decades. And it kind of takes a professor. There will always be room for professors at AIE to reintroduce concepts that you should know. But you didn't know because you didn't go to college for this. You were kind of reinventing for first principle. So context engineering, graph engineering, a lot of this has an old name. Right now it's anthologies. Yes, that's incredible. Swix, thank you so much for joining Thursday. As always, to recap, AI engineer, it's been a tradition now at this point, and I think it's lovely that we get to hear from you directly about how you're getting the Zygate as well. Yeah, and always thanks for coming to every AI. This is the very first AI. I have the T-shirt yesterday. Swix, thank you so much for, first of all, building this thing, which is now incredible, and inviting me always to help show people the behind the scenes, etc., and for joining yourself and telling us about AI. Thank you, man. I really appreciate it. All righty, folks. You heard from the man himself, Mr. Swix, about AI engineer.

He joined our show. We brought you the unfortunate news. Let me bring back. Nistan and Wolfram are still with us. It's time for us to wrap up, but I will say, this was an insane, insane week in AI. We always say insane, But the thumbnail for this show was me just going like this. And this is how we all feel where Jensen Huang no longer is the CEO of NVIDIA, the company that brings AI to everyone and brings the GPUs, the liquid gold to everyone. But then after this, we had an incredible show. We had so many speakers that are going to try and butcher the order. We had Ivan Leo. We had Michelle Chan. We had Swix. We had Zach Lloyd. We had Emma from Nuance Labs. And I think I'm missing a speaker. but it was an incredible show back to back always the engineer the first virtual the first AI yes the first the first AI virtual guest we had Ray Fernando who helped us a lot we have the folks from Eyes On help us produce the show

Eyes On thank you guys for the incredible production I really really appreciate it it was incredible we obviously thank AI engineers we thank all of you for joining us live from Thursday Yeah, if you missed any part of the show, the AI Clippers will clip a bunch of these. Please follow our YouTube. Please follow our ex as well. Definitely check out yesterday's walkaround because yesterday's walkaround was also incredible. We chatted with Lydia Halley that's now exploding. Did you guys hear 200,000 views since last night online? And she talked about Cloud Code. We chatted with Lauren Tan Potato and we chatted with a bunch of other folks as well. If you are at AE, come say hi. If you're not, this is the signal for you. Please sign up. We'll see you next time in London. Wolfram, thank you, Nistan. Thank you so much, everybody who joins. And we'll see you here next week. Hopefully less intense weeks. Good releases, less bad news potentially for open source. Thank you, everybody, for tuning in. I really appreciate you here. And we'll see you here next week.

Bye-bye, everyone. Bye, guys.