I see a large IPO on the horizon.
You're surrounded by journalists. Hold your position.
Overnight success.
The place of 3,000. Double. Right. Let's... misinformation.
Courtney declaring order inbound. Let's just roll.
We are surrounded by journalists. Hold your position. Come on, get up. Trust the experts. Buy token. We are experts. Founder mode.
I see multiple journalists on the horizon. Stand by. UAV online. Blaze. Double blaze. Triple blaze. Double kill. Wrong. Cop is... please. Team Deathmatch. We are experts. Triple blade. Let's just roll. Right.
Marky clearing order inbound. Come on, get up. We are surrounded by journalists. Hold your position. Break one. Strike two. Activate golden retriever mode.
Marky clearing order inbound. 5. I see multiple journalists on the horizon. Stand by. Founder.
You're watching TVPN. Today is Thursday, February 5th, 2026. We are live from the TVPN Ultradome. Yahoo! The temple of technology, the fortress of finance. Let me tell you about ramp.com. Time is money. Save both. Easy. Use corporate cards, bill pay, accounting, and a whole lot more all in one place. We have quite the lineup today. Let's pull up the linear lineup. Meet the system for modern software development. 70% of enterprises on enterprise workspaces on Linear are using agents, and we have a lineup for you today. We got Sam Altman from OpenAI joining Sholto from... If you're not familiar, Sam Altman is the CEO of OpenAI. They are the makers of ChatGPT.
Yes. I did actually explain to someone. I was like, oh, do you know... I was explaining the show to some random person. I was like, oh, they were like, what do you do? What do you talk about? And I was like, oh, do you know like OpenAI? They're like, yeah. I was like, oh, we're having Sam on the show today. They're like, who? No, no, no, no. They were like, oh, that's cool. Like, that's good. Like, I understand what the show is, right? Then we have Dan Barcelo from T1 Energy talking about building solar panels in America, and of course our lightning round, which I'm very excited for. So, um, we've been thinking more about the Super Bowl that's coming up. We've been thinking more about ads, the response to the ads, the back and forth with the ads. Run had a good post here. He said, 'Putting on my media observer hat on. Anthropic ads are pretty brilliant because they're dishonest in a way that's only going to rage bait OpenAI heads and certain industry insiders, but are funny and striking to everyone else when you're a call option.' Calling them a call option. Kind of a diss.
Mario Kart blue shell. Are you familiar? Have you ever played Mario Kart? Do you understand? Not enough to get the... I played Mario Kart, I think, so long ago, blue shells didn't exist. But I've been playing with my kids and I've since learned the importance of the blue shell. The blue shell, it targets just the first player, just whoever's in first. It's a great metaphor for what's going on here. When you're not in first place, you get the blue shell. You can take a shot at the leader without even needing to call them out. So, you can just say the category is bad and everyone assumes you're talking about you know who. Uh, so I thought that was an interesting thing.
Trey says, 'Sam Altman, the connoisseur collector.' Yes. Yes. Yes. That one. Yes. Deep down. Anyway, so everybody had a take on this yesterday. Uh, it was perfect in how much kind of controversy it generated. It was wildly entertaining.
Uh, I wanted to kind of... I'll read through kind of like my updated take. Uh, I got a little bit of pushback. I said they were playing dirty. Signal responded to me and said, 'Not dirty at all.' So, I wanted to address that.
Okay, dig in. But first, let me tell you about Console. Console builds AI agents that automate 70% of IT, HR, finance, support, giving employees instant resolution for access requests.
Uh, anyway, so yeah, I was processing this more. We obviously watched some ads yesterday. We watched the Get a Mac campaign. We watched the Bud Light special delivery one, which is about Bud Light is in a castle. They get an order of corn syrup. They're like, 'We don't use corn syrup. That must be for Coors Light and other competitors.' Uh, and so like I was processing them and like the difference there is that those advertisements are truthful, right? Like people that have had a PC have probably gotten a virus, right? So when Mac is like riffing on that, it's truthful, right?
And they had some data to back it up. It's not deceitful. I would say if I'm putting on my steel-manning Microsoft in 2007 hat or 2003, I would say, 'Hey, we do have Windows Defender. We're fighting viruses.' And is it possible to get a virus on a Mac? Probably. Is it possible to not get a virus on a PC? Yes. And on Microsoft, people are like, 'Yeah, no one makes viruses for your computers because you don't sell very many. It's not very ROI positive.'
That's a good point. Um, anyways, and then Bud Light's campaign was truthful even though it was aggressive in that you could look up the ingredient list of their competitors and see that they did in fact use corn syrup, right? And you can make your own decision on whether or not you like that ingredient, but they were just drawing awareness to it.
And so my point is that I think that Anthropic's ads are closer to political attack ads and that they're sort of intentionally trying to be deceptive, right? They haven't broken any laws. They don't name ChatGPT. They're just sort of like throwing mud at the whole category. Uh, I asked... yeah, so anyways, I said they were playing dirty. I got some pushback on it. I asked Claude, I said, 'Claude, how would you define playing dirty?' And Claude said, 'Playing dirty generally means achieving your goals through tactics that are deceptive, unethical, or that violate the understood rules and norms of a given context, even if you're not technically illegal.' It's the gap between what you can do and what you should do. A few dimensions to it. It goes into deception, misleading others about your intentions, hiding information, or creating false impressions to gain an advantage. These campaigns do an amazing job creating a false perception or impression of what ads in LLMs are going to be like.
I do think the response, just to chime in with some random stuff, but I think the response to the ads, we were wondering like, you know, outside of the TBPN, we love ads. Ads are fine and they're not going to do anything weird. We're strong supporters. Uh, what will the public's reception be like? Will Claude skyrocket to the top of the charts because these ads are so effective? Will general consumers buy the line that yes, the chat apps are going to get weird with the ads or not? Um, and I was scrolling on Instagram Reels last night completely randomly. I was not looking for Anthropic content. I think I followed the Claude account. Maybe, maybe not. It just targeted me. It hits me with a vertical version of the ad. It's called Deception, I think. Something like that. No, Violation. Violation is the one. It's not called Deception. It'd be a little too on the nose. I think there is one called Deception. There's a bunch. They all have different names. Anyway, it's called Violation. And there's a screen grab here.
Oh, yeah. Yeah, I remember. So, people were like the Anthropic Deception ad is deceptive. Yeah. Yeah. So, Violation pops up and it's this... and it's a Claude AI and it has almost 6,000 likes. Even when it just got served to me, my interpretation was like this is working. This is popular. It's not just beautifully shot, it's well edited for vertical. Yeah. It's either really resonating or they're putting a massive amount of spend behind it or both.
This is what it looked like vertically. And then this is the funny thing. So, in this ad, you see the guy struggling to do a pull-up. He asks, you know, what's supposed to be an LLM, 'Create a fitness plan for me.' And then the fitness bot says, 'Hey, you know what else could help? Uh, one-inch insoles from Height Max or something like that.' And it's like this looks maxing. It's very funny. Uh, but it's a one-inch insert that would go in your shoes. I scroll up. What's the next ad that Meta serves me? An ad for three-inch inserts. Three-inch inserts. And the ad is actually deceptive. It says the guy can go from 5'9 to 6'1. That's four inches. And so these are full shoes that have the inserts built in. And for some reason I got in even though I'm not in the market for insoles. Um, the algorithm just knows that I love these ads because they're very funny and they're very on trend with the looks maxing thing. And so I get served these ads constantly. This is all Meta shows me is these height-enhancing shoes because I think I actually clicked on them and was like digging in.
So you bought them, right? Of course. Of course. Uh, to send to my enemies. You're trying to get to seven feet. You're trying to get to seven feet. That would be good. That would be good. Tyler, do you have something on this? Uh, yeah, I was just going to say, so I saw on Instagram as well. I saw some of the Claude ads and in the comments, I mean, people were riding with Claude. They are. Yeah. They're like... it was like normies, like totally, totally, like they're winning the vibe. They've been winning the vibe with developers and they've been winning the vibe war on X and now it feels like they're about to win the vibe war in the public square. Really quickly, let me tell you about Figma. Figma isn't your average vibe coding tool. It lives in Figma, so outputs look good, feel real, and stay connected to how teams build, create codeback prototypes and apps fast. Sorry.
Sorry, I said two things can be true about the campaign. It's brilliant, well-timed, and incredibly strategic for a few reasons I'll outline below. And it's designed to plant a false impression of ChatGPT's forthcoming ad product in the minds of hundreds of millions of Americans. Yep. They could argue, oh, we're not trying to do that. But you can't really kind of argue with the effect. So, Anthropic accomplishes a lot. The campaign entertains America, right? It's wildly entertaining. It's hilarious. Like, truly like the perfect like script that you can hear. You can hear the em-dashes, the pauses. It's amazing. Mother is the name of the agency that did it.
They absolutely crushed it. One, they're also putting themselves on the map ahead of the IPO. I think in some ways like certain audiences would know more about Anthropic than Claude, right? Even if you're just like generally interested in investing in AI, you're probably hearing about Anthropic more than you're hearing about Claude.
It builds their aura with insiders. If I said if they spend $100 million on this campaign and all it does is help retain a couple of like truly elite researchers, it's worth it. What are you laughing at?
I love it. Noah saying Anthropic is campaigning to get themselves banned just like with misinformation just going way too far into it backlashes. That's just funny to me. We can go back to that. It somewhat continues their like fear-based messaging that they've been kind of riding with in general back at the essays effectively rage-baits OpenAI. Like they got fully baited. Sam switched out of his like, you know, lowercase typing and was like, I got to go into uppercase for this one.
Increases like the public's like general scrutiny of the ads rollout and then Washington too is another factor.
Yeah, maybe. Yeah. I just think it'll come up.
Totally. I mean, you have to ask Mark Zuckerberg, what do you, how do you make money? Like anything's on the table these days in Washington.
Yeah. And so, and then the other thing is like it's going to broadly damage consumer trust in LLMs. Some people will just be like, 'Wait, like they've been kind of like making money on me without me knowing, right? Or how can I trust every output as like actually good advice or am I being monetized?' So yeah, and this is the one that you think could come back to bite them. The other five are potent potentially, right? But it depends what their consumer strategy is. Right now they're saying we don't care about consumer. They've said that a lot, but their actions kind of speak differently in some ways. So
I said Anthropic has consistently told the market they don't care about consumer, but I'm not sure. The argument for ads is that they'll make LLMs free for people that can't afford to pay a subscription, but Anthropic has already lost the race to serve billions of people, right? I don't think that they're, when you look at Gemini's sort of traction, OpenAI's chatbot traction, like it seems like the race to get to three billion monthly actives is kind of over. I don't believe that Claude's going to come behind and get there. And they wouldn't be able to do that without doing ads, right? Because there's like, you can kind of run the number, the blue shell take out in front.
So, the question that kind of where I was taking this is can they deliver a luxury product to a smaller cohort in the hundreds of millions to kind of iPhone numbers, right? There's like roughly like one and a half billion iPhones that are like active in the world. Those people could all buy a cheaper Android and just cheaper devices, but they've paid a premium for the iPhone because they can and for many people it delivers a better experience.
So I said the iPhone was not the first smartphone. Claude was not the first consumer LLM. The iPhone did differentiate on specs early, not unlike a model card, but Apple did eventually pivot to more emotional arguments for why you should be seen with an iPhone. It tells people you care about the environment, that you don't have adult apps flooding your app store, and that you take privacy seriously. These have had varying levels of success. Every tech company was able to tell an ESG story. And I can't imagine an Apple exec even saying the word porn today, even though Steve Jobs was very pointed about it back in 2010. He said, you know, there's a porn store for Android. You can download porn. Your kids can download porn. That's a place we don't want to go, so we're not going to go there. Calling out the competition by name and dropping that is like
Yeah. And the key thing here is that it was factual. It was true. It was true.
Like it wasn't deceptive.
And so I don't think that was edgy, but he wasn't playing dirty.
And so Apple did however carve out a solid messaging strategy around privacy and allowed them to put the screws to advertisers. The market adjusted and the ad industry obviously survived, but average Americans still feel that Facebook is listening to their conversations to target ads. Consumers deserve choice. It's great if they want to pay for ad-free tiers. Most don't.
1% in Europe for Facebook, by the way. That's the stat. They all have the option to pay for Facebook, for ad-free Facebook, and only 1%.
So consumers deserve choice, but they should not be misled about how ad platforms work. Android has generated an immense amount of value for the world. So has Google broadly. Let consumers choose, but let them choose intelligently.
Yes, I have a rebuttal, but I'm going to tell you about Lambda. Lambda is the super intelligence cloud building AI supercomputers for training and inference that scale from one GPU to hundreds of thousands. So my rebuttal, my steel man is that
And to be clear, I'm not saying Anthropic shouldn't have done this. I'm just saying that it was a little dirty.
Yeah, they're rolling around in the mud. It's good. They're in the trough. We love the trough. We live in the trough. We live for the trough. My steel man is that they didn't cross the line. They didn't play dirty because they didn't call out ChatGPT directly. Okay, you can take that whatever you want. But will
The other thing is they're punching up. They are punching up. Yes. But third, there is a world where something like this will happen. There is a world where the ads do get integrated in such a seamless way. If you look at the evolution of Google's 10 blue links, it started with 10 blue links, no ads. Then it was a very clear yellow box with ad and it was very clear that it was an ad. And over time the UI evolved to be a lot less aggressive about telling you that it's an ad. And the ads on Meta platforms do get creepy sometimes. You talk about something and then you see the ad and maybe that's just confirmation bias or some sort of cognitive. You only notice the ones that are weird. So they all feel weird. You see a lot of stuff that you weren't talking about. That doesn't trigger anything. But when you see the thing that you were just talking to your friend about, I was just talking to you about sweaters and I see an ad for a sweater. I'm like, how did it know? And realistically, it knows because you just went on Facebook, you found that sweater, you bought it, it knows that we're friends, we're DMing, we're talking, we're literally friends on the platform. And so it's like, well, if Jordi likes this and they're hanging out all the time, sending each other DMs, why don't I just show John what I just sold to Jordi? That makes perfect sense. That's something that can be done with just stock vanilla machine learning at, you know, core AI inside Facebook and Meta and they do that very effectively, but it can feel sort of creepy sometimes and some people get creeped out by it and they talk about it. And so the idea that an interaction like that might happen is not complete science fiction. It is possible. And so they are sort of warning that hey, if you want to make sure that this never happens, it's our pledge that that won't, that's not even on the table.
Now the big question is when's Anthropic launching ads?
We got to get them to launch ads.
Well, I don't think they can now.
No, they have to. It's okay. I give them permission. I will say if everyone's going to be dunking, oh, you went back on your promise. No, I will be your strongest soldier. I will be your stronger soldier and I will say it's the right thing to do. Put ads in Claude Code. Put ads in Claude. Put ads in the comments of the code that you write. So if someone reviews the code, they're seeing an ad. That's what I want. That's the future I want to live for. Anyway, TurboPuffer, serverless vector and full-text search built from first principles and object storage. Fast, 10x cheaper, and extremely scalable. Oh, and you know I got to tell you more about ads. Vibe, where DTC brands, B2B startups, and AI companies advertise on streaming TV. Pick channels, target audiences, and measure sales. Just like on Meta, we love ads here, and we love doing ads for ads.
Zach Kukov says, 'Every time I see Anthropic and OpenAI try to distinguish themselves with comms marketing, I realize how much we are replaying the PC wars from the '90s. Anthropic: tasteful, elegant, opinionated, prosumer, expert, enterprise. OpenAI: populist, broadly appealing, low consumer plus typical enterprise.'
Yeah. So, Anthropic is the Apple and OpenAI is the Microsoft and they're also aligned with Microsoft or owned in part by Microsoft. Yes.
Yeah. I mean, it would be interesting to hear Dario just talk for an hour purely about just the risks of advertising and AI.
Right. Because that would be powerful. Certainly wouldn't have been as effective as dragging OpenAI in front of hundreds of millions of people.
Yeah. But I mean, doesn't Anthropic have a podcast? Anthropic podcast. I think they do.
TJ was helping Sam with some comms. He said Anthropic might think more seriously, he said, fix it for you. Anthropic might think more seriously
About adding ads if they had any consumers using their product.
Just taking a very tactful response to a situation. Just being like, 'What if you amped it up, brother?'
Yeah. I mean, you've already lost if you're just dropping a massive word salad.
Maybe. I don't know. I think there's a lot of nuance here and it's good. I do think it's important to not mold, to not be, you know, like angry and made mad and let people spike your cortisol.
Exactly. Like they're going to be gesture maxing at the Super Bowl and you can't let it affect you. You got to focus on the new models which have launched today. There's a bunch of new stuff. So GPT 5.3 codec launched today and there's a very very cool update to or new product Frontier, which is a product for building AI co-workers that I'm very excited to talk to Sam about because it feels like the first glimpse of an orchestration product which we talked about yesterday. Gas Town has been taking off. Let's give some credit to John.
Yeah. No, it might seem like John had early insight on today's OpenAI launch. We did not.
We didn't. No, we actually didn't.
But in hindsight, you called it perfectly.
One day of alpha. Yeah. If you had launched an enterprise focused orchestration platform yesterday.
Yes. You could have raised money, got sold secondary.
All right before OpenAI immediately launched a very...
Are you going to steal from OpenAI? They're not even thinking about this. They launch it the next day.
Who knows? Who knows? Anyway, let's move on to Shopify. Shopify is the commerce platform that grows with your business and lets you sell in seconds online, in store, on mobile, on social, on marketplaces, and now with AI agents.
Matt Turk says, 'Regular startup. We closed a few customers and shipped some new features. Good week so far.' Anthropic. We destroyed our main rival with our Super Bowl ads and tanked the entire software category in public markets by announcing some plugins. Good week so far.
Oh yeah, we barely even talked about this, but Anthropic launched a lawyer in your pocket. They launched a legal tool or they announced it. I don't know. Is it actually available in the app yet? Because it feels like, okay, maybe this competes with Harvey. I'm not seeing it yet in the Claude app. But there...
I don't really think it competes with Harvey.
No, no, no. So I don't think it does because they're selling it direct to consumer, I believe at least. But it seems like an amazing product. Like it seems like the demand for this would be incredible. So many...
It will compete with LegalZoom. Like LegalZoom is down 15% since this announcement.
It's now a 1.38 billion company.
Yeah. Sort of in Chegg mode. Yeah, and not just for... I mean LegalZoom is a little bit different because they can actually file your incorporation documents and they've faced pressure from Stripe Atlas for a long time on that, LLCs and whatnot. But I mean truly, like if you're getting a job and your employer gives you an offer letter, like taking that to a lawyer can be really expensive. If it's your first job you're probably not going to review it. A lot of people are probably just copy pasting it into Claude, asking their parents, 'Take a look at this.' A little bit of that. And then they can... but just being able to just forward the email in or integrate your Gmail and just say, 'Hey, I got this offer letter. Like does anything in here look weird? Is there anything I should ask about? I don't have a ton of leverage but I want to understand this document.' Claude should be able to do that and it makes a lot of sense and I expect OpenAI to launch this product like ASAP anyway.
Yeah. I mean, so I think all that was actually launched for the Claude legal thing was just a plugin in Claude Code,
Which means it's basically just like MD files, right? It's like skills. It's like skills,
But it feels like it goes beyond skills because they probably had to do some legal work of their own to make sure that they're like not giving legal advice and they're couching things properly. I think that's also why ChatGPT Health is in a different area because they don't want any risk of like pre-training on your testosterone levels. And then I go into the next version of ChatGPT and I say, 'What are Tyler Cowen's testosterone levels?' And it just knows it because it learned it from the chats that you were sending it. There's a whole bunch of private information beyond testosterone levels, of course. Anyway, Railway. Railway is the all-in-one intelligent cloud provider. Use your favorite agents to deploy web apps, servers, databases, and more while Railway takes care of scaling, monitoring, and security automatically. Kev over on X said, 'Uh, tbh, the Anthropic ads are good, but I think they're a bad idea. Normies are not going to think, wow, this is what ChatGPT is going to be like. I better subscribe to Claude.com.' They're going to think, 'Wow, this is what AI is going to be like.'
Rune agrees. He's not biased at all. He says, 'Suicide bombing strategy. It's bad for them, but worse for OpenAI. You almost have to respect it.'
What was the text I sent you? I said, quote, 'I don't know how to get on Anthropic.'
Yeah. Yeah. John was asking a friend that's outside of tech and they were like, 'What's Anthropic? How do I get on it? I don't know how.'
Yeah. Let's go to Eric Sefcovic. He said this Anthropic ad is simply obnoxious. And then we'll move on to other stories. Anthropic generates most of its revenue from enterprise business so it can afford not to maximize its consumer revenue opportunity. But the sanctimonious moralizing here presents advertising as a cynical business model choice. It's neither cynical nor a choice. The premium digital advertising supported model is the only repeatedly proven pathway for a consumer technology product to reach humanity at scale. There are just folks who can't pay $20 a month. OpenAI's advertising revenue will enable it to offer access to larger models for free tier users. To the extent Anthropic thinks chatbots powered by frontier models are valuable to society, it should aspire to provide access to as many people as possible. Now, they'd probably say, 'Hey, we're premium. We do give access to chatbots powered by frontier models to free users. They just only get a certain amount and then they have to upgrade.' But Eric's point still holds here for sure. Ads allow for that to an extent that no other consumer technology business model demonstrably can. This kind of condescending attitude toward digital advertising is economic chauvinism and it represents a desire for technological gatekeeping and the idea that ads will necessarily influence chatbot content is overstated, borderline anachronistic. Anachronistic because you don't do it that way anymore. You wait until, hey, this person's buying, you know, insoles or lifts and they're asking about the Roman Empire and this is the perfect time because they're just chilling reading the Roman Empire and they're thinking, 'Oh, okay. Yeah, I did need to buy that.' And then they switch gears. Like, you don't need to put the ad right next to the content that relates to it. That's just an antiquated way of thinking about online advertising. It ignores functional... Yeah. Again, I think Sam Altman did say in October of 2024, 'I kind of think of ads as a last resort for us as a business model.' So again, he's going to have to eat his words on this one.
They get to have a little bit of a victory lap there.
Yeah. So in that sense, criticism is fair.
Anyway, before we move on, CrowdStrike. Your business is AI, their business is securing it. CrowdStrike secures AI and stops breaches. So Claude also announced a new model from Anthropic, of course, introducing Claude Opus 4.6. We have Sholto coming on the show at 12:30 to discuss that. They say it's the smartest model and it got an upgrade. Opus 4.6 plans more carefully, sustains agentic tasks for longer, operates reliably in massive code bases and catches its own mistakes. It's also our first Opus class model with a 1 million token context window in beta. So very exciting. They put a number of... they put the model card together. They did particularly well. What was the benchmark that stuck out to you, Tyler? You said...
Yeah, I think it was ARC-AGI-2. It's now at like 69%. I think previous was...
I believe it was 5.2 which was at I want to say 55 around.
So like pretty sizable upgrade. Yeah.
But I mean with these things it's always so hard to tell. Like it's really just like qualitative differences at this point where the models are like you have to use it for like an hour or two and then you can kind of tell what the differences are.
Yeah. And also like in new contexts, like in terms of just the vanilla go and ask a question, it's been able to get you a pretty good answer for like years now. But it hasn't been able to go and pull a bunch of financial data together, fact check it all, put it in an Excel sheet. This is what they're pushing and this is what Coatue Investments is so excited about. Ooh, says Coatue. Anthropic updates AI model to field complex financial research. So Opus 4.6 is designed to carry out financial research and other work-related functions. The company's expansion into new areas, including legal service, has rattled Wall Street and sparked concerns about which companies and services will be disrupted by AI. The SaaS apocalypse is upon us. We will be asking Sam Altman, is software dead or...
Are we back? Who knows? Claude 4.6 Opus is still best. It still has the best SVG results out of all the models. Just incredibly high taste, says Lisan Al-Gaib. And these are some pretty beautiful pictures. Are you familiar with SVG programming? So you basically draw each square and line in code. So you have to say, I want a square that goes from this pixel to this pixel to this pixel to this pixel. And you add...
and layer all those up. It can be extremely time-consuming if you do it by hand. But Claude can just sort of one-shot it and it looks pretty on brand.
It's an interesting benchmark because this isn't something you necessarily ask. These are not generative images in the diffusion sense. This isn't the Midjourney model. This isn't what Sora uses. This is a different thing, asking the model to basically write code that generates an image. So it really has to understand the back and forth between what it's building. So cool benchmark. I like that. I enjoyed that.
I also like the New York Stock Exchange. Do you want to change the world? Raise capital at the New York Stock Exchange. Then it's as easy as that. Anyway, the Anthropic handle being owned by a guy who only posts his favorite form of AI safety: Wordles. He only posts Wordles. Oh, and he just has @anthropic. That's very funny from Paul Graham.
So this is a strange thing with X. They have policies around their handles that they really stand by in that Ramp is not owned by Ramp. @anthropic is not owned by Anthropic. And when you look at the accounts, you think, is this really the best use of a handle like this in comparison?
Paul Graham, who is at Anthropic, is emphatically not an AI company. Ohioan, liberal, bookworm, news hound, celeb.
If he's a bookworm, he probably loves sitting on the Anthropic handle given that Anthropic has been worming their way through some books, chewing through some books, much like a worm.
Yes, we didn't get to that article. That was a funny, funny article. You got to feed Claude. It's okay. The books got to feed, you know, and they have to... Yeah, potentially bigger than the water thing. Like, it's just so visceral. People love books and watching them be destroyed.
Yeah, if you didn't see... destroyed book, yes, I agree. But people don't like destroying books. They're sensitive about that. The very funny thing is that Paul on Twitter, on X, has migrated to Bluesky for many things, but he didn't get the Anthropic handle on Bluesky. He's anthropic42 over there. So shoot him a follow if you're hanging out looking for Wordle progress.
People are playing around with the nominative determinism. Yes. Say Google's CEO is named Pichai. His purpose is to pitch AI. Dario's last name is literally AI model.
Maybe I should change my name to pitch MongoDB. John, pitch MongoDB. Choose a database built for flexibility and scale with best-in-class embedding models and rerankers. MongoDB has what you need to build what's next.
Dylan Patel is sharing some data. He says 4% of GitHub public commits are being authored by Claude Code right now. At the current trajectory, we believe that Claude Code will be 20% of all daily commits by the end of 2026.
That feels low to me. I was expecting 99%. Sorry.
Well, I mean, so I think the way you can see this is because whenever you actually do a commit with Claude Code, it adds itself. So there's a ton of people that are using Claude Code but then not attributing back to Claude.
Yes. And we also have some news. Dylan Patel from SemiAnalysis will be joining the show tomorrow to break down Claude Code. We might have to pull him out of his psychosis. He is addicted to Claude Code. He's running multiple instances at all times and he has updated his profile picture in the SemiAnalysis Slack to reflect his devotion to the Shog, I suppose.
Anyway, so OpenAI unveils Frontier, a product for building AI co-workers. This is in the Wall Street Journal and OpenAI also posted about it. The new platform, launched amid market fears over AI's disruption to software, is aimed at helping businesses develop AI agents that work alongside humans. And so there are some interesting questions here about...
Think of these as like AI co-workers that are actively trying to take your job.
Like they're trying to help, but they also want your title. They want your comp. Yes. They want to learn everything about what makes you great and they want...
Maybe, or maybe they're just trying to empower you, Jordi. Maybe they're just trying to make you a better you. I don't know. Maybe they're just coming over, just cracking jokes, trying to distract you while learning so that they can...
They can literally do both. They can literally do both. Before we dive into this, let me tell you about Restream. One live stream, 30 plus destinations. If you want to multistream, go to restream.com.
So, Frontier works with OpenAI's previously announced AI agent building tools and makes it easier for businesses to combine sources of data that agents need to perform tasks. The AI company said the agents will be able to process information from various sources and complete tasks like working with files and running code. OpenAI said so no more copy-paste everything into your ChatGPT Enterprise edition. It should have access to your network, plug into all your different systems. You'll be able to write API bindings, I imagine, and there might be some forward-deployed engineers or some associates from OpenAI that are helping you actually onboard fully to the agentic workflows that have been promised.
Yeah, they're hiring how many consultants to help with this go-to-market?
They are hiring hundreds of AI consultants to boost enterprise sales. This is information... job creation.
Yeah, this is a great gig. Like, I would highly recommend jumping on this if you're in this market. OpenAI is hiring hundreds of new staffers to expand a technical consulting team that helps large corporations develop custom AI applications and agents to automate employee tasks. According to a person knowledgeable with the company's plans, the hiring effort could help it beat back competition from arch-rival Anthropic, which has also upped its game in catering to enterprises. It comes as OpenAI prepares to launch a new enterprise offering that would unify businesses' efforts to use AI. The chatbot maker is expanding its number of technical consultants, also known as forward-deployed engineers, who can customize OpenAI's model using a client's own data. A person said these engineers can, for example, help T-Mobile develop AI to respond to customer service requests or help Intuit provide its customers with tax preparation services.
So you have all this data, you want to do long-context reinforcement learning on it. Long-context reinforcement learning has been very, very successful in the coding world because Git has a complete history of every line of code that's been written, every comment, why it happened. You have this perfect record of everything that happened when you built a piece of software and there's a ton of open-source repositories. You can download all of GitHub basically and see how software is developed, so you can train the model on that. Well, you can't really get the same level of free data with enterprises. A lot of this data is locked up and a lot of it's specific to a specific business process.
Yeah. Let's pull up an image. There's a little graphic here that they made from OpenAI's Frontier, openai.com/index/introducing.
Yeah. And if you can zoom in a little bit, at the bottom you have your system of record and you have business context, agent execution, evaluation and optimization. Your agents, OpenAI agents, third-party agents. That feels significant, right? They want to be the orchestrator, right? And if you want to bring in some other folks in to help out, great. At least for now.
And then they have interfaces, which are ChatGPT Enterprise, OpenAI Atlas, and other business applications.
I'm a little upset that they didn't go with a Mad Max theme like Gas Town. I like the Polecats. I like the Mayor. I like the Deacon. I thought that was a fun metaphor. They went with something a little more enterprisey, but I think Frontier is a good name. I don't know, it flows, of course. It's not Frontier 5.2 High, so we got to count our blessings here. But I think Frontier works, you know. Are you working on a Frontier integration project? Oh, those consultants from OpenAI, they're getting us onboarded to Frontier. It sounds good. Why is Tyler laughing?
Yeah, I'm getting flamed in the comments. What's up with Tyler?
Yeah. What happened? Did you use shampoo or conditioner?
I need a haircut. Maybe you need a hat. Maybe you can grab one of those TDP hats over there. You'll be good. Don't worry. Let me tell you about AppLovin. Profitable advertising made easy with Axon AI. Get access to over 1 billion daily active users and grow your business today.
We'll get more from Sam on Frontier.
Yes, I think we can move on.
So the last thing on 5.2, the last benchmark that's interesting is GPT-5.2 with high, not extra-high, reasoning effort has a 50% time horizon of around 6.6 hours on our expanded suite of software tasks. This is the highest estimate for time horizon measurement we have reported to date and it's right on track, doubling every 128 days. It depends on what you're looking for, but the implications of this are, you know, more and more tasks are suitable for these models, although I do think we're going through a shift with the orchestration thing where we might need a new benchmark because what happens if the individual model, GPT-5.2 High, can do one six-hour task basically, but you can deploy five that talk to each other and when you combine those it adds up to 20 hours of work. You just get a jump in the graph, but it's not technically the model. So you need sort of a new benchmark, supposedly.
Jessica Lessin likes Sam Altman's post. She says, 'This is excellent comps from Sam Altman. An eye-popping stat in an easy-to-understand claim. And I hate though this budding movement to frame the business model debate as elitism versus anti-elitism when chatbots launched with subscription business models. I thought it would be a great new era of consumer tech blended with ads and subs, much as we've seen with Netflix and Spotify. It would be a shame if the OpenAI versus Anthropic rivalry wipes that away and makes it seem like...
Subscriptions are only the domain of the elite. So many other consumer businesses show otherwise, and they are better off for the balance. I like that. Signal says either you or I have lost the ability to process reality.
Yeah, I mean, not in response to this. In response to that. Okay. Yeah. No, a lot of people didn't think this was great comms at all. Like, I mean, I would say like most of the posts I saw were like, 'Hey, like this feels like you got baited.'
Yeah. Yeah. Because you could have just not responded, I guess. Yeah. Or you could have said he said later in the day, 'I'm excited to share that Codex has 1 million users.' Yeah. Which product they launched this week, right? Yeah. Yeah. Just focus on yourself, maybe. I don't know.
Meanwhile, over in Silicon... Greg says GB200 has really been enabling us to do some amazing things.
Yeah, it is remarkable how, you know, relatively slow the rollout has been. Obviously it's been very quick, but it does take time from the time that Nvidia announces these chips to actually fabbing them, racking them, wiring them up, getting ready to train new models. And so, anything? Yeah, I mean, I think Dean Ball had a new post today and he said like, you know, companies don't even have Blackwells yet. Like, so we haven't seen, there's no models that have been trained on Blackwells.
Yeah, pretty. So it's like it's going to get ready to see some AI progress, but we got Mike Isaac in the chat. He says not good comms. Not good. I'm with Mike. Expand, Mike. Expand. I'm with Mike on that one. We'll take, we'll see. Let me, I'll look up Mike's.
And I will tell you about Google earnings because we touched on this, but it's important in the context. So, Alphabet sales hit record, spending to double. They're going all in on AI. Google parent Alphabet reported an 18% jump in fourth quarter revenue driven by growth in digital advertising.
Before we go into this, I did, Mike had a banger yesterday. He said, 'By the number of OpenAI employees I see tweeting about Anthropic's Super Bowl ad, Anthropic should be paying OpenAI earned media fees.' Yeah, that was a good take. Got him. That was a good take. Got him.
Sales reached nearly $114 billion ahead of analyst expectations. Net income $34.5 billion, a 30% increase compared with the period a year earlier. The company reported a record $43 billion in sales for 2025. Profit $132 billion. Not too bad. Alphabet shares were sort of all over the place. The Wall Street Journal printed that they went down 1% in after-hours trading, but there's a lot going on. Google, like other technology companies, plans to spend tens of billions of dollars to develop AI models and build the data centers needed to train and run them.
The company said it expected to spend between $175 billion and $185 billion in capex in 2026, up from $91 billion to $93 billion in 2025. So, they're like all doubling, which is, that's exponential growth, baby. Get ready for some AI progress.
Yeah. Somebody, I don't have it pulled up, but somebody was saying that the 2026 projected capex will be more than the lifetime capex for Google up to 2021. So, in a single year, they're going to eclipse that, which is just insane.
Buco had a good take. He said, 'Google capex on purpose. Tell the market this is what it will take to defeat us before IPOs hit.' Yeah, certainly nerve-wracking if you're competing with them, but they have the edge on the capital side in the capital war over Anthropic and OpenAI, but the race is still real.
Yeah, let me tell you about Phantom Cash. Fund your wallet without exchanges or middlemen and spend with the Phantom Card. The Wall Street Journal had some more context on Google. Google leans hard into its AI winter status. AI winter status, not AI winter. Ad and cloud growth acceleration justify the recent surge in Alphabet stock, but blowout capex forecast still takes one's breath away. The motto for the artificial intelligence race today should be: if you've got it, spend it.
That's a message that Meta Platforms took to heart during its fourth quarter report last week when the Facebook and Instagram parent announced plans to spend up to $135 billion on capex compared to about $72 billion last year. Google managed to up the ante Wednesday with its own plan to spend as much as $185 billion this year, which would be about double last year's outlay. Google's annual revenue has now topped $400 billion, about twice as large as Meta's. Still, that new spending target, even for a company that has been firing on all cylinders lately, takes one's breath away.
The stock price slipped in after-hours trading after its fourth quarter report and conference call. Google has both the political and financial capital to lay such a bet. The company's Gemini 3 model has put it on top of a heap of performance for AI models. While the unmatched distribution of its search engine and products like Gmail have quickly driven adoption, Google said Wednesday that it has more than 750 million monthly active users just on its Gemini app, which only represents a portion of Gemini's actual users because it's vendored into all the different products. What's going on in the chat?
40 Army has entered the chat. Interesting. They're hitting the chat with #keep40. Okay. They want to be heard by Sam Altman. We'll be... It's not keep 40. Hasn't 40 been deprecated? And so, shouldn't they update the hashtag to say like bring back 40? Revive 40. Revive 40. It's actually not leaving till February 13th.
Oh, okay. So, there is time to reverse the decision. I am interested to know about the operating cost of keeping 40 alive because it feels like if it rolls, clearly something people were willing to pay for. Yeah. Can you just kind of leave it in a corner and it just kind of runs? It was, you know, maybe more of like a user behavior decision than a financial one. Because it seems like running legacy models gets really costly. Okay.
The strong reception of Gemini along with Google's victory over the federal government's efforts to break up the company have cheered investors when sentiment on technology and AI is faltering. Alphabet stock price has jumped around 20% in the last three months. Nvidia, Microsoft, Amazon.com, and Broadcom have all lost ground during that time. But capital matters, too. And here is where Google's business model pays off the most. The company's advertising arm has long been a lucrative cash cow that is still managing double-digit growth rates. Growth has actually accelerated with ad revenue up 14% in the fourth quarter compared to 13% in the previous one. It's exactly what we saw at Meta. The ads are getting better, baby.
The company's cloud computing division was even more impressive with revenue growth jumping 48% year-over-year to hit $17.8 billion. Google Cloud hasn't seen growth like that since early 2021 when the business was still less than a third of its current size. Google Cloud turned in a record $5.3 billion in operating profit in the last quarter, a figure that was 45% higher than Wall Street's targets. The company's booming business produced nearly $165 billion in operating cash flow in 2025, the highest in the S&P 500. These strong results will help investors digest the latest investment plan. But spending what could be 40% of annual revenue on AI chips and related infrastructure is still a sizable gamble. Such investments will sharply elevate depreciation charges which in the latest quarter reduced net income by 18%. Google isn't the bargain it was less than a year ago when breakup fears had the stock trading at less than 16 times projected earnings.
My question is like, when does the AI buildout stop? Because if you're constantly investing more and more, where does the cash flow come from? Like as an analyst, you look at this and you say, okay, they're spending, you know, 50% of their revenue or 40% of their revenue on the AI buildout on data centers, but how long will they need to do this? Like if they have to do this forever, then you just permanently have a worse business because you're just constantly buying hardware.
Yeah. I mean, you get a massive capability increase, lots of labor moves into data centers and eventually the revenue spike. Yeah. You can enter a scenario where it makes sense to continue to increase capex because revenue is accelerating even faster. Yeah. Yeah.
Notable that Nvidia is still, you know, trading down today even after that update on earnings. Like you would expect people to look and see like, hey, you know, Jensen had given some pretty kind of wild projections, right, looking out over the next couple years. And you have to... updating here makes it a lot more... spend a lot on TPU, but still they will be buying a lot of Nvidia chips for sure.
Really quickly, let me tell you about Plaid. Plaid powers the apps you use to spend, save, borrow, and invest securely, connecting bank accounts to move money, fight fraud, and improve lending now with AI. Richard says Google is a company that doesn't do hype. For them to go and increase capex from $90 billion to $180 billion is probably the most bullish thing long-term investors can see as it shows the scale of future revenue growth.
I'm shocked that at this stage most still don't understand this. Yeah. So excited about it. What did Doug say over at Fabricated Knowledge? He said Google made $163 billion in CFO last year and is now planning to spend $180 billion in capex next year. Now, I think acceleration is coming, but wowza, I'd rather go bankrupt than lose the race. Quoting Sergey, the trailing 12 months of free cash flow is staggering over there. Going to make $200 billion in cash flow this year. Still less than capex for now. Also going to have $80 billion in net cash and probably $100 billion in myriad minority illiquid equity stakes. So yeah, they own like SpaceX and a bunch of... Waymo. With current rates of growth, they can raise $300 billion in 2030 without even going into debt. Totally agree, but this feels pretty close to the pedal to the metal to me, says Doug.
So, Joe Weisenthal says, 'The average person on Earth is watching 25 YouTube Shorts every single day. YouTube Shorts average 200 billion daily views.' That's a lot. That's just insane. That's a lot. I mean, you can watch a YouTube Short in like what, 5 seconds on average, right? Because you skip one, you watch one for 5 seconds, one for 15 seconds. So, I mean, 25 YouTube Shorts, that feels like maybe 5 or 10 minutes of browsing, but still like pretty staggering.
Yeah, I haven't migrated any... I don't scroll on YouTube Shorts yet. I find if I search for something, like let's say I'm looking up a car, it's nice to get a 60-second if I don't have time to watch a review, but I'm not like sitting there scrolling. Yeah, it still serves clearly a lot of people, but I rarely scroll through, whereas on Instagram I will scroll the feed of reels.
Yeah. Staying in the content world, Khi shared that just in YouTube generated over $60 billion last year, more than Netflix. And Polymouse says, 'YouTube is beating Netflix with this really sneaky content strategy in which their creators make stuff people want to see and are then rewarded for it with views and money.' And it is a simple encapsulation of the YouTube... hates this one... one weird simple trick.
UGC is a big, big business. Who would have thought? But yeah, the scale of the quality. I'm hoping we're going to talk to MrBeast about his outlook on YouTube and how it's changing soon because he's taken it in such an incredible direction where the content is Netflix quality. I mean, he sells it to Amazon Prime, right? So, he's certainly there.
In the antitrust division of the Department of Justice posted an update today. The DOJ antitrust division filed notice that it will cross-appeal from the remedies decision in its case against Google's unlawful monopolization of internet search and search advertising. So they're still hashing that out. Let me tell you about Labelbox. Reinforcement learning environments, voice, robotics, evals, and expert human data. Labelbox is the data factory behind the world's leading AI teams.
So big tech is throwing cash into India in its quest for AI supremacy. We talked about this story a little bit yesterday. We couldn't get to the bottom of... Bloomberg. Yeah. Bloomberg had a story that... it's about the facts. Yeah. Effectively Google's scaling square footage. Yes. In India pretty dramatically. Yes. And the question was, you know, why are they scaling it up at all? They're investing so much in capex and servers and digital workers. But of course they need humans as well.
So, American tech behemoths are racing to establish leadership in artificial intelligence, not just in the US, but also around the world. India is welcoming them with open arms, says the Wall Street Journal. India is becoming one of the hottest markets globally for US AI titans looking to cater to the country's massive and digitally savvy population. Looking to attract more tech investments, the Indian government announced plans over the weekend to give tech firms a 20-year tax break on overseas revenue gleaned from global data services based in India.
So you go and you set up a data center or you set up a business, a tech business, you're making money off of that whole, you know, data center or whatever you're selling in terms of software and you don't have to pay taxes to India for 20 years. That's a long, long time. Techno Chief called this out yesterday.
Yeah. The move is part of the Indian government's push to make the country a major provider of AI services, including low-cost tools to solve local problems while leaving cutting-edge innovation to deep-pocketed firms in the US and China. 'This will give India the opportunity to become a major AI hub,' said India's technology minister. In the past few months, US tech companies have unveiled tens of billions of dollars in investment in Indian data centers as they race to build AI infrastructure around the world.
In October, Google announced a $15 billion investment in data centers in southeastern India as well as undersea cable links in what the company described as its largest single AI hub outside the United States. In December, Microsoft unveiled its largest ever investment in Asia with a $17.5 billion pledge to develop the country's cloud and AI infrastructure. Amazon also pledged $35 billion across its operations in India up until 2030. The big US data center firms known as hyperscalers are drawn to a country whose 1.4 billion consumers are some of the most prolific users of data and AI chatbots. Very interesting.
Anyway, there's more news from Nvidia in The Information. Nvidia is delaying a new gaming chip due to memory shortage. This is the line you don't want to cross. The gamers are going to rise up and they're going to storm the data centers and they will be putting the screws. They'll be calling their representatives, calling their senators. Pause AI has a new cohort. If you don't give me a new gaming graphics card to play the latest games... it's... Yeah. So it's the first time in 30 years it won't release one in the calendar year.
Yeah. Every year the... it's over. 1080 Ti, 2080, 3080. These were great graphics cards and now no longer, just delaying it. I mean, I guess you can use cloud gaming or something. But also it does feel like a lot of video games have sort of maxed out what they can do with stock graphics cards and so there isn't as much need to upgrade every year. What do you think?
Yeah. I mean, also with Genie 3, it's like we're getting new advancements in gaming. Yes. Yes. Yes. So, they should be happy about this. You'll own nothing and be happy. How much Genie 3 did you play yesterday, Tyler? It's fun, right? It's fun. So, you play it. I thought about playing it. Oh, okay. You thought about playing it, but the mechanics weren't there, right? You need Genie 4. I'm going to beat DA. That's the goalpost currently. I want mechanics. I want tire changes, gas refill as I'm driving the Nürburgring in my Genie 3 car simulator, race simulator.
Anyway, memory chips are a key component of GPUs. Very excited to ask Sam Altman about the various pieces of the AI bottleneck since... we should ask him about world models, too. That would be interesting. I, you know, sure he's not one to not launch. Yeah, he's launched a lot of stuff.
And yeah, Ben Thompson has had this take for a while that AI-generated video is the metaverse and that's what you're seeing with Meta Vibes. That's what you're seeing with Sora. This is sort of where things go. Nvidia is also slashing production of its current line of gaming chips, the GeForce RTX 50 GPUs, because of the memory shortage. One of the people said prices of Nvidia's latest gaming GPUs have already risen at retail stores and websites due to their scarcity over the past year.
I wonder how much it costs to actually build a modern gaming PC these days because it's been a while. It used to be like a few thousand dollars would get you sort of like a top-tier setup. And I wonder what it's looking like now with all the shortages. It's possible to be sure that Nvidia executives could still change their mind and release a gaming chip if the market improves as the company is known for being flexible and moving quickly, says The Information.
Demand for computer memory chips has skyrocketed due to the AI boom as they are needed in large quantities to train and operate machine learning models. Memory chips act as warehouses for storing data. The memory chip shortage is expected to lead to higher prices in consumer electronics. Last week, Apple CEO Tim Cook said the rising prices of memory chips would have an impact on the company's March quarter margins. He's also saying that he's having trouble getting line time at TSMC to some extent. And so there's a lot of bottlenecks that are working their way through all the technology markets.
He hinted that the impact would be greater in the future, noting, 'We do continue to see market pricing for memory increasing significantly. As always, we'll look at a range of options to deal with that.' Gaming and AI chips use different types of memory, but both are made of the same raw materials coming from one of three main suppliers: Samsung Electronics, SK Hynix, and Micron Technologies.
Well, without our first guest of the show, Sam Altman, the CEO of OpenAI. He's in the Remo waiting room. Welcome to the show, Sam. How are you doing? Welcome back. Good. Thank you guys for having me back. Thanks so much. Big day. Big day. Kick us off with... Where should we start? Should we start with the model or Frontier?
Can we start with the model just because I'm having fun with it? Yeah, absolutely. Break it down. What'd you launch? So we launched 5.3 Codex. It is, I think, the best coding model in the world. We took a lot of the feedback that people had about 5.2 Codex and got it into one model. It is much smarter at programming, but it's also way faster. You can interact with it mid-turn. I think it's got a much better personality.
It's really good at computer use. So it feels like a very big step forward. It was funny as we were deploying it this morning, a couple of very, like, extremely experts at using these models noticed and said, 'Man, something's really different with Codex.' And they caught it mid-deploy. So I think you can really feel it quickly. That's great. Oh, you're saying people outside of OpenAI, just everyday use. Yeah. Like the start of an hour that we put it out before we, you know... talk about interacting with a mid-turn. How does that work? Why is that important? What does that unlock?
So people are starting to use these tools for very long pieces of work at one time, you know, multi-hour tasks. And sometimes you don't specify it correctly, right? Sometimes something's not set up right. Something just screws up. The ability... they can do amazing things with no steering, but they can do much more amazing things if you steer them along the way. So this has been... this is one of the things I felt most new about this model.
So talk about orchestration, how this fits into Frontier because I imagine... one second, it's notable like if you see a coworker making a mistake and you don't interrupt them, that's rude, right? Like it's deeply inefficient. It is incredible what these models can do without any feedback. Like if you think about a new coworker, especially, you know, you train them and you give them a lot of feedback early on and they learn the job and you correct them and they kind of get practice. In the models, they will soon do that, but right now they don't do that. So we just rely on either they get it right one shot or we collect them, correct them along the way.
Yeah. So, you know, I think there's a lot of people that are running multiple agents and multiple tabs. They're starting to think about orchestration. Feels like Frontier is a piece of that. But if you're interacting with a model that's running mid-turn, like how does the user experience change for developers with 5.3? And then what will it look like in the Frontier world? I think we will be heading towards a workflow where a lot of people just feel like they're managing a team of agents and they'll keep, as the agents get better, they'll keep operating at a higher and higher level of abstraction, which at least watching what's happening so far is a jump that people are going to make pretty well.
The models are so good now. There's such a capability overhang that building better tools to let people do that, which the Codex app that we launched on Monday was a great step forward for, will be very, very important. But you will be managing very complex workflows. The agents will keep getting better. So you'll keep working at the maximum of your management bandwidth or cognitive ability to keep track of all the stuff. And the tools to make that easy to do will matter, I think, more than intelligence for a little while because there's such an intelligence overhang already.
What's the role of a forward-deployed engineer today at OpenAI towards the end of the year? Capability overhang that feels like raw meat for a forward-deployed engineer. It's like they solve that problem, right? Yeah. Yeah, I mean, look, eventually the models will get so good that they'll help companies deploy themselves and the forward-deployed engineers will again get to work at a higher level of abstraction. But for now, you go into a company that is not AI-native and say, okay, you've said you want to... you know, they say they want to deploy AI. They really are not sure what to do. How do I hook this up to my systems? Do I need to fine-tune a model on my codebase? How do I think about orchestrating agents and using things from different companies?
Most of all, or at least what we hear most frequently, is how do I think about security of my data and how do I know that these AI co-working agents are not going to go access a bunch of information and share it in ways they shouldn't or get, you know, a context exploit or something like that. So the forward engineers take this incredible new technology and a platform like Frontier and say, we will connect your company to an AI platform so that you can use all these agents and workflows and everything else you want.
How important are these metaphors or how temporary are they? I was very interested in reading about Gas Town and you have these pole cats and it's this whole Mad Max world and that feels like maybe just a temporary aberration where you're setting up agents for specific tasks, but also that could be incredibly valuable in explaining to a large corporation how they're going to integrate AI across the whole organization. Yeah, I suspect like everything else that's happening when an industry is moving so fast, all of this is somewhat temporary on a long enough time scale.
And you, as these models become more capable and these agents are operating on very long time horizons with the ability to just kind of figure it out and our trust and their robustness keeps going up, then maybe you don't need a lot of the abstractions we need today. You know, maybe you just have a single AI bot that runs at your company and you can say, 'Hey, I want to launch this new product,' and it does everything an ambitious person would do. But that's not where we are today. So today we have to put in a little more work to get the pieces put together.
Yeah. How are you thinking about the METR benchmark for long task horizon? You're at the top of the charts. At the same time, it feels like we might need a new chart if we're talking about agent swarms because they'll be able to do things that go for weeks, but they will subdivide the work. There's some subdivision that happens within a reasoning model, but it doesn't truly parallelize, at least that I'm aware of. So what does it look like in a world where you go to a model, but now it's spinning up a whole bunch of different models underneath?
I think there's two of the key insights of the whole field are in this question. Number one, the implication that no chart in AI lasts more than a few years is right. And like this one kind of, you know, we'll see how much longer it's really useful. Yeah. The second is a lot of people thought that, okay, we're going to just need a super long task and a super long task horizon, so we need a super long context. And definitely what people have already seen with coding agents is by agents breaking up work, orchestrating it well, farming it off to sub-agents, even with the current limitations of the technology, we can do something which should not be surprising because it's similar to how people do things and get amazing amounts of work done. So that's been cool to watch. I think that will keep going.
A joke that some people at OpenAI make is that soon the chart that matters is just going to be GDP impact. Yeah. And then the question is what's the one that comes after that? But everything else, these, a lot of these proxy metrics, there's now so much economic value in what the models are doing. What do you think could come after? I have no idea. Do you have an opinion? Happiness. I don't know. When you look back at some of your blog posts from 10 years ago, your predictions were usually pretty on point. Yeah. Maybe it's hard to predict. Thank you. Thank you.
Specifically the merge, basically there's a bunch that are good. Like right now we're getting, you know, thousands of messages in the chat about 40 and you predicted in 2016 that people would become, you know, very attached to a chatbot. Yeah. I'm working on like a big prediction blog post for the next 10 years seems too far, but the next five. But because it's like, I'm sure a lot of it will be wrong and it's still fun to try.
The sort of like relationships with chatbots, clearly that's something now we got to worry about more and is no longer an abstract concept. Even the question of what comes after GDP, like one reason I think that's interesting is the way we measure GDP now could start going down even though quality of life goes way up. Sure. And we don't have a lot of practice with things like that. But the massively deflationary... not just Europe, we want quality of life going up too, not just going down, you know. Yeah.
Switching gears, what do you think about the work that the Neo boom, the research efforts that are happening all over Silicon Valley? It feels like there's an acknowledgment that there's research breakthroughs that need to happen and everyone's taking different shots at those. Do you think that those companies will just find a breakthrough and join a lab or launch their own products? I think it's fantastic. First of all, one of the meta things that we wanted to do when we started with OpenAI, when we started OpenAI, is like there had been a period where the technology industry in Silicon Valley in particular was amazing at new research labs or just doing new research in industry in general, and then it kind of fell apart and there hadn't been a good one in a while. And part of what we hoped to show, and this was not only us, like a lot of people were excited about new research labs, is that industry could do research again. Mhm.
So seeing that now become like fashionable in all of these new labs, I think it's totally awesome. Some will succeed, some will fail, some will go into some other effort, but having industry support research in, you know, startup style, I think it's wonderful. Over the next two years, would you expect to acquire more individual product companies or more research labs? Good question. I don't have a strong opinion there. I would bet... well, I'd say the very best ones will often look like a mixture of both. Like the one that I have in mind right now is something that very much looks like a mixture of both. So maybe that's the shape of things to come is the really truly extraordinary product work will have more and more research component and it'll be kind of more of a hybrid thing.
Is data the new oil? Is there value? We were joking a couple days ago that they have a bunch of data, but they don't understand AI. They don't know how to monetize. Well, and that phrase was effectively wasted a decade ago. Yeah. And so to say it now sounds really silly, but it feels like it could be more true now than ever. Yeah. You know, certainly. Yeah, man. They really did waste it a decade ago. I was just thinking of the kind of people that used to say that and came out of it. TED talks, you're not supposed to call them out. They're nice people over at TED.
You know, definitely like the sort of the magic relationship of this last eight years, whatever you want to call it, has been like, you know, that we can put in more and more resources, compute, data, new ideas, whatever, into creating an artifact and it gets like the log of it gets better. Mhm. So you can throw in, you know, that's why we have this huge exponential increase in resources, but we keep getting better and better models. And for all of the concern people have about it's going to like top out or it's slowing down or whatever, like no one's been right about that. I mean, sometimes it looked like they were for a couple of months as we digested a new model or came to a new form factor, but it has been an incredibly smooth last six or eight years of this.
What those resources are. There can be some trade-off between, you know, sometimes it's better to spend your money on better data, sometimes on more compute, sometimes something else. On the whole, compute power is the new oil is the statement that feels closest to true to me, but there will be other parts too. Makes sense. Is software dead? It's different. It's definitely not dead. But what software, like how you create it, how you're going to use it, how much you're going to have written for you each time you need it versus how much you'll want sort of a consistent UX. Yeah. That's all going to change.
You know, there have been a number of these like big sell-offs of SaaS stocks over the last few years as these models have rolled out. I expect there will continue to be more. I expect there will be big booms in software. I think it's just going to be volatile for a while. Mhm. As people figure out what this looks like, the statement that someone said to me that has stuck in my mind most these last couple of weeks is that every company is an API company now whether they want to be or not. Oh yeah. Because agents are just going to be able to write a CS. Yeah.
We had D from Uber on yesterday and he had a pretty refreshing kind of approach. We were asking about integrating agents with Uber and he recognized that, yeah, the ad business could potentially be threatened if you can order an Uber in ChatGPT, but he basically said like you have to think of the consumer. The consumer wants to order an Uber via their preferred agent, you should let them, otherwise you're going to have other problems. Yeah, that is the right take for sure. And I don't... or I think so at least. Yeah. And you know, we've been through platform shifts like this before where you... I mean, Uber wouldn't have existed without one. It wasn't until the iPhone where you could like have it make sense to order an Uber to where you were as you were out in the world.
So I think there will be totally new things that happen. Other things you'll use in new ways. But definitely as I've started using Codex, my excitement about having agents go off and do things for me and still use other services, pay other services. I'm sure we'll need to figure out new business models and how revenue gets shared around, but that will happen. Yeah. Talk more about Codex Desktop. One more question on SaaS. Have any public market SaaS companies tried to get a soft landing with OpenAI? And do you think there's any value just let's say... not that I'm... no one has... no public SaaS companies that I'm aware of have tried that with OpenAI.
Look, some of them will certainly be durable and are on sale right now and potentially just need new energy and need kind of a new, an entirely new approach and maybe OpenAI could provide that. Yeah. Yeah. I think some will be incredibly valuable. Some do feel like a thinner layer now. But I don't know, like I was talking recently to a bunch of SaaS companies and they do not feel unexited. Like they're like, we're going to go through a big transformation here and, you know, yeah, sure, other people can instantly write software now, but so can we and we got a great system of record and seems reasonable. Some won't make it, of course.
Yeah. Yeah. Talk more about the Codex Desktop rollout. It feels like a successful amount of downloads, but a key shift for people who are maybe lightly technical but don't have time to set up an IDE and configure an environment to actually start writing software. I want to know about plans to integrate to the phone. That was a big moment I think for a lot of people with the ClaudeBot, MoltBot, OpenClaw thing was like, oh, I can text something and it will go and write code and that's valuable and that unlocks a new agentic experience. Like where do you see the Codex Desktop ecosystem going?
I am so... Codex Desktop has been somewhat of a surprise to me in terms of how much people love it, including how much I love it myself. You know, I think it's a great example of 10% of polish of the experience of using these models, especially when there's so much capability overhang, goes an extremely long way to what you can build and how you interact with this stuff. Of course, we should have an ability to kick off new tasks from mobile and we'll do that. I mean, really what you want is like your single AI that's working for you on a unified backend, access to all of your data and your ideas and your stuff and your memory and all the context and the ability to work across a lot of surfaces. And often you'll be at your desktop. Often you'll be on your phone and you just want to add something in.
But it is a pretty profound shift in my own workflow, not just for coding tasks, but more general purpose tasks. It's still kind of hard to use if you're not at least reasonably technical. But obviously, we'll find a version of this product that can do other knowledge work tasks and control your computer and things like that where you don't have to be. And it'll bring the magic of building stuff really to a lot of people because even if you never look at code, you'll be able to build something reasonably sophisticated.
One of the things that I have built when I was playing around with the new Codex app is this thing I had always wanted, just like this magic autocompleting to-do list. Oh yeah. I really work with to-do lists and this idea that I could just put tasks in and it would try to go do them. If it could complete them, it could complete them. If it needed questions, it would ask me questions. If I had to do something, I could still do it the old-fashioned way. But an interface like that where, you know, all the stuff you want to do, you just sort of explain to a computer or your AI and it tries to go off and do them. And sure, if you're on your phone, you're going to just add a task on your phone or, you know, if you want to easily import something from email, you're going to do that. Like, it feels really good. So I'm excited about all of the ways that this will just become a general knowledge work agent.
Yeah. Were you unsurprised to see a product like OpenClaw come from open source? Because it's certainly that kind of user experience is not... I would imagine this is something that you knew would be a thing. And yet I think part of the magic of OpenClaw is that it would be very, very difficult for a large tech company... Peter didn't make many phone calls to hyperscalers to say, hey, I'm going to be integrating your API. It just went. And you guys, you know, and when I think back on like the Sky acquisition, this kind of experience was probably very top of mind and things that you're working towards internally.
I love the spirit of everything about OpenClaw. Yeah. And you are totally right that it's much easier to imagine a one-person open source project doing something like that than a company who is going to be afraid of lawsuits and data privacy and everything else. You know, I think this is kind of how innovation works there. Something like that starts. It's clearly amazing. There will be a way to make a mass market version of that product, but letting the builders build the equivalent of the Homebrew Computer Club spirit go here is so important.
Yeah, totally. Can we switch to social? I feel like if I Google Sam Altman social, I get pure AI and Sora and then also demand or predictions about a human on social network. Where do you see social going broadly? How do you want to integrate it with it and power it in the future? The Moldbook thing was like a very interesting social experiment to watch and I think points to agents interacting in some sort of social space, hopefully on behalf of people at least in some degrees, could be quite interesting. I don't think we know what to do there yet, but it feels like social is going to change a lot.
And I am interested in the space of what a social experience can look like when, you know, your agent is talking to my agent and coming up with new stuff. Clearly putting like a lot of AI bots on the existing social platforms is just making everyone crazy and not that fun. So that's not the right answer. Yeah. But I think we can design something new for what this technology is capable of that will feel good and useful.
Yeah. Is there a solution to the bot problem that's just all the labs sort of integrating with all the other platforms and even if you can't detect it's AI generated, you can literally say we just generated those tokens, like those exact tokens are in our database? Can't do that because they're open source models are like good enough to write at this point. Yeah. I am excited about sort of like assertion of humanity instead of trying to like detection of AI as a thing here.
I don't know if the social platforms, it's like in their interest to solve this. It is because it's creating like, at least in the short term, it creates like a lot of engagement and increased usage. So I believe they could solve it if they wanted to. I'm not sure it's in their interest. I'm actually not even like, it is, I don't like it, but it is, some people do seem to enjoy it.
Yeah. Can you talk a little bit more about where Sora as a video generation model is going? It feels like tool use is maybe under discussed. You know, adding reasoning. It's not just the diffusion model. It's giving these models the ability to make linear cuts and overlay motion graphics. And when I scroll the Instagram reels that I see, they're like vibe reels with cutouts and it flips negative and it's all color graded. And it's stuff that like, yeah, you could probably diffuse it all, but...
It's pretty cool just to teach a model to also use After Effects or whatever video, you know, motion graphics suite you want to use. Is that an interesting unlock? What do you see going on?
So all of that stuff will happen and I agree with you. The models will get really great at doing that. People love generating videos. I would say people, we have not yet found a way that people really love watching other people's videos. This is true for a lot of other AI, like they love to, you know, people love talking to ChatGPT or whatever. It's not that compelling for most people to like read other people's generation. So I think there is something...
But isn't that for all writing and all video?
Yeah, it seems stronger to me in this case than the general case, but maybe you're right, maybe this is...
Yeah, but if somebody says, hey, I generated a 15-minute video, I'm really excited for you to watch it, and you watch the first 10 seconds and you're not that captured by it, I don't care that it was human.
Yeah, maybe you're right and it's not a special case.
Do you see that in the data with Sora downloads? Because I've noticed that I'll generate stuff on Sora, download it and share it to a group chat, and then it's this little in-joke that me and five other people get, and we see this like family group chats of, oh, it's our dog and our kids, but like there's not really like, okay, yeah, this is a business, you know, everyone likes this.
Absolutely. I would say that the most common use case is something like that. You know, like memes on group chats is a real killer use case of sort.
How is the Disney rollout going? I was super excited about it. Jordi was extremely bullish on it from a business strategy perspective.
Yeah. When you look at how image models have grown various LLMs historically, and now you're going to have an image and video model that can do something that no other LLM can do, at least legally.
Yeah. And is Bob Iger joining OpenAI? Bob, I love it. No credit for that.
He's going to be looking for a job and free agent here. He's a free agent. Pick him up.
Hit him up for us. You know, do some recruiting. That'd be great.
I think that generating characters in images and videos is going to be very important to people and they really like that. Like we were saying otherwise, I don't think many people want to watch me and some Star Wars character doing something together, but I might think it's cool.
It's, you know, there's like a real trend going on right now with ChatGPT where it's make a caricature of me and my job based off of everything you know about me, and those kinds of things people actually do like looking at other people's media.
Yeah, it's almost like a face filter or something. There's like enough of it's the Studio Ghibli moment, like there's enough of the human still in that. It's not, yours is not the same as mine. So it's still personalized.
It's personalized and it says something about you. And you know, a lot of what's gone viral before with image gen, I think it's like if you can make people look a little bit more attractive or cool than they are in real life.
Without sort of having to ask for that.
How are you thinking about the actual rollout? We were debating between like open the floodgates, you can generate any Disney property, versus like it's Spider-Man week and everyone's posting Spider-Man, and then it's Mickey Mouse week and there's another viral moment.
I'm not sure what the team is planning there. I know Disney's had some different opinions about what they want to do and try to be a good partner there, but I'd be excited to open the floodgates personally.
Oh, that'd be fun. Cool. Talk about your first speaking video, your first Super Bowl ad. It felt like not generated, lots of motion graphics, the black dots coming together. Like what was the goal with that ad? Who were you trying to speak to? It didn't feel like a direct response QR code, download the app. No. What was the mission?
I love that ad. I think that was such a cool one. It was clearly not meant to be like a mass market or direct response ad. But you know, speaking to the people who are at the center of this revolution and just trying to celebrate everything that has come before and everything that will come after.
We didn't hear a lot about it from like average users of Chat, but we heard a lot about it from like researchers in the field and a lot of resonance there. It was definitely not generated. It was done the old-fashioned way. And you know, a lot of people loved it, a lot of people hated it, and then many people in the middle didn't get it, and I felt okay about that.
It's a great encapsulation.
I like our ad for Sunday.
It's about Codex, no surprise. But...
Yeah, talk about the evolution of the advertising to be more just clear about the actual use case, the value. Like what are you trying to say with your advertising strategy now as it relates to video?
Well, the thing I would most like us to say, and I think this is a new challenge given where the models are, is to teach people what they can go do with AI. I think AI is now unbelievably capable in most of the world. It's still like asking basic questions on ChatGPT. Everyone can go build amazing things now. Everyone can go do all kinds of work. Scientists are going to make new discoveries. And I'd like to, you know, to the degree that advertising we do can teach people how to use this, I think that'd be awesome.
Yeah. So the KPI is like reduce the capability overhang broadly.
I think that should be a general KPI for us, not just of our ads. Yeah. The products that we build, how we teach people to use those products, that feels very important.
Yeah. Anthropic also has a bunch of ads in the Super Bowl. Seems like they run a ton. What do you think they're getting wrong about their characterization of how ads will roll out in chat apps?
Well, it's just wrong. The main thing that I think is we are not stupid. We respect our users. We understand that if we did something like what those ads depict, people would rightfully stop using the product. Our first principle with ads is that we're not going to put stuff into the LLM stream. That would feel crazy dystopic, like a bad sci-fi movie. So the main thing that's wrong with the ads is like using a deceptive ad to criticize deceptive ads feels... I don't know, something doesn't sit right with me about that. I asked Claude what the definition of playing dirty is, and it said, what did it say? Misleading others about your intentions, hiding information or creating false impressions.
Yeah. Thought it was a little dirty. I thought it was well played, but it was...
It was well played for sure. And it was a funny ad and the stuff about the ChatGPT personality that most annoys me, which we'll fix very soon, I thought they nailed in the ad. So that part was funny. But I don't know, you know, like I also think it's great for them not to do ads. We have a different shaped business. I did notice that they said in their thing like we may later revise this decision and we'll explain why. So...
Yeah, the blog post kind of did a good job of disarming the pro-ad people, gave themselves an out in the future. Do you think they care?
I do matter. Like I think it's a sideshow. You know, people are excited for a food fight between companies, but the amazing capabilities of these models, the product, the groundswell of excitement around Codex, that feels way more important to me.
How do you stop the pausing that happens in voice mode? Do you need new hardware for that or is it a model capability thing?
We need a new model. We may need some new hardware too, but mostly we just need a new model. I think we'll have a great voice mode by the end of this year.
What's the bigger bottleneck? Energy or chips?
It goes back and forth. Right now again it's chips. But it could be different at different times.
Is there anything we, like society, America, should be doing more aggressively to increase the supply of fabs?
Yeah. I think it may get solved on its own. Normal capitalism may solve it, but I think somehow deciding as a society that we are going to increase the wafer capacity of the world and we're going to fund that and we're going to get the whole supply chain and the talented people we need to make that happen would be a very good thing to do.
Do you think there's an upper bound on model IQ? Like the race right now is you're smart, but you're not smart for days. You're smart for hours. Can you go much farther and get much smarter?
It seems certain. Upper bound, I don't know. I don't know how to think about that question yet.
I can't even reason about what 2,000 IQ looks like, you know? Like I don't know what that means.
It's funny you say. I mean, I can't reason about what it means to think about a problem for like 10,000 human years.
That's another good one. Yeah, that's crazy.
But maybe IQ is going to feel even weirder. I don't know. I feel like this isn't going to feel as strange as it sounds. And for a bunch of reasons, we're so focused on other people. We're so focused on our own lives. We have such a human-centric nature that like, okay, this thing is really smart. It's inventing new science for us. It's running companies for us. It's doing all this stuff. And that sounds like it should be impossibly weird and I think it'll just be very weird.
Do you think space data centers will provide a meaningful amount of compute for OpenAI in the next two to three years, five years?
You just keep going. 10,000 years.
I wish you luck.
Okay. The funny thing about the whole back and forth about ads is that in our world the criticism is that you didn't launch ads early enough. Is there a world where you wish you launched earlier? How is the actual rollout going? Are advertisers happy? Do you have like a really long roadmap or do you think you'll be faster at catching up to sort of what's frontier in ads?
We haven't even started to test yet. We start to test soon. But you know, it's going to take us some number of iterations to figure out the right ad unit, the right way this all works. Do I wish we had started earlier? We have gone from like not a company, you know, three years and three months ago or something like that, we were like a research lab and now we are like a pretty big company with a lot of products. So there's many things I wish we had done faster. I think we were correct on the trade-off here of how we balance things that we need to do. I wish, you know, we launched this very cool enterprise platform this morning. Wish we had done that earlier too but had to deal with the monstrous growth of ChatGPT and Codex and all that other stuff.
Good problems to have. Last question for me. What happened to that internal writing model that you used to write the essay? That feels like something that was really cool, but we never really saw the light of day.
We're going to get a lot of that spirit into a future model. Again, it's like there's so much stuff happening. We have to make these hard prioritization decisions. I would love a cool writing model, not as much as I would love a cool coding model. And what is possible now for coding, for science, that's the thing I'm most excited about. For accelerating all kinds of research, AI and otherwise, for really accelerating the economy. I think that's like the right thing for us to most prioritize in terms of new capabilities. But yeah, you want a model that can write beautifully because it means it well, you want a model that can write beautifully if it can also think very clearly and express that very clearly. That's just useful in normal work.
Yeah, that makes a ton of sense. Last question for me. How have conversations been with the broad OpenAI leadership team? You guys are in a position where any single word or sentence you say in any situation can be spun into a headline immediately and then you guys have to go on damage control kind of correcting the narrative. But of course the original message is often or at least the original news is often seen more broadly than the correction. It seems like an interesting challenge.
It is a strange way to live. And I don't know of any private company that has ever been so in the news and so under a microscope. At some level it's frustrating and we're so squarely in the sights of everybody's anxieties and every competitor trying to take us down and everybody's like just what is going to happen with AI to their part of the business or their own lives. That there's like a lot of plasma looking for an instability to collapse on. In some other sense though, the subjective experience of it is we are so busy on so much exciting stuff that it often feels like there's this crazy hurricane turning around us and when we sit it's like fairly calm. You know, the media or Twitter goes insane about something one day, they're talking about a crazy meltdown. We're like that is insane. Like okay. And people talk about it all day and then later find out it's wrong and sort of seemed like a lot of wasted energy, but we're just like we have this great new model coming. People are building incredible stuff. Companies are transforming. We're trying to figure out how to get more compute and deal with this compute crunch. And we just kind of like keep going and we're busy. And then if we like open Twitter, pop up our heads and look at the news, it's like, wow, that is an insane crazy thing happening completely divorced from reality or 99% divorced from reality. And like, okay, someone will correct it, but then we get back to work and people flip out again. And it is much, it is weird to watch when we look outside, but it is less chaotic internally than I think you would imagine from reading the media reports.
Yep, that makes sense. Well, thank you so much for taking the time to come chat with us. Congrats on the... I'm excited to see the Codex ad.
Me too. Please try it. The app and 53 have been like I think the coolest thing we've done in a while.
Yes. With one prompt, I rebuilt the tbpn.com homepage to look exactly like Berkshire Hathaway and it was just immediate. It was very fun.
Plain text. It was very easy. Immediately one-shot it. Did not really push it to its limits, but I'm having fun. So, thank you so much for coming on the show. We'll talk to you soon. Great to catch up.
Cheers.
Let me tell you about ElevenLabs. Build intelligent real-time conversational agents. Reimagine human technology interaction with ElevenLabs. And I'm also going to tell you about Sentry. Sentry shows developers what's broken and helps them fix it fast. That's why 150,000 organizations use it to keep their apps working. And without further ado, we have Sholto Douglas from Anthropic. He's a member of technical staff and he's here in the TV. Sholto, how are you doing?
I am doing great. Seems like you're doing better. Tell us the news. What did you release? What happened?
All right. Well, once again, we've got a new model. It's really fantastic. The camera's a little bit laggy, a bit strange, but you know, should be good. This model is really, really, really fantastic. So, I think people have been comparing the previous generation of OpenAI and Anthropic models, and they've noticed there's some differences, right? The OpenAI models were a bit better at trying really, really, really hard on tough problems, but the Anthropic models were much faster and so forth. And so, they worked on speed while we worked on making the models much, much better at really, really tough problems. And so that's where I think the model really shines, is it's able to think for a really, really long time. It's able to expend a lot of test-time compute on thinking hard about problems. It's also a massive step towards the coworker that we've been working on. I think I mentioned last time that we're seeing this continuous progression towards models that are as capable at the rest of knowledge work as they are at coding. And I think I'm really, really excited about that as well. You can see it's a lot better computer use. We've got Claude for PowerPoint, claiming to become actually quite capable at Excel. You know, it's not perfect there yet. But huge steps, huge progress.
Do you have internal benchmarks for things like PowerPoint now?
We have benchmarks for everything that we work on.
Then but what makes it to the model card?
What makes it the model card? Oh, the model card is typically only publicly released benchmarks. Because for our internal benchmarks, we usually hold them out for internal testing.
Oh, sure, sure, sure. What's the internal process for benchmarking a PowerPoint because it's not...
I can't talk too much about benchmarking a PowerPoint but you know, I'm not sure if you know back a couple of years ago I did work at McKinsey so you know sometimes maybe they just need...
So they show them to you and you say it's good or...
Yeah, probably a lot of people involved. Yeah, talk to us about what a difference between 4.5, 4.6 means. You know, people might be familiar with pre-training, the different buzzwords that go into changing a model. Like what technically happened at Anthropic over the last few months to enable this.
Yes. So, we actually decorrelate the model release versioning from any specific technical details. It's more about the overall capability level that we're excited about. And actually, I've been asked by many people why we didn't just name 4.5 as five because it was such a step up. And I think maybe we under-anticipated. I mean, when I was last on here I don't think even I fully appreciated how much of a jump it was from four to 4.5.
It was great because by the time there was the 4.5 hype cycle we could repost a bunch of clips of you talking about it and people were like wow they covered it today. It's like no we actually had them on the day it launched and so I'm sure we'll be doing the same thing with 4.6. Talk about Task Horizon, the METR survey. Is that important? Obviously it's a public benchmark. We've talked about it before, but is that something that is a different trade-off to the other? Because I see it as like there's the IQ and then there's how long something can stay focused, like the locking in. And is there any tension between those? Do you ever get a model that's like, oh, it's really smart, but it can't stay coherent for that long or vice versa, it does, you know, dumber but longer tasks.
I mean, I think the METR eval is possibly the best eval currently out there, right? But you're right in saying that different parts of the chart measure different things. Initially, it was just testing IQ, but then as time goes on, it starts to test things where humans fail at tasks not because they aren't smart enough, but because actually persevering on a task for six, seven, eight hours is really hard. And you're right that these come apart. There's a sort of like models you should almost expect to have superhuman perseverance. But at the same time, maybe context coherence or something like this fights against that. And that's a different axis from raw intelligence or knowledge on given domains.
So METR eval, when I'm referencing AI progress to people. Recently we held this physics conference where we invited a whole bunch of physicists. We did it with Google DeepMind. We invited a whole bunch of physicists to try and convince them that AI was a really big deal. And that was the chart that resonated most deeply with people and I think is just quite conceptually graspable as oh okay things that would take me X hours the models are now capable of semi-relatively... It's an interesting benchmark considering that so many types of knowledge workers will never actually work on a task for seven hours straight. Like it's just like oh I worked on it for 20 minutes and I jumped into a meeting and then I took a look at it again and then I had lunch and then I went for a walk and then I came back I worked on it then I got somebody called, I had to work on something else. So it's like very possible that a lot of people listening to this, you know, maybe outside of some software engineers have just never even in their career worked on something for six hours straight.
The flow state. I mean, you chain things together and so forth. But yes, I agree. It's in that respect. And what does even a week-long task mean? Surely that's sort of inherently composable into a year-long work of progress. I mean I certainly don't think of myself as doing a year-long task when I'm at work, right? I'm doing these day or week-long time horizons usually which feed into a several month-long time horizon.
So talk about orchestration. It feels like there's a moment happening. I really like the Gas Town analogy and the world that he's built there. You see all over, I have four Claude Code instances running at the same time. It feels like there needs to be a new layer of abstraction. How do you think that gets solved? Is now the time to start learning those things? What metaphors will be valuable going forward?
Yes. So, right now the agents are still at this point where you need to manually multiplex across them. When I'm working, the stark thing about going from 4 to 4.5 and now 4.6 was that I went from maybe like 10% of my lines of code being written by me to 0% of my lines of code being written by me. But the thing is I need to actually sit there and constantly switch between these windows to make sure they're on track and give them guidance and feedback and you know I'm multiplexing maybe five at once or something like this when it's going really well. But you still need to be in there in the details. And I think the long-term right way to think about it is this constantly moving up levels of abstraction. So ultimately you only want to be talking to one agent that's maybe synthesizing the feedback from models as they come back and say I got stuck or I was unclear on this and that model can act as what I was doing before and I can take a step back and sort of act at an even higher level of abstraction. It doesn't feel like we ultimately want to be playing, you know, Age of Empires or StarCraft with the models, right? We shouldn't be APM-bound. Instead, it should be the models surfacing information as you need it. So that you constantly act on whatever the most important thing is and you context switch appropriately.
Speaking of gaming, do you have an update for us? When is the game shipping? Tell us about the game. What inspired this? Give us the backstory and then tell us what's going on.
Well, I think like many other people, I really wanted to test the limits of the models over the holidays, right? I mean, we saw all of that from people saying, I spent three, four days doing X. And I think that's really where almost the hype cycle for the most recent generation of models started in some respect when people got a chance to properly test the limits.
Yeah. Tells you a lot about adoption and how diffusion takes time because you need to have the space.
We need more long weekends.
Seriously, that's the bottleneck. Long weekends are the bottleneck.
Yeah. Good for GDP, I guess. Holidays.
Yeah. Yeah. Yeah. Sorry. Anyway.
Anyway, so Dylan's been getting us all into Age of Empires with these five matches on the downstairs table. And I wanted to see if I could make an Age of Scaling where, you know, you build solar panels and data centers, you train AI models and drones and so forth instead of farms and mining and all this. You start in the industrial era, you go all the way up to Kardashev. I only got like 70, 80% done and all the mechanics work, but it turns out it's actually really hard to make a game that's fun. Like actually just clicking make solar panel 100 times is not fun. Surprise, surprise. And it's fun for some people. There are some of those crazy clicker games where you do that and it's just like a brain rot game and they're pretty successful. But I think you should stay away from that. Not even Factorio. These are like mobile games that are like really, really bad. Anyway, I don't think that aligns with your mission.
Yeah.
But we talk about the grand geopolitics and strategy of this era that we're going into, right? Where the most valuable resources on Earth are, you know, well, basically the economy is in some respects reorienting around compute. And so I wanted to capture some of that dynamic. The late-night discussions of San Francisco in a game.
Yeah, I love it. How are you processing the work that's being done in world modeling and generative worlds? I know that you're not generating images, but Claude can take in an image and process it, and that has value. You text a screenshot and have them implement it or have him implement or it implement it. But what do you think about world models? Is that going to play an important piece, a link in the chain of where you're going?
Yes. So I think there's two different things here. There's the direct path to AGI, which I think is very much coding, and then AI research and general science and so forth. And I don't think that requires world models. But I do think that world models are incredibly exciting for a number of other things. I think they're very exciting from a gaming perspective. I mean, we all saw those demos are incredible with the Genie. It's just truly mind-blowing. And I also think they're probably the unlock to training robotics properly. It feels to me like there was this era of, oh, we're just going to need to fill like basically do incredible amounts of behavioral cloning of people operating robots. And it feels like the scaling is models but I don't think it's on the critical path to AGI basically.
Yeah. Can you unpack the concept of a software-only singularity?
Yes. So in this world it's one where the models are far better at digital tasks than they are at physical ones. And so we see rapid change in the digital world with relatively little change in the physical world. So information and software changes dramatically. And this ends up having some pretty weird effects. It means that maybe like the drivers of what have been the last couple decades of progress in the economy turn around like get changed very rapidly. And I think we'll see that flow on into the physical world, but at a delay. So you get much better at doing chip design. You get much better at training AI models. AI models get a lot faster. Chips get a lot better. The general economy gets a lot more efficient because the sort of information and message passing that is much of the rest of the economy ends up becoming much more efficient. But at the same time, you don't yet have robots providing limitless physical abundance. Science probably progresses really fast up to the degree that you need interaction with labs or larger particle colliders or something like this and then you go okay well I need to build the robots to...
But at the same time automated labs feel more near-term than unlimited robots in the real world manipulating, you know, the earth.
Yeah. I think maybe they actually arrive at similar-ish times. I think you need pretty competent robots for the labs or at least like no one's yet managed to figure out how to automate a lab without... turns out there are all these really weird little tasks that require a lot of manual human dexterity that are currently part of biological protocols. And so a lot of biologists will say no like you just actually need something that's capable of human-level dexterity. Maybe not for all experiments but for enough that it becomes annoying.
And in this software-only singularity, like how are you defining singularity? You know, the Kurzweil formulation of like more computing power than human brains, that's like one equation. There's also just like this point beyond which we cannot see. How do you think about singularity in that context?
I think there is a pretty tricky event horizon that I at least haven't found anyone that's made incredibly strong or good predictions. It just feels like at that point you have as many digital intelligences or more and they're as smart or more smart than human intelligences. What does that even mean for the world? It's incredibly hard to predict. It's something we spend a lot of time trying to think about and trying to prepare the world for all of the eventualities. But it's I think difficult to make sort of topline predictions of what exactly that looks like.
Yeah. In the original Kurzweil formulation, like that was almost the definition of the singularity was that you can't make predictions beyond it and so you know that like once the predictions break down then you're there, which is...
I mean I think in the near term, you know, I went down with Dwarkesh to his Elon interview in Austin and I think a lot of the sort of reflection of that in the physical world perhaps is what ends up happening like you do end up basically trying to climb the Kardashev scale and capturing more of the energy of the sun and so forth. But I think that takes some time.
Is your current takeaway we're more chip constrained, energy constrained? What's the biggest bottleneck to AI progress?
I mean, I think right now where I think I agree with Sam's earlier answer that we're more chip constrained in AI progress. But I think it's sort of interesting to roll forward two years and be like, okay, well, if you want 100 gigawatts or, you know, in four years and you want a terawatt, what does that look like? Where do you put that? And this is why Elon's going after the data centers in space, right? He thinks it's going to be the easiest way to get a terawatt in space in 2030. And maybe it's space, maybe it's the, you know, a giant desert somewhere. It's the Atacama desert, the Australian desert, somewhere in Texas, you know, maybe Texas has a lot of solar panels. Hard to know.
Yeah. Jordi, how are you guys thinking about free usage limits ahead of the Super Bowl?
It's really tough. I mean like one of the struggles of this is that compute is so constrained.
Yeah. I noticed you guys didn't say download the Claude app. There was no direct call to action.
Yeah. I mean I think the purpose of the ad was very much one for provoking like thought and discussion. And it was certainly successful already. Provoke a lot of thought and discussion. And I think I mean I don't think you need to do an explicit call to action for people to download things or to consider things.
Yeah, that makes sense. But again it sounds like the GPUs will be on fire Sunday is the expectation.
The GPUs will be on fire. As has been the case for the entire industry for the last year. I mean I think we've all had to make incredibly difficult trade-offs on exactly how compute is used.
Yeah. Can you zoom in on your experience because I feel like everyone's sort of seen like rough growth curves for Anthropic on the revenue side or on the tokens generated side or whatever and it feels exponential but smooth-ish and then simultaneously you have these constant worrying about bottlenecks. Is there enough capital? Are there enough chips? Are there enough data centers? Is there enough energy? Are we in an age of research? Are we in a plateau? How have you balanced those two narratives out, your experience of smooth growth even though it's exponential with constant fear around something like a cloud hanging over the industry and maybe a slowdown?
I mean I think so to note on the smooth or exponential growth. I think one of the things that really blew me away was when SemiAnalysis did that analysis and found that I think we went from like 2% a month ago to or 6 weeks ago to 4% of GitHub commits done by Claude Code. And for the amount of GitHub commits done by Claude Code to double over a few weeks is truly it's ludicrous, right? Like and there's no real visceral way to feel that. You almost feel like it's like feels like a number on a screen, but you can't viscerally feel it.
For us, we've always had a very, very, very strong conviction. And in many respects, Anthropic is a bet on this being true that scaling is continuing. And that progress continues unabated. So in many respects the external numbers are only a reflection of the conviction that we've had internally for a long time about exactly how we expect all the trends to go. I mean I think we're broadly on trend for like, you know, if you look at Situational Awareness, you know, I think pretty sure the power and like energy and flops predictions are bang on. It kind of feels more like we're just hitting each milestone as we expect. Roughly what we expected to have happened has happened so far.
In terms of diffusion of the technology, do you think that there's a role for forward-deployed engineers to go in and change organizations? We saw some news that OpenAI is hiring a bunch of folks. Obviously a lot of enterprises are using Claude now. But at the same time that phenomenon of if you don't have a free weekend or a long weekend, you might never get around to implementing it. And so having a little bit of extra horsepower and knowledge around the office might actually pull forward capability. Is that something you could see growing at Anthropic?
Yeah, I think it's a great idea. I think it's like very clear that people don't know how to hold these things. And also the fact that we like it's ludicrous. 3 months ago, we didn't have this most recent generation of models. We didn't have 4.5 and like you're meant to adjust your business strategy over the course of the holidays basically because the models are suddenly capable of doing things that they couldn't do right before you enter the end of the year.
Are you saying we could see a new white-collar job creation?
Maybe. I mean, look, I think these are really valuable jobs. I think that there's clearly a hell of a lot of value to be unlocked with them.
