LIVE CLIPS
EpisodeĀ 11-24-2025
Finding stuff, reporting back to the main model. It's a great way to decrease costs, make things faster. So we've seen a lot of value there. I think long term there is maybe a little bit of a tension between RL as a service and some notion of really cracking continual learning. I think it's a little bit of a race between RL as a service and can the labs crack continual learning. That being said, and maybe like one final note is that I've said this for a long time is I do expect things to eventually get to the point where large models only use as much computation as is actually necessary to achieve the task. Now, OPUS is one step in this direction, right? It only uses as many tokens as it thinks it needs to solve a given task and as a result it's more efficient. And I think that ultimately will take away a little bit from the sort of comparative advantage of small models as they get as large models get more and more and more efficient at only using the sort of right fraction of themselves to do things. But you know, I said that two years ago and it still hasn't happened, so. So maybe.
At the moment, even if the model was able to achieve a task technically by passing the test or nominally achieving your goal, it often doesn't code it in a way which is beautiful and allows you great abstractions that let you build on it in future. Often this is, at least in my own personal experience. It's not that the model is too dumb to do things, it's that it doesn't set things up well for future code. And so I think there are things not measured here, right? Yeah, but it's a pretty good proxy and I think it's a. It's a very good proxy for, for progress. Now, I think a lot of the tasks in IT are particularly machine learning research tasks. As AI models get better at that, I do expect the labs to hold back some of the capabilities there. Like, if a model is capable of, you know, writing out a whole new architecture, that's a lot better, you don't want to release that to your competitors. Right. Even if it's just capable of writing all the kernels for them, you probably don't want to really set your competitors. So in that case, I think they'll need to measure a broader array of tasks. And I'm also just very interested in seeing general software engineering tasks along this or other tasks in the economy, because I think that would be really informative for actual progress. I think GDP.
They've really gone and tried to fix this problem as well. Right. And it's tough at the scale of a billion users. But I think this is a good example of the kinds of things that are really tricky where there's trade offs and where you need to make sure that you don't have the incentive structure that allows you. That sort of like pushes you to maximize user minutes in this way and is a good microcosm of the alignment difficulties that we'll get as the models take on more and more and more responsibility in our world. Yeah, I mean, I completely agree with that. The user minute question like completely snuck up on me because I always assumed that everyone was going to be paying for this stuff as the $20 a month plans rolled out, the $200 a month plans rolled out. But of course, you know, you get to a certain scale of the Internet and it winds up being about attention and advertising and all these different. Yeah, and if you're, if you're building a digital coworker, people don't typically like rate their co workers by how much time they take up. Not like this. I love this employee. I love so much. I love my time every week, four hours every day on my calendar. It's the best. Steve just constantly talking to me. Okay, speaking of long running tasks, I want to know how confident should we be in that meter chart?
Post a screenshot of Roy Lee back in February of 2025. So literally just like months before he started Cluli. Very grateful to Kian and the Nucleus genomic teams for taking a chance on me the summer before Columbia and introducing me to the startup world. If I've ever seen a trillion dollar company in team, it's Nucleus. And Cremieux says he's obviously lying to cover up after getting caught doing fraud. An additional piece of background information people should know is that this fraud also employed the guy behind Cluley, the cheating company. And so Cremieu is being very, very hardcore in his assessment. Just actually calling Keon a fraud straight up is, is much more aggressive than, than just saying like, you know, some of the, some of their claims are maybe not legitimate. It's unclear. Like, you know, fraud is technically a crime that you need to be convicted of in the court before you are a fraud. But it's certainly he's putting his credibility on the line because if Keon comes out and says like, yeah, I'm actually not a fraud. I did it. I proved it wrong. The main thing here is it appears that the customer review.
Yes, that's purely an allegation that has. There's no proof. Yes, yes, exactly. Yeah, sometimes there's DMs that leak and there's evidence, or someone comes forward and says, like, yeah, I was actually paid to post that. But. So he says, I've been informed that Cremio. I don't know how to pronounce that last thing. Cremio, who's been on the show, also he claims he's a race scientist in chief has been paid off by the competitor to promote this nonsense against Nucleus. For the independent scientist repeating the term denied. Denied that as well. I would encourage you to do more diligence on who you're aligning yourself with. Our scientific team will issue a point by point response, which I believe they did. Unfortunately, though, this isn't about science. It's about. It's a concentrated attempt to cancel Nucleus on the backs of our successful campaign and build in efforts to build and advance the industry which benefits the very people attacking us. The mob are trying to cancel Nucleus. Keep tweeting, stay mad, we'll keep building and serving patients. P.S. we won the injunction. Link below. So they were sued by their competitor, but. So they won the injunction. But they didn't mean they won the case at all. I mean, that's a classic thing if you're getting sued to be like the case.
So narrowing it down, but full station blitz of Broadway, 1,000 plus street ads across New York City, 1,000 plus subway car ads, dozens of urban panels throughout SoHo. And apparently they're not actually. They're not able to offer the service in New York. Saw that in here. So it's really just an image of a controversial phrase on a New York subway is more likely to go viral. So you do it there because it looks like you're on the global stage and then you pull it. There's a high density of people that have a large following audience. Yeah, following. And so it's just the way to start a viral trend and own the moment. It's the reason why, you know, so many tiktokers are in Manhattan now doing stuff like man on the street stuff. It just like it has more like, aura almost. Well, Dr. Shelby liked the mindshare grabbing that Nucleus did. Says every biotech founder should be seeing this and understanding how to get one. 10 the mindshare of Nucleus. I have a playbook for you below. A lot of people are like, I love the playbook. I don't love this example because the company's getting dragged. I don't know if it's good or bad with the rage bait thing. I think usually it's a negative thing.
IVF done right in the subway all over New York City. There's a ton. Of debate going on and to be clear, accurate, I think it was intentionally trying to make some percentage of the population angry. Yes. To drive enough energy and attention. So this was, yeah, I would call it rage bait. So I would call it rage bait marketing, not necessarily rage bait. Not at the product level. But IVF as a category is a controversial category and so it's much easier to wrap it in a campaign that will go viral for upsetting reasons. You can upset people and you can get a lot of attention from that. This is an example from Kath Korvac. She says so eugenics is profitable now and so being able to wrap something that is just a scientific process that's been worked on for a long time. Seems to be somewhat friend.com inspired. Keon's original post. He says Nucleus Genomics announces the largest genetic optimization campaign ever.
Investment on that assumption. And if I'm being lied to, then that's potentially, that's potentially securities fraud. And so there is a, there is a question about if you go, I mean, I've, I've seen pitches for companies that, where they've said, like, proudly, like, you, you should invest in this because we're not training our own model. It would actually be a mistake. And you. And there's another company that's a competitor to us that is training their own model and you don't want to invest in them, you want to invest in. Us because we're going to burn your dollars. Yeah, we're getting much better economics. Yeah. And so all of it just, the only thing that matters is like being upfront with the investor for sure. And then to some degree, you do need to be upfront with the, with the, with the customer, because if the customer, there is a marketing value to, oh, if you work with us, you're working with these, you know, genius AI scientists who are going to build their own models. And if it's just repackaged ChatGPT, that might not be what you want to pay for because you, at that point you might just say, hey, like, actually, if I can just get this directly from OpenAI, I'll just go buy it from them anyway. Well, thoughts and prayers to friend of the show, John Palmer. He says he just found out my wife is leaving me. She said I'm not letting him.
Put lettuce, extra pickle. Who made that turkey? I think it's more like you go to the store and you bring a sandwich to the office. It's like, where did the sandwich come from? It came from the store, but like, you brought the sandwich, it came from you. You brought the sandwich. Yeah, this is better. I think you're right. Anyway, 22% of the time it was a hugging face model with Alora. 18% of the time was anthropic Claude with a prompt library. 8% of the time it was literally just GPT4 with system prompts. And. And 7% they actually train something from scratch. And all of these are odd. And there's a lot of debate over how this would actually happen because he's basically saying, just open DevTools, go to the Network tab, interact with the AI feature. If you see API OpenAI, API Anthropic or API Cohere AI, you're looking at a wrapper. They might have middleware, but the AI isn't theirs. And so it just opens up this debate about, you know, what is the value of the wrapper. I mean, certainly if you can resell something for 100x because you have some sort of clever workflow prompt or workflow, more power to you. Yeah, it's not exactly.
Even though it has been historically it has been up to this date like in consumer AI, it's very clear that OpenAI has run away with it but it feels like Google does have a little bit of a chance to catch up in consumer because there's just so much less of a network effect. Like the network effect just is bolted on. It's not real yet and maybe it'll never be real. Maybe Google can catch up there but I just really want to see where DAUs and user minutes actually grow because there's so many different tweaks there and every chart and data point is definitely going to be analyzed to death. Yeah one thing that's notable, Google's going super hard and it's for students so you can just get Gemini Pro for a year free and again I think that's just a bet on get people hooked on the workflow. OpenAI is clearly battling that out too. One of the big value props of using the Atlas browser is you get more advanced thinking queries like they will up your limits.
Monopoly aggregator thesis that I was picking up on was for the last year there have been a lot of features and theses around different things that could create lock in so stuff like personalization or memory or even the chat functionality between what you've linked your custom instructions the different like I think at this point I've synced ChatGPT or Auth ChatGPT with a number of different services so it should have more data, it should know all these different things I've given it even custom instructions just saying like hey cool it on the EM dashes and I didn't miss any of that. Like there was at no point where there were plenty of points where I was like oh like chatgpt is definitely better than Gemini still. But at no point was I like it's because it doesn't have personalization and I think that if I went in Gemini and I was like oh yeah, you can go take a peek at my Gmail to get personalized to understand how I write or understand what I'm interested in one I could snap my finger and Google could be way more.
Coding. Yeah. The app should be flawless. If I find one bug in the cloud consumer app, it's over. Do you guys ever use the voice to voice, like the real time audio thing on ChatGPT? No, I don't like that at all. You've never used it? I've used it. I've used it a bunch. I've used all of them, but it's just not the preferred way of interacting. Yeah. You were testing it out, Tyler, by talking with it for like eight hours a day, right? Yeah. And you were on the X Xai one. Yeah. With Ani. Was that. Imagine running constantly with a VR headset. With a VR headset and a full immersive suit in a sensory deprivation tank. Yeah. No, no, no. Why do you bring it up? I actually, I've started using. I've started using it. Like, it's pretty good. I think the model is actually much worse. Like the underlying model. Yeah. It has to be faster, right? Well, it's not the speed. It's like the actual intelligence of the model seems lower. Yeah. Like, the answers aren't as good, but I find it's useful for when I'm trying to learn, like a specific topic or something. And then I explain it back and then it tells me like, oh, is that correct or not? That's pretty good. I like that. It's really remarkable. I mean, the Gemini app.
Apple. I mean, I can't imagine how big their hardware group is, but it has to be, you know, in the thousands, I would imagine. Yeah, let's try to find out. Huge organization. So OpenAI is poaching left and right from Apple's hardware engineering group, hiring around 40 directors, managers and engineers in the last month from nearly every relevant Apple department. Mark Gurman says it's remarkable. So from what I've heard, this is Mark Gurman. Apple is none too pleased about OpenAI's poaching, and some consider it a problem. The hires include key directors, a fairly senior designation, as well as managers and engineers. And they hail from a wide range of areas. Camera engineering, iPhone hardware, Mac hardware, silicon device testing and reliability, industrial design, manufacturing, audio, smartwatches, vision pro development software. They got one from every single. Sampled. Every single. Every Single division, I suppose. Gemini is estimating that Apple has between 15,000 and 20,000 hardware engineers total. That seems like a lot. I don't know. In other words, OpenAI is picking up people from nearly every relevant department. It's remarkable, says Mark Gurman. Very interesting. I wonder, I wonder how the. How the comp structured, how everything will come together on those teams. I mean, there's a lot of people from Apple who, going over to OpenAI, it's a greenfield project, probably really fun, probably really exciting, probably not the. The most mercenary scenario, but there's always that if you're working at Apple and you're.
Do you think that's something that happens in 2026? Because right now there's a pressure to just be state of the art, like be at the frontier, basically. There's a vibe war happening and it's very important to constantly be topping all of the benchmarks. Didn't llama release that same user agreement where it was like, if you have less than 400 million DAUs, you can use the service and it JS Lee excluded all of their capabilities? Yeah, but I think that's some. I think that's pretty imperfect because there'd be a lot of ways that you could still get benefit without necessarily. But yeah, but how are you thinking about it? Well, I mean, there's a suite of capabilities here. Right. Like obviously you want. I think for general software engineering, everyone in the world should be able to use that. That's great. Let's say, like, if we train the models to get really good at assisting our own AI research, if we're teaching them mathematical tricks that we've thought about and we don't, we're not confident that anyone else knows or we're teaching them sort of like how we do our infrastructure, ultimately we want them to know those things. We don't want the rest of the world to be able to recreate infrastructure from scratch as a result. I think this is also similar to how we think about bio.
Nucleus did says every biotech founder should be seeing this and understanding how to get one. 10 the mind share of Nucleus I have a playbook for you below. A lot of people are like, I love the playbook. I don't love this example because the company's getting dragged. I don't know if it's good or bad with the rage bait thing. I think usually it's a negative thing. Usually it's hard to come back from. Occasionally it can be done in a way that's, you know, slightly enraging. But people enough people are in on the joke that they appreciate what's happening and they appreciate that it breaks through or it's enraging to someone who's not the core audience, not the actual customer. And so it's okay. But it's a big debate because Sichuan Mala posted a long essay all about the claims made by Nucleus. Kiyan says everything levied unto Nucleus by Sichwan is false. Worse than false. It appears to be architected by a competitor that has repeatedly published misstatements and inaccuracies. Sichuan is compromised, but it gets worse. Yeah. To be clear, no evidence has been provided that it was being levied by a competitor.
Leak anything. But I think there'll be some news soon, hopefully. I'm very excited for them. Yeah, we got to hang with Vincent Saturday. So OpenAI has an announcement. They're introducing Shopping Research, a new experience in ChatGPT that does the research to help you find the right products. They clearly were listening to me on the show just a few days ago when I was saying I would be using this for the holiday shopping period. Very exciting. I wonder how it will actually play out. You, of course, had that problem with cars and bids. ChatGPT was not identifying the fact that that GT3Rs had been sold two years ago. Thankfully, we have Doug Jumeiro here in the restream waiting room about to join the show. We can talk about cars and bits. We can talk about cars. We can talk about artificial intelligence. Welcome. Welcome to the show. How are you doing? I'm good. Thank you for having me. Gentlemen, thank you so much. So great to have you on. We are a technology and business show, but we have your name has come up probably 100 times independently, just when we're talking about 100% cars. So it's so great to have you on the show. Yeah. Thank you. We're huge fans. Thank you. Thank you. I'm thrilled to be here. I really am. Thanks so much. I'd love to know. I mean, we were just talking about this OpenAI shopping research in ChatGPT. It feels like there's really no substitute for watching a Doug DeMuro video to actually understand the quirks and the features of the car that you're considering purchasing. Yeah. I had been critiquing it. The context is like, I was trying to use ChatGPT to find a specific car, a GT3RS. A GT3RS. And it fully missed every car for sale on the Internet. There are many of them. And it found one that had sold on cars and bids, like, two years ago. I was like, thank you. Thanks for nothing. But good job showing up in the results. Are you getting any leverage out of AI tools at all for what you're doing these days? Well, that's an interesting question. The guys who are on the business side probably are using AI a little bit more than I realize from AI.
What is the key unlock for Opus 4.5? I mean, a lot of people are throwing around biggest. Obviously the benchmarks are very good, but this whole idea of more parameters, more data, more compute, more money, more electricity. How do you even think about allocating resources to push a model forward in 2025, in late 2025, when perhaps we're past this paradigm of like, oh, just more parame. I don't know if we are past the. I think it's important to call the paradigm like scaling in general, depending on what access you're actually scaling. Tbd, but general, the scaling paradigm, I don't think we're past that at all. Right. I mean, I think we're still seeing massive returns to scaling in all its variants. I think that we're generally, things work. We scaled, it works. The models just want to learn, right? It's as if you said like 10 years ago, yeah, the models just want to learn. And I think the hardest things with questions of focus are often on how we split and allocate people. And this model is hundreds of people worth of effort, right where they poured their lives into it over the last six months. And I think that is working out what we prioritize is really tough. I've said this before, but these models always feel like when they launch, it's exciting, they're great. But you sort of think back to, oh, my God, there's all these things that we could do better. And everyone right now is going and working on those things. It's just everything still works. So is this a refutation of what some folks might have been picking up on from the last few Dwarkesh guests? The Carpathia episode, the Sutton episode, There's been a little vibe shift around. Like, okay, maybe when we say scaling, we mean more inference diffused all over the world and small models and custom RL environments here and there. And, like, we're going to get the value from AI and we are going to continue to scale dollars to economic value, but it's not just going to be bigger and bigger pre trains forever. And then we get. God, I don't know what it's going to be bigger versions of, but as far as we're seeing, scaling still works. I don't know if you guys know this, but Dwarkesh, Dylan and I are actually a housemate. So we have this debate all the time, a dinner table discussion of, like, are we. Oh, he's slowing down. And, you know, I've often joked that the most impactful thing that, you know, one of us could do is go and crack the problems that, like continual learning or something like this that Dwarkesh focuses on so we can then go switch the narrative back to progress. Yeah, just switch up the dinner table conversation. Conversation. Exactly. So did, did, did, did the anthropic crew, like, never lose faith?
You said earlier you could imagine a scenario where labs would kind of hold back frontier models because they would be effectively handing their competitors an advantage. What's your timeline around that? Do you think that's something that happens in 2026? Because right now there's a pressure to just be state of the art, like be at the frontier. Basically, there's a vibe war happening and it's very important to, you know, constantly be topping all of the benchmarks. Didn't Llama release with that same user agreement where it was like in. If you have less than 400 million DAUs, you can use the service and it j. Excluded all of their. Yeah, but I think that's some. I think that's pretty imperfect because there'd be a lot of ways that you could still get benefit without, you know, necessarily. Yeah, but how are you thinking about it? Well, I mean, there's a, there's a. Suite of capabilities here. Right. Like, obviously you want. I think for general software engineering, everyone in the world should be able to use that. That's great. But let's say if we train the models to get really good at assisting our own AI research, if we're teaching them mathematical tricks that we've thought about and we're not confident anyone else knows, or we're teaching them how we do our infrastructure, ultimately we want them to know those things. We don't want the rest of the world to be able to recreate our infrastructure from scratch as a result. I think this is also similar to how we think about biology. And this is actually a line. I think we, we, we need to like, do some work on exactly how we draw the line here. But at, you know, sort of house view anthropogen, we're quite worried about the ability of models to become much better at, at biology and, and sort of producing old viruses and this kind of thing. And so as a result that we, you know, we have like all these safeguards around whether or not the model is able to do and, and help people with biology. It actually, at the moment, I think the safeguards are a little bit. They err on the overactive. I know many biologist friends who are frustrated because it doesn't quite. They're like, please, I can be trusted with biological super intelligence. Exactly. I will not create. I will not create a virus pandemic. Yeah, yeah. We're erring on the line of safety here and we're sort of navigating to finding the exact right pathway there. Yeah. Probably the most important question I have, what are your timelines around a humanoid robot beating a human at fencing. Oh, very good question. So, I mean, I'm as an expert. As an expert at fencing. The unitary robots are pretty good at backflips and stuff. They're a lot better than backflips than I am. So. I know, but fencing takes grace and finesse and all these things that we're not seeing in these. Also physical size. Right. Isn't height an advantage in fencing and reach. And. And I believe those unitary robots, I think sholta. I think me and you got. Got a foot on them easily. We do. People don't know, but everyone on this call is over six feet. People just assume. They see talking head and they think, oh, bunch of five, five, five guys. Yeah. Not true. Maybe. Maybe sports are hard. Maybe mid mid 2030. Mid 2030. Whoa. I'm really excited for. I'm not feeling the acceleration. I'm sorry. Sell everything. I think we get dropping co worker in two years and I think fencing robot takes a little bit longer. Well, that's your fallback plan. If you lose your job as an AI member of the technical staff, you go back to fencing what, swordsman in the world. It's going to be great. Yeah, that'd be fun. I can't wait to tell you. Operate a robot like manga style and fight. It's going to be. Yeah, that's going to be wild. Yeah, for sure. Yeah. Somebody was saying that some of the bull cases for some of those humanoid robots is that you just all get in VR and you just get to go hang out with your friends as robots and do whatever you want and you're just hanging out in person. Very funny. Last question for me, actually, from our intern Tyler, who's wearing the thinking cap. Thank you for sending it over. He's a huge fan. There we go. Do you expect mechanistic interpretability research to make meaningful contributions to capabilities, not just safety of the models, like actual capability results? Yeah. Great question. One of the interesting things about mechan work so far is that it's already lent, I think, to a lot of capabilities progress because of the mental models that are provided. Actually, after the original transformer circuits papers, it was interesting how the language of that paper ended up really dominating the mental models and the way that people thought across multiple labs about what actually was going on inside Transformers. And it led to, I think, a much deeper and richer understanding of what they are. So. So I think it's already helped in quite diffuse way. Not in a concrete way, but in a diffuse way in terms of the concrete ways dial up the smart neuron or something like this that I haven't really seen yet. And I think it's mostly sort of future work is going to be mostly in an alignment direction, but the sort of rich understanding and the rich understanding has helped us a lot in terms of actually understanding how to train these models. I have one extra question. Go for it. Tell me a little bit about Dario's communication style. I was hearing a story about, I think Jensen has no direct reports or no, everyone reports to him and no one reports to him. He has no meetings or all the meetings. And he likes 60 direct reports. 60 direct reports, but no big meetings. And he has. But he reads everyone's to do list like every single day or something. What's it like at Anthropic? What is Dario like as a leader these days? Yeah, Dario has a really, really cool communication style, which is that he quite frequently puts out very, very well reasoned essays. And then throughout Slack, we'll have giant essay length like comment debates with people about different ways. It's really great you get these. But the essays are really nice because one, you can go back and read all the past ones and it tells this history of Anthropic. Yeah, it's, you know, I think in many respects like it will be one of the better, you know, in a decade from now to chart the history of AGI. We'll be reading these like compendium of essays. Yeah. And, and, and there's like, there's incredible comment threads on either side of them and so forth. But also throughout Slack, whenever where he's very open and honest with the company, whenever we're debating different things, he will lay out the pros and cons and how he's thinking about them and, you know, why this one's attention and why that one's moral struggle. And people will write back big essays on why they think we should do X or Y. And he'll respond. It's really, it's quite a joy. It's a very recent communication style. Yeah. As a result, it means that many people, or really the entire company have a good model of how he's thinking. Yeah. And that really helps because it means that you sort of have a coherent sense of direction across the entire company. Yeah, that makes a ton of sense. I like that a lot. Yeah. Cool. So many examples of successful founders who have adopted the written culture and seen great results, I think. And he's a great writer. I mean, read Machines of Love and Grace and it's just such a Brilliant essay. That's great. You're absolutely right. Have you ever caught him using AI? Has he ever been like, oh this one, he's phoning it in. Not yet, but maybe soon. I mean it's kind of a bull case. If he does wind up just saying Claude, like handle it. I'm going on vacation for a couple days. I'm the drop in coworker. I'm pretty sure we measure loss on, on his essays. That's good. Yeah, yeah. But right now, I mean there's a high bar. High bar. But congratulations. Thank you so much for taking the time to hop on the show. Yeah, super impressive. Congrats to the whole team. We'll talk to you soon. Good to see you. See ya. Ciao. Bye. Back to the show, back to the timeline. Back to linear. Meet the system for modern software development Purpose built tool for planning and building products There is more OpenAI news, of course, more tech news of all times. OpenAI's hardware division says Mark Gurman built around Jony I've secretive startup has ramped up the hiring of Apple engineers to the group has brought on about 40 new people in the last month or so with many of them coming from Apple's hardware group. Yeah, hearing that Sholto interview, I'm disappointed. I don't think we're getting ads from Anthropic anytime soon and I don't think we're going to get a mobile device. Well, we are actually talking today to Quinn Slack, the CEO of AMP and sourcegraph. AMP is a frontier coding agent and AMP is free. They introduced AMP Free which is ad supported and has a no cost mode and so you can now use their coding agent for free with ads. 40 people. That does not seem like cause for concern for Apple. I mean I can't imagine how big their hardware group is but it has to be in the thousands, I would imagine. Yeah, let's try to find out. Huge organization. So OpenAI is poaching left and right from Apple's hardware engineering group, hiring around 40 directors, managers and engineers in the last month from nearly every relevant Apple department. Mark Gurman says it's remarkable. So from what I've heard, this is Mark Gurman. Apple is none too pleased about OpenAI's poaching and some consider it a problem. The hires include key directors, a fairly senior designation, as well as managers and engineers. And they hail from a wide range of areas. Camera engineering, iPhone hardware, Mac hardware, silicon device testing and reliability, industrial design, manufacturing, audio, smartwatches, vision pro development software. They got one from every Single sample. Every single, every single division, I suppose. Gemini is estimating that Apple has between 15,000 and 20,000 hardware engineers in total. That seems like a lot. I don't know. In other words, OpenAI is picking up people from nearly every relevant department. It's remarkable, says Mark Gurman. Very interesting. I wonder, I wonder how the structure, how the comp structured, how everything will come together on those teams. I mean, there's a lot of people from Apple who, going over to OpenAI, a greenfield project, probably really fun, probably really exciting, probably not the most mercenary scenario, but there's always that. If you're working at Apple and you're excited about AI and you've been there for the last three years, watching all this progress happen at the application layer and the model layer, and not being thrilled with the progress happening at the hardware layer. This is like a. Yeah, it's a, it's a wide open opportunity to like be working right at that intersection of the models and the hardware. There's a lot of AI engineers who have made moves because they don't want to be a GPU poor company. And it's weird because Apple's in this, in this scenario where they're partnering with Gemini now, they're clearly going to survive. It's not a serious threat, at least not yet. Maybe if this device is incredible. But right now Apple looks pretty strong. The new iPhones are selling well, everything's good. But what's like from an AI perspective, it's gotta be one of the worst gigs because you were in this sort of like openly hostile environment to LLMs, to scaling, to building large GPU clusters and then, yeah, they're sort of playing catch up now, but they're certainly not calling up Oracle for, you know, a trillion dollars of compute. You go over to OpenAI, you're just gonna be immersed in a lot more higher risk taking higher risk on. I wonder. Yeah, Gabe is asking if wouldn't that be bearish as the hardware group at Apple is responsible for the terrible Vision Pro. I do wonder. I do wonder. It depends on who it was. It was a, it was an incredible technological feat. I just think they built the wrong. I was just watching a thing about A guy who 3D printed an adapter so you could use the strap from the Apple Vision Pro on the quest 3 from Meta and like. And that really speaks to the fact that like the Apple Vision Pro, although it was too heavy, it has this screen on the outside that I don't think anyone wants. Like, there were pieces about it that Were clearly like the best. Like the screen is just the best. No one's debating that. The band, the knit band is very cool. It has this amazing device where you rotate it and it tightens up. There are a whole bunch of things that are amazing. It's just like as a package, it didn't deliver. But if you just want to, if your job is just like, hey, we gotta put a screen on this and it's gotta be the highest resolution screen. Like, well, go to the place that developed the highest resolution screens. Like, they did a good job. Well, Sam Altman replied to one of our cards we put up on November 22. TVPN posted on this day, Sama was rehired at OpenAI. Got his badge back and Sam Altman replied and said, cannot believe this was only two years ago. Subjectively feels like five. Yeah. What a turnaround to go from defenestrated to back in the seat in so much and have so much control over their organization that you're able to raise at massive valuations. Strike broker. All these deals move the entire market. Just a remarkable run. Put on such a master class in deal making that people are now sitting here being like there's no way that this would be a $500 billion company if Sam wasn't in the driver's seat. First, let me tell you about fall build and deploy AI video and image models trusted by millions to power generative media at scale. Danny Zhou, who founded the reasoning team in Google Brain, now part of the Gemini team at Google DeepMind. He says game over and carried no interest. Friend of the show quotes it and says I genuinely think OpenAI equals equals Yahoo. He's not assigning the variable, he's equating it. I've migrated almost all my workflows code off their APIs now. Ironic that Google will probably do it twice. Lmao. And I don't know about this. The pattern matching on the Yahoo example. We should have him on the show. And actually yeah you were saying the other thing with the Yahoo example is it wasn't like there wasn't like the company was valued at pretty tremendously for a longer period of time. It wasn't like this binary like one. Moment exactly like the peak market cap for Yahoo 125 billion during 2000. That feels like it's just hard to pattern match perfectly to this. But I mean it certainly would be poetic if that's the way it played out. It is funny. Carried says, ironic that Google will probably do it twice. They actually created the transformer they released the transformer paper and chromium to inspire themselves to find harder sort of challenge. Giving everything a freeway. Give out all the alpha. Yeah, yeah, for sure. Just kind of like find their fire again. Yeah, there was a, there was an interesting article on Medium that was sort of burning up hacker news that I thought would be fun to go through. First, let me tell you about graphite.dev, code review for the age of AI. Graphite helps teams on GitHub ship higher quality software faster. So this person who no one has, no one can really understand who this person is. They don't necessarily exist on the Internet fully. So there was like a question about that. This is like a very like, you know, hacker news and turmoil segment. But Tay Ha says, I reverse engineered 200 AI startups, 146 are selling you repackaged ChatGPT and Claude with new UI. And so basically the thesis of this article is that this fellow wrote a piece of code that looks at the marketing copy and says, what are they claiming? And then looks at the calls that happen when you actually interact with their AI feature. So if there's a chatbot on this particular startup's website and you are near chatting with it, and you look into the trace that's happening in Chrome, is it going to the startup server or is it going to OpenAI server, or is it going to Anthropic server? That's telling. And then there's also a little bit of API fingerprinting. Basically, OpenAI has a specific pattern of rate limiting and it's exponential. So if you're spamming the OpenAI API, it will, according to a unique pattern, tell you, hey, you've sent too many messages, cool off for one minute, and then the next time you do it, cool off for two minutes. The next time, cool off for four minutes, then eight minutes, then 16. Right. And it gets exponentially longer and you're on progressively more longer and longer timeouts. But the shape of that curve and the specific timings are unique to OpenAI. And so if I'm a startup and I have the exact same like back off and timeout curve, well, then it's probably just OpenAI under the hood. At least that's the claim that's being made here. And so the finding in this article is that 73% had a significant gap between the claimed technology and the actual implementation. And so out of the 200 AI startups that this fellow analyzed, 54 companies either had accurate technical claims. They said, hey, we're using like we have a custom AI model that we trained and they did. Or they're transparent about their stack. They say, hey, this is a wrapper. Like we're a wrapper company. And so, you know, our AI is powered by chatbots, we're partnered with OpenAI, we're partnered with Anthropic or whatever. Now 146 companies, that's 73% according to him, were sort of misrepresenting their technology. So either they said they had proprietary AP AI, proprietary AI. And yet when he dug into it, it was OpenAI API plus prompts. Tyler? Yeah, I mean it's like kind of what, what do people expect? Like, like is if you fine tune. If you use the OpenAI API to fine tune the model. Yes. Which you can do. Yes. Is that proprietary? Like no one else has that fine tune? Yes. You're still calling the API. It's like I don't expect startups to train their own full language models. That's like pretty unrealistic and doesn't really make sense. Yeah. So I'm kind of confused, I guess. It's a very cool study, but this tracks with exactly. Like I would guess that 73% of AI startups are just re skinning. Yes. And so 19% of the overall companies, the 38 that were analyzed in this study found that the startup said they had in house models and it was actually fine tuned public models. And so there's a question, is like whose house is it in? It's technically in OpenAI's house. So fine tune, as I mentioned, an open source model that's public, does that count as a public model? An open source model? Let's assume yes. And then, and then last 8% would. So they had a custom ML pipeline and they were in fact using standard cloud services. That's even wishy wash here. I think that's totally fine because like you can totally have a custom ML pipeline that's wiring together OpenAI and Gemini and AWS and you know, a bunch of other. If I'm using a startup, I don't want them to train their own language model because I don't think they're going to like in 99.99% of the case, like they're not going to be able to do a better model than OpenAI, Anthropic, Gemini, Grok. Like I want them to use the best model. Yes. And it's like, okay, you can fine tune it over. I agree, that's totally fine. And so the author also agrees with you. He says here's what really shocked me, I'm not even mad about it. Every time I saw the phrase our proprietary language model, I knew what I was going to find. And I was right 34 out of 37 times. And this is where it gets weird because he says, here's the technical signature. And so the user submits the query, it posts to API generate and then with wrapper logic it posts to API.OpenAI.comv1chat completions. And I have no idea how he's seeing the back end. It makes no sense how he would be able to do this unless there was just like a massive security vulnerability. Because what I would assume is happening is that the users over here, the startups website's here and then the user goes to the startup's website and then the startup's website on the back end talks to OpenAI and comes back. And maybe you could understand that like, okay, the amount of EM dashes, like there's a. This is telltale signal. But that's not what he's doing. He's saying that he was able to just literally hit like the Chrome Inspect developer tools. Look at the, look at the chain of calls and see that it was calling OpenAI from the front end, which is crazy because I didn't even know you could do that. It feels like if you were calling it directly from the front end, you would like potentially leak a key that would be able to put you on the hook for a bunch of bills. I would think you would want to authenticate that on the back end. And so he gives a bunch of examples of rag and then he's exposing some margins, which is actually very bullish for these companies because he breaks down some of these and says that a GPT4 API is $0.03 per thousand input token, $0.06 per thousand output token. So the cost per query for this hypothetical startup was $0.03 and they charged $3 or $300 per month for 200 queries. And so the wrapper economy, this is 75 times direct costs. That's extremely bullish for printing for that company. He found another one that was maybe 1000x API costs that's doing some pinecone embedding. And he also says this is pattern number three. The quote, we fine tuned our own model. Reality check. Fine tuning sounds impressive and it can be, but here's what I found. 45% it was OpenAI's fine tuning API, which, that sounds right, right? It's a little bit of a step to be like we fine tuned our own model. It's like no, you fine tuned OpenAI's model and then you got your own model from that result. It's a little bit. Yeah, you're still fine tuning it. I don't think there's that big a difference between fine tuning. Are you fine tuning it or is OpenAI fine tuning it? There's a fine tuning API which you use to and then a new but. Who'S doing the fine tuning, you or OpenAI? Well, what do you like? You're not interfacing directly with the gpu. It's like I went to the store and I bought a sandwich. Who made the sandwich? Well, I told them what the used turkey, I told them put lettuce, extra pickle. Who made the turkey? I think it's more like you go to the store and you bring a sandwich to the office. It's like where did the sandwich come from? It came from the store. But like you brought the sandwich, it came from you, you brought the sandwich. Yeah, this is better. I think you're right. Anyway, 22% into the time it was a hugging face model with Alora. 18% of the time was anthropic Claude with a prompt library. 8% of the time is literally just GPT4 with system prompts. And 7% they actually trained something from scratch. And all of these are are odd. And there's a lot of debate over how this would actually happen because he's basically saying just open dev tools, go to the network tab, interact with the AI feature. If you see API, OpenAI, API, anthropic or API cohere AI, you're looking at a wrapper. They might have middleware, but the AI isn't theirs. And so it just opens up this debate about what is the value of the wrapper. I mean certainly if you can resell something for 100x because you have some sort of clever workflow prompt or workflow, more power to you. Yeah, it's not exactly like bearish on. The company's yet, but the debate that was surrounding was more around so. So this, this author claims that after posting this, seven founders reached out privately. Some were defensive, some were grateful. They asked for help transitioning their marketing from proprietary AI to built with the best in class APIs. Because some of these founders did did I guess feel like using proprietary AI as a marketing tagline was disingenuous. And then someone else, I think I saw something that was one VC reached out and said like I'd like you to audit my portfolio because I have been told that I was investing in companies that were training their own AI. And I made the investment on that assumption. And if I'm being lied to, then that's potentially securities fraud. And so there is a question about if you go, I mean, I've seen pitches for companies that, where they've said like proudly like you should invest in this because we're not training our own model, it would actually be a mistake. And there's another company that's a competitor to us that is training their own model and you don't want to invest in them, you want to invest in us. Because we're going to burn your dollars. Better. Yeah, we're going to much better economically. And so all of it just the only thing that matters is like being upfront with the investor for sure. And then to some degree you do need to be upfront with the customer because if the customer there is a marketing value to, oh, if you work with us, you're working with these genius AI scientists who are going to build their own models. And if it's just repackaged ChatGPT, that might not be what you want to pay for because at that point you might just say, hey, like actually if I can just get this directly from opening, I'll just go buy it from them anyway. Well, thoughts and prayers to friend of the show, John Palmer. He says he just found out my wife is leaving me. She said I'm not legible to capital. The legible to capital meme is fantastic. I do think that will he made a new meme. Fantastic coinage. I love it. It's an absolute ripper. We'll be using it. Let me tell you about Fin AI, the number one AI agent for customer service. It's AI that handles your customer support. Timeline is in turmoil over Nucleus. Former guest on the show two or three times. Keon has founded Nucleus for IVF and he put up a subway campaign that says IQ is 50% genetic, height is 80% genetic. I completely disagree with that one. It's entirely skill based. For me the genes did not matter. I had to grind for this view. Grind my growth plates, I suppose. Have your best baby is what it says. And it says IVF done right in the subway all over New York City. There's a ton of debate going on. And to be clear, accurate, I think it was intentionally trying to make some percentage of the population angry to drive enough energy and attention. So this was. Yeah, I would call it, I would call it rage bait. So I would call it rage bait marketing. Not necessarily rage bait, not at the product level. But IVF as a category is a controversial category. And so it's much easier to wrap it in a campaign that will go viral for upsetting reasons. For you can upset people and you can get a lot of attention from that. This is an example from Kath Korvac. She says, so eugenics is profitable now, and so being able to wrap something that is just a, you know, a scientific process that's been worked on for. A long time seems to be somewhat. Friend.com inspired. Keon's original post. He says, nucleus Genomics announces the largest genetic optimization campaign ever. Which is, which is just funny because, you know, friend.
And it will sort of take longer. What about model routing? How important is that within the context of coding agents, Claude code just the surface area of what Anthropic is building. How many layers will this have over time? How are you thinking about the development of actually routing to the most efficient model? Because it sounds like it's happening within Opus 4.5, but then there are also times when you might want to go to just a different model entirely. Yeah. And it's similar there where I really think that ultimately something like routing is a little bit of a medium term hack, I guess one could say across different model sizes. Where ultimately you want everything to be like an end to end learnt system. Right. And it's similar to. I think we'll see a similar lesson as Tesla saw, where they're like, okay, actually everything's just one giant end to end LUT system as opposed to discrete components that have different purposes. But it takes time to get there. You said earlier you could imagine a scenario where labs would kind of hold back frontier models because they would be effectively handing their competitors an advantage. What's your timeline?
Case was dismissed and it's like one of five cases against you. Yeah, yeah. But in this case, they won a preliminary injunction, which means that the case is just still progressing and they still have to fight it. It's not, it's. A lawyer would, would file a preliminary injunction because they believe they had such a slam dunk case that they could prevent a lot of, a lot of like, basically going a lot further and spending more money in the case. And so a judge might say, hey, this is actually, it's not clear enough for me to make a decision right now. We're still going to proceed with the case and give both sides an opportunity to continue to make, to make their case. Yeah. And then there was a little bit of like a twist in the fact that Roy Lee, the founder of Clulee, apparently had worked at Nucleus, Nucleus Genomics. Very great. And Cremu posts a screenshot of Roy Lee back in February of 2025. So literally just like months before he started Cluli. Very grateful to Kian and the Nucleus genomic teams for taking a chance on me the summer before Columbia and introducing me to the startup world. If I've ever seen a trillion dollar company and team, it's Nucleus. And Cremieux says he's obviously lying to cover up after getting caught doing fraud. An additional piece of background information people should know is that this fraud also employed the guy behind Cluley, the cheating company. And so Cremieux is being very, very hardcore in his assessment. Just actually calling Kian a fraud straight up is. Is much more aggressive than just saying, like, you know, some of the, some of their claims are maybe not legitimate. It's unclear, like, you know, fraud is technically a crime that you need to be convicted of in the court before you are a fraud, but it's certainly he's putting his credibility on the line because if Keon comes out and says, like, yeah, I'm actually not a fraud, I did it. I proved it wrong. I mean, the main thing here is it appears that the customer reviews are potentially fixed.
Screens, and you can still. You can still watch TikTok, but you can at least act like you're off the grid. When I think about it, I was actually thinking about it this. This. This weekend, because I was out there and seeing all this, and there's a camper in every parking lot. And I do wonder if, like, dad's being more involved. When I was a kid, dad had a sports car. Not my dad, but in general, like, dad had a sport, he would use it to get away. You would take it on a drive and take it to a show, and. Sure. And now, like, dads and families want to do this stuff together. And you, sports cars, you can't do that with your kids. And I've even watched the vintage suv. I got a Ferrari FF to try to get around that when I had my first kid, and it did not work. The electrical issues were like, if I drive this thing to the farmer's market and I stop it, it might not start again. So I got to make sure I started at home and stop it. But I am watching, like, Grand Wagoneers, right? And old. And these are things you can now do with the whole family. And I think that there's been, like, sort of a renewed push among. Among. Among dads and among families to kind of do, hey, let's do the Vint thing together. And it's really. It's really become kind of a thing, and that's part of it, but I think there's probably many parts of it. Speaking of dads, speaking of firing dads, what's the best modern car that screams, I'm coming to town to fire your dad? We have this big thing on our podcast that, yeah, we think that there are some cars that just say, I'm here, but fire your dad. Such a. Like, it's such a big old. Like, this is. This is a rich guy who doesn't. The electric Rolls Royce, which is called the Spectre, that's a big one because it's got one of the things I like, fire your dad. Cars, to me, are cars where it takes up a lot of space but is unbelievably impractical. So that car is a massive vehicle, but it's got two doors. And that guy owns the factory. That guy, he comes in, he's like, I don't care what's. This whole factory is closed. Be gone. And he takes the Rolls Royce off. It's also so tall. We saw one at the airport next to a Urus, and we were like, it's a car that's than that SUV that's been lowered for performance. It's like.
Bring it in. That one was really cool and looked really cool. Although I have to say, just a couple weeks ago I reviewed the new Hyundai Crater, which is an off roader concept car. And from a perspective of a car that I think will sell, they should make that body on frame. Off roaders. Off roading has become such a thing in the last 10 years with the Raptor and the G wagon and the 4Runner and the Wrangler and the Bronco and off road trim levels of every single car in existence, including minivans. Hyundai has started to come out with some off road trim levels. But it's pretty clear to me that like there should be an. They should go after the 4Runner. Forerunner is selling pretty easily, millions, literally millions of units. And it sense to me that, that these automakers are not going after it more. I think Honda's insane for not pursuing it, but Hyundai and Kia have been much more aggressive and willing to be aggressive in these kind of new segments. And I think that Crater is previewing a car that will exist in some form. Probably not exact like that, with graphic displays on the windshield, but something sort of similar. I bet you they're cooking it up. Where do you think that trend came from?
To be ready to vlog the whole thing, which almost looks like he did a deal with authorities, which is like, hey, we're gonna arrest you. We're not. We need to make a big scene out of this. We're not gonna really throw the book at you, but we need to do this marketing for it. I think it was all legit. I think he is always ready to vlog. I think his boys are, like, always at the shot. You know what I mean? Yeah, yeah, I'm sure. I'm sure. But, yeah, boy, that was a big deal. So many people in the community do the Montana thing, and I really think he did a big thing about it. It's on our pod last week. I really think with the advent of license plate readers and with states just starting to realize how pervasive this has become, it's just really going to start becoming more and more difficult to do. And as a guy who's got four very expensive cars registered to my actual home addresses, I watched the Montana thing, and I think to myself, f you. If I have to pay, you have to pay. Exactly. I do feel bad. I mentioned this on the pod. I do feel bad for the guys who are doing it to get away from the stupid regulations. Like smog on a $15,000 car. Meanwhile, it's being ruined by guy after person after person doing it on expensive cars. There's a little hypocrisy.
Effect ready for tomorrow. Keller said that he's launched zipping points. He can pick up packages and deliver them autonomously with the zipline. Autonomous drones. This is the private plane for your burrito, folks. It's arrived. We're here. We're in the future. Future is here. The future plane, the flying car is here, and it will deliver you chipotle in 15 minutes. In four minutes while it's still warm. So here's a zip. Grabbing a package from one of our restaurant partners. It'll take so many cars off the road over the coming years. That's great news for environmentalists, for congestion, for anyone who wants to be able to really let it loose on the roads. If we're getting less congestion, maybe the speed limit goes up to 80 miles an hour, maybe 120, maybe 160. Maybe we get up to 200 and you can really let it loose. I would like that. I wonder if they'll need to, like, will you need to prove that you're at a certain address in order to get stuff delivered there? Because it's such a funny dimension to mess with people and just be like, hey, look on your lawn. And there's just like. There's like a burrito. A burrito. Just chilling. Well, people do that with pizzas, right? They prank call, I'd like a dozen pizzas delivered to this address. I'll pay in cash. This is like a famous prank. And then you show up and it's like, I don't need all these pizzas. I'm being pranked. I think that that's been mostly resolved by modern payment solutions, but I'm sure there will be oddities around these drones. Sheesh. Sheesh. Theo Vaughn, Flamin Hot Tweets is joking around and says, imagine how cool it will be to shoot one of these out of the sky to get a free meal. Going hunting for your Chipotle burrito. That, of course, is extremely cyberpunk and hilarious, but it will be massively illegal. And Keller breaks it down. He says, we're regulated by the FAA, so the consequences are similar to shooting at a 7. Taking off from the airport, not a good idea. Also, communities love the service. And I imagine he's not. He's not saying, like, the details, but if you shoot at a 737 as it's taking off from an airport, I think you're going to jail for a long time. And I think you will not just be able to shoot one of these out of the sky and pick up a free burrito. With a 22. But what about, like, a really big net? If you own the land, how high can you build a net? I think that would be the same as throwing a really big net at a 737 on takeoff. Tyler. I don't know. I don't think. I think for anything, the consequences are similar. Like, yes, it's illegal to shoot something out of the sky, but you're not. I think somebody could shoot one of these things down and get a misdemeanor. I think with a good lawyer, you're getting a misdemeanor. You're not going to. Whereas the 737, you know, if you. Get John Quinn on your side, you're eating burritos for free. Yeah. So David. David. David Chang was saying, like, it's just. There's like, he, He. It wasn't that he was, like, bearish on. On the tech, but he didn't think it solved, like, delivering it in the right form factor. Yeah, we gotta ask Keller about that tomorrow. Because right now, if you order food, depending on where you are between the time that it's cooked and picked up and delivered, it's like, there's a big gap. Well, speaking of coast, it has to be shorter. Speaking of cooked, there was a wild job description that hit the tiny.
Pretty photo real and so you want to stay safe out there. It's going to be Joe Weisenthal asked Nanobanana to create a really annoying LinkedIn profile. This is what I was talking about and I couldn't tell is this a real person? I have no idea because at this point we're way past the Turing test for images in the sense that this looks perfectly edited. But this could also just be a straight up screenshot. I would need to fact check this, but instead of fact checking it, I'm going to tell you about Privy. Privy make it easy to build on crypto rails, securely spin up white label wallet sign transactions and integrate on chain infrastructure all through one simple API. Kalyn Sterling I don't think is a real person says about I don't do small talk. I do deep dives. My journey is a quantum leap through the liminal spaces of tech and spirituality. Chief Visionary Officer TEDx Speaker Professional Storyteller Democratizing the metaverse one DAO at a time. 10X Growth Alchemist so did Joe actually ask nanobanana to do this? I want to see the prompt if that's true. Yeah, let's try to. Yeah, we need to replicate this. Try to replicate it because I actually did. So this was the example that I gave. I went to nanobanana and I said make a My prompt was just create a really annoying LinkedIn profile. But I forgot to check the Nano banana box. And so even though it was multimodal, according to Tyler over there, it did not generate it. It generated text. And so you know, oh, it's like I can only do text. Very questionable. We'll see. We'll keep playing around with that. And in the meantime, I'll tell you about adquick.com out of home advertising made easy and measurable plan.
Super impressive. Congrats to the whole team. We'll talk to you soon. Great to see you. See you. Ciao. Bye. Back to the show, back to the timeline, Back to linear. Meet the system for modern software development. Purpose built tool for planning and building products. There is more OpenAI news, of course, more tech news of all times. OpenAI's hardware division says Mark Gurman built around Jony I've secretive startup has ramped up the hiring of Apple engineers. The group has brought on about 40 new people in the last month or so, with many of them coming from Apple's hardware group. Yeah, hearing that Sholto interview, I'm disappointed. I don't think we're getting ads from Anthropic anytime soon and I don't think we're going to get a mobile device. Well, we are actually talking today to Quinn Slack, the CEO of AMP and sourcegraph. AMP is a Frontier coding agent and AMP is free. They introduced AMP Free which is ad supported and has a no cost mode. And so you can now use their coding agent for free with ads. 40 people. That does not seem like cause for concern for Apple. I mean I can't imagine how big their hardware group is, but it has to be in the thousands, I would imagine. Yeah, huge organization. So OpenAI is poaching left and right from Apple's hardware engineering group, hiring around 40 directors, managers and engineers in the last month from nearly every relevant Apple department. Mark Gurman says it's remarkable. So from what I've heard, this is Mark Gurman Apple is none too pleased about OpenAI's poaching and some consider it a problem. The hires include key directors, a fairly senior designation, as well as managers and engineers. And they hail from a wide range of areas. Camera engineering, iPhone hardware, Mac hardware, silicon device testing and reliability, industrial design, manufacturing, audio, smartwatches, Vision Pro development software. They got one from every single sampled. Every single, every single division, I suppose. Gemini is estimating that Apple has between 15,000 and 20,000 hardware engineers in total. 15,000. That seems like a lot. I don't know. In other words, OpenAI is picking up, picking people from nearly every relevant department. It's remarkable, says Mark Gurman. Very interesting. I wonder how the comp structured, how everything will come together on those teams. I mean there's a lot of people from Apple who going over to OpenAI. It's a greenfield project. It's probably really fun, probably really exciting, probably not the most mercenary scenario, but there's always that root. If you're working at Apple and you're excited about AI and you've been there for the last three years watching all this progress happen at the application layer and the model layer and not being thrilled with the progress happening at the hardware layer. This is like a, it's a wide open opportunity to be working right at that intersection of the models and the hardware. There's a lot of AI engineers who have made moves because they don't want to be a GPU poor company. And it's weird because Apple's in this, in this scenario where they're partnering with Gemini now, they're clearly going to survive. It's not a serious threat, at least not yet. Maybe if this device is incredible. But right now Apple looks pretty strong. The new iPhones are selling well, everything's good. But from an AI perspective, it's gotta be one of the worst gigs because you were in this sort of openly hostile environment to LLMs, to scaling, to building large GPU clusters and then yeah, they're sort of playing catch up now. But they're certainly not calling up Oracle for, you know, trillion dollars of compute. You go over to OpenAI, you're just going to be immersed in a lot more. Higher risk taking higher risk on. I wonder, yeah, Gabe is asking if wouldn't that be bearish as the hardware group at Apple is responsible for the terrible Vision Pro. I do wonder. It depends on who it was. It was a, it was an incredible technological feat. I just think they built. I was just watching a thing about A guy who 3D printed an adapter so you could use the strap from the Apple Vision Pro on the quest 3 from Meta. And that really speaks to the fact that the Apple Vision Pro, although it was too heavy, it has this screen on the outside that I don't think anyone wants. There were pieces about it that were clearly the best. The screen is just the best. No one's debating that. The band, the knit band is very cool. It has this amazing device where you rotate it and it tightens up. There are a whole bunch of things that are amazing. It's just as a package, it didn't deliver. But if you just want to, if you're, if your job is just like, hey, we got to put a screen on this and it's got to be the highest resolution screen. Like, well, go to the place that developed the highest resolution screens. Like, they did a good job. Well, Sam Altman replied to one of our cards we put up on November 22. TVPN posted on this day, Sama was rehired at OpenAI. Got his badge back and Sam Altman replied and said cannot believe this was only two years ago. Subjectively feels like five. Yeah, what a turnaround to go from defenestrated to back in the cedin so much and have so much control over the organization that you're able to raise at massive valuations, broker all these deals, move the entire market. Just a remarkable run put on such a masterclass in deal making that people are now sitting here being like there's no way that this would be a $500 billion company if Sam wasn't in the driver's seat. First let me tell you about fall build and deploy AI video and image models trusted by millions to power generative media at scale. Danny Zhou who founded the reasoning team in Google Brain, now part of the Gemini team at Google DeepMind. He says game over and carried no interest. Friend of the show quotes it and says I genuinely think OpenAI equals equals Yahoo. He's not assigning the variable, he's equating it. I migrated almost all my workflows code off their APIs now. Ironic that Google will probably do it twice lmao and I don't know about this. The pattern matching on the Yahoo example we should have him on the show. And actually yeah you were saying the other thing with the Yahoo example is is it wasn't like there wasn't like the company was valued at pretty tremendously for a longer period of time. It wasn't like this binary like one. Moment exactly like the peak market cap for Yahoo 125 billion during 2000. That feels like it's just hard to pattern match perfectly to this. But I mean it certainly would be poetic if that's the way it played out. It is funny Carried says ironic that Google will probably do it twice. They actually created the transformer. They released the transformer paper and chromium to inspire themselves to find harder sort of challenge giving everything a straight away give out all the alpha. Yep, yep for sure. Just kind of like find their fire again. Yeah there was an interesting article on Medium that was sort of burning up hacker news that I thought would be fun to go through. First let me tell you about graphite.dev, code review for the age of AI. Graphite helps teams on GitHub ship higher quality software faster. So this person who no one can really understand who this person is, they don't necessarily exist on the Internet fully. So there was like a question about that. This is like a very like you know, hacker news and turmoil segment but Taeha says I reverse engineered 200 AI startups, 146 are selling you repackaged ChatGPT and Claude with new UI. And so basically the thesis of this article is that this, this fellow wrote a piece of code that looks at the marketing copy and says, what are they claiming? And then looks at the calls that happen when you actually interact with their AI features. So if there's a chatbot on this particular startup's website and you are near chatting with it, and you look into the trace that's happening in Chrome, is it going to the startup server or is it going to OpenAI server, or is it going to Anthropic server? That's telling. And then there's also a little bit of API fingerprinting. Basically, OpenAI has a specific pattern of rate limiting and it's exponential. So if you're spamming the OpenAI API, it will, according to a unique pattern, tell you, hey, you've sent too many messages, cool off for one minute, and then the next time you do it, cool off for two minutes. The next time, cool off for four minutes, then eight minutes, then 16. Right. And it gets exponentially longer and you're on progressively more longer and longer timeouts. But the shape of that curve and the specific timings are unique to OpenAI. And so if I'm a startup and I have the exact same like, back off and timeout curve, well, then it's probably just OpenAI under the hood. At least that's the claim that's being made here. And so the finding in this article is that 73% had a significant gap between the claimed technology and the actual implementation. And so out of the 200 AI startups that this fellow analyzed, 54 companies either had accurate technical claims. They said, hey, we're using like, we have a custom AI model that we trained. And they did. Or they're transparent about their stack. They say, hey, this is a wrapper. Like, we're a wrapper company. And so, you know, our AI is powered by ChatGPT, we're partnered with OpenAI, we're partnered with Anthropic or whatever. Now, 146 companies, that's 73%, according to him, were sort of misrepresenting their technology. So either they said they had proprietary AI, proprietary AI. And yet when he dug into it, it was OpenAI API plus prompts. Tyler. Yeah, I mean, it's like kind of what do people expect? Like, is if you fine tune. If you use the OpenAI API to fine tune the model. Yes. Which you can Do? Yes. Is that proprietary? Like no one else has that fine tune? Yes. You're still calling the API? It's like I don't expect startups to train their own full language models. That's like pretty unrealistic and like really make sense. Yeah. So I'm kind of confused, I guess. Yeah, I guess it's a very cool study. But this tracks with exactly like I would guess that 73% of AI startups are just reskinning. Yes. And so 19% of the overall companies, the 38 that were analyzed in this study found that the startup said they had in house models and it was actually fine tuned public models. And so there's a question, is like whose house is it in? It's technically in OpenAI's house. So fine tune, as you mentioned, an open source model that's public, does that count as a public model? An open source model? Let's assume yes. And then last 8% would. So they had a custom ML pipeline and they were in fact using standard cloud services. That's even wishy washier. I think that's totally fine because like you can totally have a custom ML pipeline that's wiring together OpenAI and Gemini and AWS and you know, a bunch of other. If I'm using a startup, I don't want them to train their own language model because I don't think they're going to like in 99.99% of the case, like they're not going to be able to do a better model than OpenAI, Anthropic, Gemini Grok. Like I want them to use the best model. Yes. And it's like, okay, you can fine tune it or whatever. I agree, that's totally fine. And so the author also agrees with you. He says, here's what really shocked me. I'm not even mad about it. Every time I saw the phrase our proprietary language model, I knew what I was going to find. And I was right. 34 out of 37 times. And this is where it gets weird because he says, here's the technical signature. And so the user submits the query, it posts to API generate and then with wrapper logic it posts to API.OpenAI.comv1chat completions. And I have no idea how he's seeing the backend. It makes no sense how he would be able to do this unless there was just like a massive security vulnerability. Because what I would assume is happening is that the user's over here, the startup's website's here, and then the user goes to the startup's website. And then the startup's website on the back end talks to OpenAI and comes back and maybe you could understand that, like, okay, the amount of EM dashes, like there's a. This is telltale signal. But that's not what he's doing. He's saying that he was able to just literally hit like the Chrome Inspect developer tools. Look at the chain of calls and see that it was calling OpenAI from the front end, which is crazy because I didn't even know you could do that. It feels like if you were calling it directly from the front end, you would potentially leak a key that would be able to put you on the hook for a bunch of bills. I would think you would want to authenticate that on the back end. And so he gives a bunch of examples of like rag. And then he's exposing some margins, which is actually very bulky bullish for these companies because he breaks down some of these and says that a GPT4 API is $0.03 per thousand input token, $0.06 per thousand output token. So the cost per query for this hypothetical startup was $0.03 and they charged $3 or $300 per month for 200 queries. And so the wrapper economy, this is 75 times direct costs. That's extremely bullish for that printing for that company. He found another one that was maybe 1000x API costs that's doing some pinecone embedding. And he also says this is pattern number three. The quote, we fine tuned our own model. Reality check. Fine tuning sounds impressive and it can be, but here's what I found. 45% it was OpenAI's fine tuning API, which. That sounds right, right? It's a little bit of a step to be like, we fine tuned our own model. It's like, no, you fine tuned OpenAI's model and then you got your own model from that result. It's a little bit. Yeah, you're still fine tuning it. I don't think there's that big a difference between fine tuning. Are you fine tuning it or is OpenAI fine tuning it? There's a fine tuning API which you use to. And then. But who's doing the fine tuning, you or OpenAI? Well, what do you like? You're not interfacing directly with the gpu. It's like I went to the store and I bought a sandwich. Who made the sandwich? Well, I told them what the used turkey. I told them, put lettuce, extra pickle. Who made the turkey? I think it's more like you go to the store and you bring a sandwich to the office. It's like where did the sandwich come from? It came from the store. But like you brought the sandwich, it came from you. You brought the sandwich. Yeah, this is better. I think you're right. Anyway, 22% of the time it was a hugging face model with Alora. 18% of the time was anthropic Claude with a prompt library. 8% of the time is literally just GPT4 with system prompts. And 7% they actually trained something from scratch. And all of these are odd. And there's a lot of deb how this would actually happen because he's basically saying just open dev tools, go to the network tab, interact with the AI feature. If you see API OpenAI API Anthropic or API cohere AI, you're looking at a wrapper. They might have middleware but the AI isn't theirs. And so it just opens up this debate about what is the value of the wrapper. I mean certainly if you can resell something for 100x because you have some sort of clever workflow prompt or workflow, more power to you. Yeah, it's not exact, it's not exactly like bearish on the company's yet. But the debate that was surrounding was more around so, so this, this author claims that after posting this seven founders reached out privately. Some were defensive, some were grateful. They asked for help transitioning their marketing from proprietary AI to built with the best in class APIs because some of these founders did, did I guess feel like using proprietary AI as a marketing tagline was disingenuous. And then someone else, I think I saw something that was one VC reached out and said like I'd like you to audit my portfolio because I have been told that I was investing in companies that were training their own AI and I made the investment on that assumption. And if I'm being lied to then that's potentially securities fraud. And so there is a question about if you go, I mean I've seen pitches for companies that, where they've said like proudly like you should invest in this because we're not training our own model, it would actually be a mistake. And there's another company that's a competitor to us that is training their own model and you don't want to invest in them, you want to invest in us because we're going to burn your dollars. Yeah, we're going to much better economically. And so all of it just the only thing that matters is like being upfront with the investor for sure. And then to some degree you do need to be upfront with the customer because if the customer there is a marketing value to. Oh, if you work with us, you're working with these genius AI scientists who are going to build their own models. And if it's just repackaged ChatGPT, that might not be what you want to pay for because at that point you might just say, hey, like actually if I can just get this directly from opening, I'll just go buy it from them anyway. Well, thoughts and prayers to friend of the show, John Palmer. He says he just found out my wife is leaving me. She said I'm not legible to capital. The legible to capital meme is fantastic. I do think that will he made a new meme. Fantastic coinage. I love it. It's an absolute ripper. We'll be using it well. Let me tell you about Fin AI, the number one AI agent for customer service. It's AI that handles your customer support. Timeline is in turmoil over Nucleus. Former guest on the show two or three times. Keon has founded Nucleus for IVF and he put up a subway campaign that says IQ is 50% genetic, height is 80% genetic. I completely disagree with that one. It's entirely skill based. For me, the genes did not matter. I had to grind for this view. Grind my growth plates, I suppose. Have your best baby is what it says. And it says IVF done right in the subway all over New York City. There's a ton of debate going on. And to be clear, accurate, I think it was intentionally trying to make some percentage of the population angry to drive enough energy and attention. So this was. Yeah, I would call it, I would call it rage bait. So I would call it rage bait marketing. Not necessarily rabbit product level, but IVF as a category is a controversial category and so it's much easier to wrap it in a campaign that will go viral for upsetting reasons. For you can upset people and you can get a lot of attention from that. This is an example from Kath Korvac. She says, so Eugenics is profitable now. And so being able to wrap something that is just a, you know, a scientific process that's been worked on for. A long time seems to be somewhat friend.com inspired. Yep. Keon's original post. He says, Nucleus Genomics announces the largest genetic optimization campaign ever. Which is, which is just funny because a friend was saying this is the largest out of home campaign ever. And now Keon is saying this is the largest genetic optimization campaign ever. So Narrowing it down, but full station Blitz at Broadway. 1000 plus street ads across New York City, 1000 plus subway car ads. Dozens of urban panels throughout SoHo. And apparently they're not actually. They're not able to offer the service in New York. Saw that in here. So it's really just an image of a controversial phrase on a New York subway is more likely to go viral. So you do it there because it looks like you're on the global stage and then you pull in. There's a high density of people that have a large following audience. Yeah, following. And so it's just the way to start a viral trend and own the moment. It's the reason why, you know, so many tiktokers are in Manhattan now doing stuff like man on the street stuff. It just like, it has more, like, aura almost. Well, Dr. Shelby liked the mindshare grabbing that Nucleus did. Says every biotech founder should be seeing this and understanding how to get one tenth the mindshare of Nucleus. I have a playbook for you below. A lot of people are like, I love the playbook. I don't love this example because the company's getting dragged. I don't know if it's good or bad with the rage bait thing. I think usually it's a negative thing. Usually it's hard to come back from. Occasionally it can be done in a way that's slightly enraging, but enough people are in on the joke that they appreciate what's happening and they appreciate that it breaks through or it's enraging to someone who's not the core audience, not the actual customer. And so it's okay. But it's a big debate because Sichwan Mala posted a long essay all about the claims made by Nucleus. Keon says everything levied unto Nucleus by Sichuan Mala is false. Worse than false. It appears to be architected by a competitor that has repeatedly published misstatements and inaccuracies. Sichuan is compromised, but it gets worse. Yeah. To be clear, no evidence has been provided that it was being levied by a competitor. Yes. That's purely an allegation that has no. There's no proof. Yes, yes, exactly. Yeah. Sometimes. Sometimes there's DMs that leak and. And there's evidence. Or someone comes forward and. And says, like, yeah, I was actually paid to. To post that. But so he says, I've been. I've been informed that Cremio. I don't know how to pronounce that last thing. Crimeo, who's been on the show also he claims he's a race. Scientist in chief has been paid off by the competitor to promote this nonsense against Nucleus for the independent scientists repeating the denied. Camille denied that as well. I would encourage you to do more diligence on who you're aligning yourself with. Our scientific team will issue a point by point response, which I believe they did. Unfortunately though, this isn't about science. It's about. It's a concentrated attempt to cancel Nucleus on the backs of our successful campaign and build in efforts to build and advance the industry which benefits the very people attacking us. The mob are trying to cancel Nucleus. Keep tweeting, stay mad, we'll keep building. And serving patients. P.S. we won the injunction. Link below. So they were sued by their competitor, but. So they won the injunction, but they didn't mean they won the case at all. I mean, that's a classic thing if you're getting sued to be like, the case was dismissed and it's like one of five cases against you. Yeah, but in this case, they won a preliminary injunction, which means that the case is just still progressing and they still have to fight it. It's not. It's. A lawyer would file a preliminary injunction because they believe they had such a slam dunk case that they could prevent a lot of like, basically going a lot further and spending more money in the case. And so a judge might say, hey, this is actually, it's not clear enough for me to make a decision right now. We're still going to proceed with the case and give both sides an opportunity to continue to make their case. And then there was a little bit of like a twist in the fact that Roy Lee, the founder of Clulee, apparently had worked at Nucleus Genomics. Very great. And Cremieu posts a screenshot of Roy Lee back in February of 2025. So literally just like months before he started Cluli. Very grateful to Kian and the Nucleus genomic teams for taking a chance on me the summer before Columbia and introducing me to the startup world. If I've seen a trillion dollar company in team, it's Nucleus. And Cremieu says he's obviously lying to cover up after getting caught doing fraud. An additional piece of background information people should know is that this fraud also employed the guy behind Cluley, the cheating company. And so Cremieu is being very hardcore in his assessment. Just actually calling Kiyan a fraud straight up is much more aggressive than just saying some of the. Some of their claims are maybe not legitimate. It's unclear. Like, you know, fraud is technically A crime that you need to be convicted of in the court before you are a fraud. But it's certainly, it's certainly he's putting his credibility on the line because if, if, if Keon comes out and says, like, yeah, I'm actually not a fraud, I did it, I proved it. I mean, the main thing looks really bad. Main thing here is, is it appears that the customer reviews. Yes. Are potentially fictitious. And if you're selling a service that allows people to pick their baby and you're giving and you're showing reviews from happy customers that may or may not be real people at all, like that just feels deeply wrong. So I think that one of the first things that they could have done, I don't believe they have, is just say like, no, our reviews are real. We used AI imagery because the people, the real people didn't want their identity online tied to this service. Right. Yeah. For privacy reasons. But I haven't seen anything, I haven't seen anything like this. This guy Adi had a good point. He said, one of the core tensions in this industry is the fact that most companies recognize they're working on an incredibly sensitive topic. They know the general population will need to be slowly and tactfully acclimated to the idea of advanced family planning. Nucleus is perceived as polluting the commons with their deliberately inflammatory marketing. Their virality comes at the cost of increased skepticism for the whole industry. So yeah, a lot of, a lot of folks were not very happy about that. Keon has replied. And if you want to dig into the actual scientific claims on either side, there are long posts where you can go through there. But obviously AI generated blog posts are alleged plagiarism in the Nucleus origin, white paper errors in there, blatant falsification, terms of service are contradictory. Yeah, they also apparently they hired two people that had a non competes for 18 months. Those people just immediately started on working on Nucleus. Nucleus claimed that they weren't competitive so that the non compete didn't apply. But if you look at the companies and what they offer, it seems very clear that they are competing. So anyways, very messy, very messy, messy story. And yeah, I don't know. Will o' Brien says. David, I'm so sorry now man, but you guys are doing an absolutely terrible job at responding to this blog post and seem to be missing the point here. First of all, it is a huge claim to say that every claim by Sichuan Mala is false with absolutely zero evidence for explanation share receipts. Second of all, you make the claim that the person is paid off by competitors of yours, again, with zero evidence. Third, that you make the claim that CREMU is paid off by your competitor. This is bogus and not true. But most importantly, you guys have made zero points of substance here. Rather than just insinuating you guys are leaping ahead and others are jealous. You are selling a scientific product and someone has made a scientific critique in good faith, waiting to be corrected and explicitly saying they will make changes if they are proved wrong. And the best you guys can do is accuse them of being paid off and reply with memes. Not a great look. I want to see startups of the bold vision succeed, but how you communicate with the broader world is so important, especially with a product like yours. And how you guys are carrying on in this, honestly, pretty lackluster. So, yeah, again, I don't. Yeah, at this point, Keon. Keon's been on the show. He's very funny, high energy. We've had some enjoyable conversations. But if I'm a potential customer of Nucleus at this point and I see just these series of exchanges, I'm certainly going to wait and see. See how things evolve versus signing up to use the service to. It's just so different. Cluli. It's so different than Cluli because if I use clulee and I'm like, oh, the notes that were taken in that meeting weren't that good. Or if you go into Cluly being like, I'm going to cheat on this test, and then it's like, oh, it didn't work. Their engineers aren't good enough to really help you cheat on that test. You're like, okay, well, I probably should have been cheating on that test. It's like the lowest stakes thing possible. But when you're, when you're. This is like literally deciding who your child offspring will be. It's the highest. The quality of the product could very well, like contribute echo for generations. Literally. Literally generations. Exactly. Not even just the child's life. Yes, the life of the child's child's child. Yes, it is extremely child's child's. Child's child's. It could alter the course of history. I mean, it kind of could. It's sort of crazy. So, yeah, I mean, it's hard because viral marketing does work. Like, you know, moving fast and breaking things does work in certain contexts, but in the bio, in bloodline optimization, it's really, really high stakes. And so you gotta be extra, extra careful. Extra careful for sure. Well, let me tell you about profound. Get your brand mentioned in ChatGPT, reach millions of consumers who use AI to discover new products and brands. Will Brown has a funny post here. He's just got a recruiting email from a company explicitly mentioning that they have 75th percentile comp. That's so funny to me. He says we're assembling a B team and have raised an okayish amount of money from pretty good investors. It's so good. Somebody should actually run this. Yeah. I love Prime Intellect. There is such a fun crew over there. We got to have them back on the show soon. I think that they're. I won't leak anything, but I think there'll be some news soon. Hopefully. Hopefully. I'm very excited for them. Yeah. We got to hang with Vincent Saturday. So OpenAI has an announcement. They're introducing Shopping Research, a new experience in ChatGPT that does the research to help you find the right products. They clearly were listening to me on the show just a few days ago when I was saying I would be using this for the holiday shopping period. Very exciting. I wonder how it will actually play out. You, of course, had that problem with cars and bids. ChatGPT was not identifying the fact that that GT3Rs had been sold two years ago. Thankfully, we have Doug Jumeiro here in the restream waiting room about to join the show. We can talk about cars and bits. We can talk about cars, we can talk about artificial intelligence. Welcome. Welcome to the show. How are you doing? I'm good. Thank you for having me. Gentlemen, thank you so much. So great to have you on. We are a technology and business show, but we have. Your name has come up probably 100 times independently, just when we're talking about 100% cars. So it's so great to have you on the show. Thank you. We're huge fans. Thank you. Thank you. I'm thrilled to be here. I really am. Thanks so much. I'd love to know. I mean, we were just talking about this OpenAI shopping research in ChatGPT. It feels like there's really no substitute for a.
Super impressive. Congrats to the whole team. We'll talk to you soon. Great to see you. See you. Ciao. Bye. Back to the show, back to the timeline, Back to linear. Meet the system for modern software development. Purpose built tool for planning and building products. There is more OpenAI news, of course, more tech news of all times. OpenAI's hardware division says Mark Gurman built around Jony I've secretive startup has ramped up the hiring of Apple engineers. The group has brought on about 40 new people in the last month or so, with many of them coming from Apple's hardware group. Yeah, hearing that Sholto interview, I'm disappointed. I don't think we're getting ads from Anthropic anytime soon and I don't think we're going to get a mobile device. Well, we are actually talking today to Quinn Slack, the CEO of AMP and sourcegraph. AMP is a Frontier coding agent and AMP is free. They introduced AMP Free which is ad supported and has a no cost mode. And so you can now use their coding agent for free with ads. 40 people. That does not seem like cause for concern for Apple. I mean I can't imagine how big their hardware group is, but it has to be in the thousands, I would imagine. Yeah, huge organization. So OpenAI is poaching left and right from Apple's hardware engineering group, hiring around 40 directors, managers and engineers in the last month from nearly every relevant Apple department. Mark Gurman says it's remarkable. So from what I've heard, this is Mark Gurman Apple is none too pleased about OpenAI's poaching and some consider it a problem. The hires include key directors, a fairly senior designation, as well as managers and engineers. And they hail from a wide range of areas. Camera engineering, iPhone hardware, Mac hardware, silicon device testing and reliability, industrial design, manufacturing, audio, smartwatches, Vision Pro development software. They got one from every single sampled. Every single, every single division, I suppose. Gemini is estimating that Apple has between 15,000 and 20,000 hardware engineers in total. 15,000. That seems like a lot. I don't know. In other words, OpenAI is picking up, picking people from nearly every relevant department. It's remarkable, says Mark Gurman. Very interesting. I wonder how the comp structured, how everything will come together on those teams. I mean there's a lot of people from Apple who going over to OpenAI. It's a greenfield project. It's probably really fun, probably really exciting, probably not the most mercenary scenario, but there's always that root. If you're working at Apple and you're excited about AI and you've been there for the last three years watching all this progress happen at the application layer and the model layer and not being thrilled with the progress happening at the hardware layer. This is like a, it's a wide open opportunity to be working right at that intersection of the models and the hardware. There's a lot of AI engineers who have made moves because they don't want to be a GPU poor company. And it's weird because Apple's in this, in this scenario where they're partnering with Gemini now, they're clearly going to survive. It's not a serious threat, at least not yet. Maybe if this device is incredible. But right now Apple looks pretty strong. The new iPhones are selling well, everything's good. But from an AI perspective, it's gotta be one of the worst gigs because you were in this sort of openly hostile environment to LLMs, to scaling, to building large GPU clusters and then yeah, they're sort of playing catch up now. But they're certainly not calling up Oracle for, you know, trillion dollars of compute. You go over to OpenAI, you're just going to be immersed in a lot more. Higher risk taking higher risk on. I wonder, yeah, Gabe is asking if wouldn't that be bearish as the hardware group at Apple is responsible for the terrible Vision Pro. I do wonder. It depends on who it was. It was a, it was an incredible technological feat. I just think they built. I was just watching a thing about A guy who 3D printed an adapter so you could use the strap from the Apple Vision Pro on the quest 3 from Meta. And that really speaks to the fact that the Apple Vision Pro, although it was too heavy, it has this screen on the outside that I don't think anyone wants. There were pieces about it that were clearly the best. The screen is just the best. No one's debating that. The band, the knit band is very cool. It has this amazing device where you rotate it and it tightens up. There are a whole bunch of things that are amazing. It's just as a package, it didn't deliver. But if you just want to, if you're, if your job is just like, hey, we got to put a screen on this and it's got to be the highest resolution screen. Like, well, go to the place that developed the highest resolution screens. Like, they did a good job. Well, Sam Altman replied to one of our cards we put up on November 22. TVPN posted on this day, Sama was rehired at OpenAI. Got his badge back and Sam Altman replied and said cannot believe this was only two years ago. Subjectively feels like five. Yeah, what a turnaround to go from defenestrated to back in the cedin so much and have so much control over the organization that you're able to raise at massive valuations, broker all these deals, move the entire market. Just a remarkable run put on such a masterclass in deal making that people are now sitting here being like there's no way that this would be a $500 billion company if Sam wasn't in the driver's seat. First let me tell you about fall build and deploy AI video and image models trusted by millions to power generative media at scale. Danny Zhou who founded the reasoning team in Google Brain, now part of the Gemini team at Google DeepMind. He says game over and carried no interest. Friend of the show quotes it and says I genuinely think OpenAI equals equals Yahoo. He's not assigning the variable, he's equating it. I migrated almost all my workflows code off their APIs now. Ironic that Google will probably do it twice lmao and I don't know about this. The pattern matching on the Yahoo example we should have him on the show. And actually yeah you were saying the other thing with the Yahoo example is is it wasn't like there wasn't like the company was valued at pretty tremendously for a longer period of time. It wasn't like this binary like one. Moment exactly like the peak market cap for Yahoo 125 billion during 2000. That feels like it's just hard to pattern match perfectly to this. But I mean it certainly would be poetic if that's the way it played out. It is funny Carried says ironic that Google will probably do it twice. They actually created the transformer. They released the transformer paper and chromium to inspire themselves to find harder sort of challenge giving everything a straight away give out all the alpha. Yep, yep for sure. Just kind of like find their fire again. Yeah there was an interesting article on Medium that was sort of burning up hacker news that I thought would be fun to go through. First let me tell you about graphite.dev, code review for the age of AI. Graphite helps teams on GitHub ship higher quality software faster. So this person who no one can really understand who this person is, they don't necessarily exist on the Internet fully. So there was like a question about that. This is like a very like you know, hacker news and turmoil segment but Taeha says I reverse engineered 200 AI startups, 146 are selling you repackaged ChatGPT and Claude with new UI. And so basically the thesis of this article is that this, this fellow wrote a piece of code that looks at the marketing copy and says, what are they claiming? And then looks at the calls that happen when you actually interact with their AI features. So if there's a chatbot on this particular startup's website and you are near chatting with it, and you look into the trace that's happening in Chrome, is it going to the startup server or is it going to OpenAI server, or is it going to Anthropic server? That's telling. And then there's also a little bit of API fingerprinting. Basically, OpenAI has a specific pattern of rate limiting and it's exponential. So if you're spamming the OpenAI API, it will, according to a unique pattern, tell you, hey, you've sent too many messages, cool off for one minute, and then the next time you do it, cool off for two minutes. The next time, cool off for four minutes, then eight minutes, then 16. Right. And it gets exponentially longer and you're on progressively more longer and longer timeouts. But the shape of that curve and the specific timings are unique to OpenAI. And so if I'm a startup and I have the exact same like, back off and timeout curve, well, then it's probably just OpenAI under the hood. At least that's the claim that's being made here. And so the finding in this article is that 73% had a significant gap between the claimed technology and the actual implementation. And so out of the 200 AI startups that this fellow analyzed, 54 companies either had accurate technical claims. They said, hey, we're using like, we have a custom AI model that we trained. And they did. Or they're transparent about their stack. They say, hey, this is a wrapper. Like, we're a wrapper company. And so, you know, our AI is powered by ChatGPT, we're partnered with OpenAI, we're partnered with Anthropic or whatever. Now, 146 companies, that's 73%, according to him, were sort of misrepresenting their technology. So either they said they had proprietary AI, proprietary AI. And yet when he dug into it, it was OpenAI API plus prompts. Tyler. Yeah, I mean, it's like kind of what do people expect? Like, is if you fine tune. If you use the OpenAI API to fine tune the model. Yes. Which you can Do? Yes. Is that proprietary? Like no one else has that fine tune? Yes. You're still calling the API? It's like I don't expect startups to train their own full language models. That's like pretty unrealistic and like really make sense. Yeah. So I'm kind of confused, I guess. Yeah, I guess it's a very cool study. But this tracks with exactly like I would guess that 73% of AI startups are just reskinning. Yes. And so 19% of the overall companies, the 38 that were analyzed in this study found that the startup said they had in house models and it was actually fine tuned public models. And so there's a question, is like whose house is it in? It's technically in OpenAI's house. So fine tune, as you mentioned, an open source model that's public, does that count as a public model? An open source model? Let's assume yes. And then last 8% would. So they had a custom ML pipeline and they were in fact using standard cloud services. That's even wishy washier. I think that's totally fine because like you can totally have a custom ML pipeline that's wiring together OpenAI and Gemini and AWS and you know, a bunch of other. If I'm using a startup, I don't want them to train their own language model because I don't think they're going to like in 99.99% of the case, like they're not going to be able to do a better model than OpenAI, Anthropic, Gemini Grok. Like I want them to use the best model. Yes. And it's like, okay, you can fine tune it or whatever. I agree, that's totally fine. And so the author also agrees with you. He says, here's what really shocked me. I'm not even mad about it. Every time I saw the phrase our proprietary language model, I knew what I was going to find. And I was right. 34 out of 37 times. And this is where it gets weird because he says, here's the technical signature. And so the user submits the query, it posts to API generate and then with wrapper logic it posts to API.OpenAI.comv1chat completions. And I have no idea how he's seeing the backend. It makes no sense how he would be able to do this unless there was just like a massive security vulnerability. Because what I would assume is happening is that the user's over here, the startup's website's here, and then the user goes to the startup's website. And then the startup's website on the back end talks to OpenAI and comes back and maybe you could understand that, like, okay, the amount of EM dashes, like there's a. This is telltale signal. But that's not what he's doing. He's saying that he was able to just literally hit like the Chrome Inspect developer tools. Look at the chain of calls and see that it was calling OpenAI from the front end, which is crazy because I didn't even know you could do that. It feels like if you were calling it directly from the front end, you would potentially leak a key that would be able to put you on the hook for a bunch of bills. I would think you would want to authenticate that on the back end. And so he gives a bunch of examples of like rag. And then he's exposing some margins, which is actually very bulky bullish for these companies because he breaks down some of these and says that a GPT4 API is $0.03 per thousand input token, $0.06 per thousand output token. So the cost per query for this hypothetical startup was $0.03 and they charged $3 or $300 per month for 200 queries. And so the wrapper economy, this is 75 times direct costs. That's extremely bullish for that printing for that company. He found another one that was maybe 1000x API costs that's doing some pinecone embedding. And he also says this is pattern number three. The quote, we fine tuned our own model. Reality check. Fine tuning sounds impressive and it can be, but here's what I found. 45% it was OpenAI's fine tuning API, which. That sounds right, right? It's a little bit of a step to be like, we fine tuned our own model. It's like, no, you fine tuned OpenAI's model and then you got your own model from that result. It's a little bit. Yeah, you're still fine tuning it. I don't think there's that big a difference between fine tuning. Are you fine tuning it or is OpenAI fine tuning it? There's a fine tuning API which you use to. And then. But who's doing the fine tuning, you or OpenAI? Well, what do you like? You're not interfacing directly with the gpu. It's like I went to the store and I bought a sandwich. Who made the sandwich? Well, I told them what the used turkey. I told them, put lettuce, extra pickle. Who made the turkey? I think it's more like you go to the store and you bring a sandwich to the office. It's like where did the sandwich come from? It came from the store. But like you brought the sandwich, it came from you. You brought the sandwich. Yeah, this is better. I think you're right. Anyway, 22% of the time it was a hugging face model with Alora. 18% of the time was anthropic Claude with a prompt library. 8% of the time is literally just GPT4 with system prompts. And 7% they actually trained something from scratch. And all of these are odd. And there's a lot of deb how this would actually happen because he's basically saying just open dev tools, go to the network tab, interact with the AI feature. If you see API OpenAI API Anthropic or API cohere AI, you're looking at a wrapper. They might have middleware but the AI isn't theirs. And so it just opens up this debate about what is the value of the wrapper. I mean certainly if you can resell something for 100x because you have some sort of clever workflow prompt or workflow, more power to you. Yeah, it's not exact, it's not exactly like bearish on the company's yet. But the debate that was surrounding was more around so, so this, this author claims that after posting this seven founders reached out privately. Some were defensive, some were grateful. They asked for help transitioning their marketing from proprietary AI to built with the best in class APIs because some of these founders did, did I guess feel like using proprietary AI as a marketing tagline was disingenuous. And then someone else, I think I saw something that was one VC reached out and said like I'd like you to audit my portfolio because I have been told that I was investing in companies that were training their own AI and I made the investment on that assumption. And if I'm being lied to then that's potentially securities fraud. And so there is a question about if you go, I mean I've seen pitches for companies that, where they've said like proudly like you should invest in this because we're not training our own model, it would actually be a mistake. And there's another company that's a competitor to us that is training their own model and you don't want to invest in them, you want to invest in us because we're going to burn your dollars. Yeah, we're going to much better economically. And so all of it just the only thing that matters is like being upfront with the investor for sure. And then to some degree you do need to be upfront with the customer because if the customer there is a marketing value to. Oh, if you work with us, you're working with these genius AI scientists who are going to build their own models. And if it's just repackaged ChatGPT, that might not be what you want to pay for because at that point you might just say, hey, like actually if I can just get this directly from opening, I'll just go buy it from them anyway. Well, thoughts and prayers to friend of the show, John Palmer. He says he just found out my wife is leaving me. She said I'm not legible to capital. The legible to capital meme is fantastic. I do think that will he made a new meme. Fantastic coinage. I love it. It's an absolute ripper. We'll be using it well. Let me tell you about Fin AI, the number one AI agent for customer service. It's AI that handles your customer support. Timeline is in turmoil over Nucleus. Former guest on the show two or three times. Keon has founded Nucleus for IVF and he put up a subway campaign that says IQ is 50% genetic, height is 80% genetic. I completely disagree with that one. It's entirely skill based. For me, the genes did not matter. I had to grind for this view. Grind my growth plates, I suppose. Have your best baby is what it says. And it says IVF done right in the subway all over New York City. There's a ton of debate going on. And to be clear, accurate, I think it was intentionally trying to make some percentage of the population angry to drive enough energy and attention. So this was. Yeah, I would call it, I would call it rage bait. So I would call it rage bait marketing. Not necessarily rabbit product level, but IVF as a category is a controversial category and so it's much easier to wrap it in a campaign that will go viral for upsetting reasons. For you can upset people and you can get a lot of attention from that. This is an example from Kath Korvac. She says, so Eugenics is profitable now. And so being able to wrap something that is just a, you know, a scientific process that's been worked on for. A long time seems to be somewhat friend.com inspired. Yep. Keon's original post. He says, Nucleus Genomics announces the largest genetic optimization campaign ever. Which is, which is just funny because a friend was saying this is the largest out of home campaign ever. And now Keon is saying this is the largest genetic optimization campaign ever. So Narrowing it down, but full station Blitz at Broadway. 1000 plus street ads across New York City, 1000 plus subway car ads. Dozens of urban panels throughout SoHo. And apparently they're not actually. They're not able to offer the service in New York. Saw that in here. So it's really just an image of a controversial phrase on a New York subway is more likely to go viral. So you do it there because it looks like you're on the global stage and then you pull in. There's a high density of people that have a large following audience. Yeah, following. And so it's just the way to start a viral trend and own the moment. It's the reason why, you know, so many tiktokers are in Manhattan now doing stuff like man on the street stuff. It just like, it has more, like, aura almost. Well, Dr. Shelby liked the mindshare grabbing that Nucleus did. Says every biotech founder should be seeing this and understanding how to get one tenth the mindshare of Nucleus. I have a playbook for you below. A lot of people are like, I love the playbook. I don't love this example because the company's getting dragged. I don't know if it's good or bad with the rage bait thing. I think usually it's a negative thing. Usually it's hard to come back from. Occasionally it can be done in a way that's slightly enraging, but enough people are in on the joke that they appreciate what's happening and they appreciate that it breaks through or it's enraging to someone who's not the core audience, not the actual customer. And so it's okay. But it's a big debate because Sichwan Mala posted a long essay all about the claims made by Nucleus. Keon says everything levied unto Nucleus by Sichuan Mala is false. Worse than false. It appears to be architected by a competitor that has repeatedly published misstatements and inaccuracies. Sichuan is compromised, but it gets worse. Yeah. To be clear, no evidence has been provided that it was being levied by a competitor. Yes. That's purely an allegation that has no. There's no proof. Yes, yes, exactly. Yeah. Sometimes. Sometimes there's DMs that leak and. And there's evidence. Or someone comes forward and. And says, like, yeah, I was actually paid to. To post that. But so he says, I've been. I've been informed that Cremio. I don't know how to pronounce that last thing. Crimeo, who's been on the show also he claims he's a race. Scientist in chief has been paid off by the competitor to promote this nonsense against Nucleus for the independent scientists repeating the denied. Camille denied that as well. I would encourage you to do more diligence on who you're aligning yourself with. Our scientific team will issue a point by point response, which I believe they did. Unfortunately though, this isn't about science. It's about. It's a concentrated attempt to cancel Nucleus on the backs of our successful campaign and build in efforts to build and advance the industry which benefits the very people attacking us. The mob are trying to cancel Nucleus. Keep tweeting, stay mad, we'll keep building. And serving patients. P.S. we won the injunction. Link below. So they were sued by their competitor, but. So they won the injunction, but they didn't mean they won the case at all. I mean, that's a classic thing if you're getting sued to be like, the case was dismissed and it's like one of five cases against you. Yeah, but in this case, they won a preliminary injunction, which means that the case is just still progressing and they still have to fight it. It's not. It's. A lawyer would file a preliminary injunction because they believe they had such a slam dunk case that they could prevent a lot of like, basically going a lot further and spending more money in the case. And so a judge might say, hey, this is actually, it's not clear enough for me to make a decision right now. We're still going to proceed with the case and give both sides an opportunity to continue to make their case. And then there was a little bit of like a twist in the fact that Roy Lee, the founder of Clulee, apparently had worked at Nucleus Genomics. Very great. And Cremieu posts a screenshot of Roy Lee back in February of 2025. So literally just like months before he started Cluli. Very grateful to Kian and the Nucleus genomic teams for taking a chance on me the summer before Columbia and introducing me to the startup world. If I've seen a trillion dollar company in team, it's Nucleus. And Cremieu says he's obviously lying to cover up after getting caught doing fraud. An additional piece of background information people should know is that this fraud also employed the guy behind Cluley, the cheating company. And so Cremieu is being very hardcore in his assessment. Just actually calling Kiyan a fraud straight up is much more aggressive than just saying some of the. Some of their claims are maybe not legitimate. It's unclear. Like, you know, fraud is technically A crime that you need to be convicted of in the court before you are a fraud. But it's certainly, it's certainly he's putting his credibility on the line because if, if, if Keon comes out and says, like, yeah, I'm actually not a fraud, I did it, I proved it. I mean, the main thing looks really bad. Main thing here is, is it appears that the customer reviews. Yes. Are potentially fictitious. And if you're selling a service that allows people to pick their baby and you're giving and you're showing reviews from happy customers that may or may not be real people at all, like that just feels deeply wrong. So I think that one of the first things that they could have done, I don't believe they have, is just say like, no, our reviews are real. We used AI imagery because the people, the real people didn't want their identity online tied to this service. Right. Yeah. For privacy reasons. But I haven't seen anything, I haven't seen anything like this. This guy Adi had a good point. He said, one of the core tensions in this industry is the fact that most companies recognize they're working on an incredibly sensitive topic. They know the general population will need to be slowly and tactfully acclimated to the idea of advanced family planning. Nucleus is perceived as polluting the commons with their deliberately inflammatory marketing. Their virality comes at the cost of increased skepticism for the whole industry. So yeah, a lot of, a lot of folks were not very happy about that. Keon has replied. And if you want to dig into the actual scientific claims on either side, there are long posts where you can go through there. But obviously AI generated blog posts are alleged plagiarism in the Nucleus origin, white paper errors in there, blatant falsification, terms of service are contradictory. Yeah, they also apparently they hired two people that had a non competes for 18 months. Those people just immediately started on working on Nucleus. Nucleus claimed that they weren't competitive so that the non compete didn't apply. But if you look at the companies and what they offer, it seems very clear that they are competing. So anyways, very messy, very messy, messy story. And yeah, I don't know. Will o' Brien says. David, I'm so sorry now man, but you guys are doing an absolutely terrible job at responding to this blog post and seem to be missing the point here. First of all, it is a huge claim to say that every claim by Sichuan Mala is false with absolutely zero evidence for explanation share receipts. Second of all, you make the claim that the person is paid off by competitors of yours, again, with zero evidence. Third, that you make the claim that CREMU is paid off by your competitor. This is bogus and not true. But most importantly, you guys have made zero points of substance here. Rather than just insinuating you guys are leaping ahead and others are jealous. You are selling a scientific product and someone has made a scientific critique in good faith, waiting to be corrected and explicitly saying they will make changes if they are proved wrong. And the best you guys can do is accuse them of being paid off and reply with memes. Not a great look. I want to see startups of the bold vision succeed, but how you communicate with the broader world is so important, especially with a product like yours. And how you guys are carrying on in this, honestly, pretty lackluster. So, yeah, again, I don't. Yeah, at this point, Keon. Keon's been on the show. He's very funny, high energy. We've had some enjoyable conversations. But if I'm a potential customer of Nucleus at this point and I see just these series of exchanges, I'm certainly going to wait and see. See how things evolve versus signing up to use the service to. It's just so different. Cluli. It's so different than Cluli because if I use clulee and I'm like, oh, the notes that were taken in that meeting weren't that good. Or if you go into Cluly being like, I'm going to cheat on this test, and then it's like, oh, it didn't work. Their engineers aren't good enough to really help you cheat on that test. You're like, okay, well, I probably should have been cheating on that test. It's like the lowest stakes thing possible. But when you're, when you're. This is like literally deciding who your child offspring will be. It's the highest. The quality of the product could very well, like contribute echo for generations. Literally. Literally generations. Exactly. Not even just the child's life. Yes, the life of the child's child's child. Yes, it is extremely child's child's. Child's child's. It could alter the course of history. I mean, it kind of could. It's sort of crazy. So, yeah, I mean, it's hard because viral marketing does work. Like, you know, moving fast and breaking things does work in certain contexts, but in the bio, in bloodline optimization, it's really, really high stakes. And so you gotta be extra, extra careful. Extra careful for sure. Well, let me tell you about profound. Get your brand mentioned in ChatGPT, reach millions of consumers who use AI to discover new products and brands. Will Brown has a funny post here. He's just got a recruiting email from a company explicitly mentioning that they have 70.
Bottlenecked on our ability to generate images, we're bottlenecked still on the raw intellectual ability of the models. And that's the sort of direction that we want to push. So in that idea of the bottlenecking, What is the key unlock for Opus 4.5? I mean, a lot of people are throwing around, like, biggest, obviously the benchmarks are very good, but this whole idea of more parameters, more data, more money, more electricity, like, how do you even think about allocating resources to push a model forward in 2025, in late 2025, when perhaps we're past this paradigm of like, oh, just more parameters. I don't know if we are part of. I think it's important to call the paradigm like scaling in general, depending on what axis you're actually scaling. Tbd, but it's general. The scaling, scaling paradigm. I don't think we're past that at all. Right. I mean, I think we're still seeing massive returns to scaling in all its variants. I think that we're generally, things work. We scaled, it works. Want to learn, right? It's as if you said like 10 years ago, the models just want to learn. And I think the hardest things with the question of focus are often on how we split and allocate people. And this model is hundreds of people's worth of effort, right where they poured their lives into it over the last six months. And I think that is, is working out what we prioritize is really tough. I've said this before, but these models always feel like, you know, when they launch, it's exciting. You know, they're great. But you sort of think back to, oh my God, there's all these things that we, we could do better. And everyone right now is going, is going and working on those things. It's just everything still works. So is this a refutation of what some folks might have been picking up on from the last few Dwarkesh guests? The Carpathy episode, the Sutton episode. There's been a little vibe shift around. Like, okay, maybe when we say scaling, we mean more inference diffused all over the world and small models and custom RL environments here and there. And like, we're going to get the value from AI and we are going to continue to scale and dollars to economic value, but it's not just going to be bigger and bigger pre trains forever. And then we get, God, I don't. Know what it's going to be bigger versions of. But as far as we're seeing scaling still works. I don't know if you guys know this, but Dwarkesh, Dylan and I are actually a housemate, so we have this debate all the time, dinner table discussion of, like, you know, are we slowing down? And, you know, I've often joked that the most impactful thing that, you know, what one of us could do is go and crack the problems that, like continual learning or something like this that Dwarkesh focuses on so we can then go switch the narrative back to progress. Yeah, just switch up the dinner table conversation. Conversation, exactly. So did the anthropic crew never lose faith in pre training? There's this whole at Neurips last year.
That actually get adopted beyond the demos, beyond the benchmarks in a couple years or in a couple months. We can't even talk in a couple of years. A couple of years, maybe a couple of weeks, honestly. But like, like, like once it gets in the hands of companies, businesses, startups, you know, different folks implementing this, like how do you see, you know, do you see someone being like, yeah, it's just my daily driver for just talking to it. Even though it costs a lot, how do you think about where you're most excited to see it diffuse into the overall ecosystem? Yeah, I actually do expect this model to become a lot of people's daily driver. It's that step up in being able to delegate trust. We asked internally how much faster Sonnet 4 would have to be for you to take this, to take the switch back basically and give up 0.5 in exchange for. For Sonnet 4.5 and it was multiple times faster. It was like really quite a stark increase in speed. I think it was like four times faster or something. For people to have switched from OPUS to Sonnet, that itself is pretty stark. I think it's also highly likely to become Daily Driver just because it is a lot more efficient. There's this one plot I really like where it shows the amount of tokens it uses to get a certain score on Sweebench and it uses, I think like a quarter of the tokens as Sonnet 4.5 on Suitebench, which is a pretty impressive number. That means it's actually cheaper than Sonnet 4.5 to get the same score on suite bench. Now, TBD, how well that generalizes out to everyday use, but I'm seeing it solve problems way faster. It writes better code the first time round. I actually think in many cases this will end up cheaper because it is so much more efficient at getting to the right answer. Yeah. How are you thinking about personalization and sort of like cross pollination of data?
The rich understanding has helped us a lot in terms of actually understanding how to train these models. I have one extra question. Go for it. Tell me a little bit about Dario's communication style. I was hearing a story about, I think Jensen has no direct reports or everyone reports to him and no one reports to him. He has no meetings or all the meetings. And he has 60 direct reports. He has 60 direct reports, but no big meetings. And he has. But he reads everyone's to do list like every single day or something. What's it like at Anthropic? What is Dario like as a leader these days? Yeah, Dario has a really, really cool communication style, which is that he quite frequently puts out very, very well reasoned essays. And then throughout Slack, we'll have giant essay length, like comment debates with people about the essay. It's really great you get these. But the essays are really nice because one, you can go back and read all the past ones and it tells this history of Anthropic. Yeah, it's, you know, I think in many respects, like it will be one of the better, you know, in a decade from now to chart the history of AGI. Sure. Will be reading these like, compendium of essays. Yeah. And. And there's like incredible comment threads on either side of them and so forth. But also throughout Sleep Slack, whenever where he's very open and honest with the company, whenever we're debating different things, he will lay out the pros and cons and how he's thinking about them and why this one's attention and why that one's moral struggle. And people will write back big essays on why they think we should do X or Y. And he'll respond, it's quite a joy. It's a very written communication style. As a result, it means that many people, or really the entire company have a good model of how he's thinking. And that really helps because it means that you sort of have a coherent sense of direction across the entire company. Yeah, that makes a ton of sense. I like that a lot. Yeah. Cool. So many examples of successful founders who have adopted the written culture and seen great results, I think. And he's a great writer. I mean, read Machines of Love and Grace and it's just such a brilliant essay. That's great. You're absolutely right. Have you ever caught him using AI? Has he ever been like, oh, this one, he's phoning it in? Not yet. Not yet, but maybe soon. I mean, it's kind of a bull case. If he does wind up just saying, could Claude handle it. I'm going on vacation for a couple days. I'm the dropping coworker. I'm pretty sure we measure loss on his essays. That's good. Yeah. Yeah. But right now, I mean, there's a High Bar. High Bar. But congratulations. Thank you so much for taking the time to hop on the show. Yeah, super.
Always gets stuck trying to cooperate with everyone and then just lose all its money. And, you know, sometimes. Sometimes the good guys finish first. I certainly hope that works out. I have genuinely, even though I've never been full, like, oh, my God, I'm going to get Paperclip next year. I have enjoyed a lot of the safety research, and I've always appreciated how thoughtful Anthropic is as an organization around safety. And I think that a lot of people should be a lot more appreciative of how seriously Anthropic takes safety. Not because we didn't get Paperclip this year, but because we saw stuff like GPT psychosis crop up and we saw actual people, know individuals in the venture capital community who. It felt like they got a little crazy. And I'm wondering, do you feel like you're at Anthropic? Do you feel like you're closer to solving the problem of, like, the chatbot went a little bit too sycophantic with me, and it kind of hurt me psychologically because it feels like there's a certain amount of craziness that happens when you're operating at the scale of a billion people. Like, you just pull a billion random people, you're gonna get a lot of crazy people. But at the same time, it feels like this is an interesting place where Anthropic could be doing a lot of research. How are you feeling about solving that problem? And how much can your research kind of generalize to maybe the consumer apps that have more, even more users? But you could maybe be a leader in the space just with the philosophy, because it's like a net good to everyone. Yeah. So we put an enormous amount of effort into this. And, I mean, our models push back a lot. I think there is a tension here between paternalism and freedom, so to speak. Right. But we try and have our models be like, look out for the best interests of the user. I think Mike put it really nicely in a recent talk or podcast where he said, we never look at user minutes as a metric. That is not something that we think about as a proxy of the quality of your experience. We're just out there trying to find out, is it helping you do the things you want and is it adding value? Is it adding value? I hope that our alignment work generalizes really far. I think it's a really tough problem. I mean, I think to OpenAI's credit, they've really gone and tried to fix this problem as well. Right. And it's tough at the scale of a billion users. But I think this is a good example of the kinds of things that are really tricky where there's trade offs and where you need to make sure that you don't have the incentive structure that allows you. That sort of like pushes you to maximize user minutes in this way and is a good microcosm of like the alignment difficulties that we'll get as the models take on more and more and more responsibility now. Yeah, I mean, I completely agree with that. The user minute question like completely snuck up on me because I always assumed that everyone was going to be paying for this stuff as the $20 a month plans rolled out, the $200 a month plans rolled out. But of course, you know, you get to a certain scale of the Internet and it winds up being about attention and advertising and all these different. Yeah, and if you're, if you're building a digital coworker, people don't typically like rate their co workers by how much time they take up. I love this employee. They take up so much of my. Time every week, four hours every day on my calendar. It's the best. Steve just constantly talking to me. Okay, speaking of long running tasks, I want to know how confident should we be in that meter chart of the task doubling? Because.
Models, and that's the sort of direction that we want to push. So in that idea of the bottlenecking, What is the key unlock for Opus 4.5? I mean, a lot of people are throwing around biggest. Obviously, the benchmarks are very good, but this whole idea of more parameters, more data, more compute, more money, more electricity. How do you even think about allocating resources to push a model forward in 2025, in late 2025, when perhaps we're past this paradigm of like, oh, just more parameters. I don't know if we are past the prime. I think it's important to call the paradigm like scaling in general. You know, depending on what access you're actually scaling tbd, but it's general. The scaling paradigm, I don't think we're past that at all. Right. I mean, I think we're still seeing massive returns to scaling in all its variants. I think that we're generally, things work. We scaled, it works. The models just want to learn, right? It's as if you said like 10 years ago, the models just want to learn. And I think the hardest things with the question of the focus are often on how we split and allocate people. And this model is hundreds of people's worth of effort, right? Where they poured their lives into it over the last six months. And I think that is. Is working out what we prioritize is really tough. I've said this before, but these models always feel like, you know, when they launch, it's exciting. You know, they're great. But you sort of think back to, oh, my God, there's all these things that we. We could do better. And everyone right now is going, is going and working on those things. It's just that everything still works. So is this a. Is this a refutation of what some folks might have been picking from the last few Dwarkesh guests? The Carpathy episode, the Sutton episode. There's been a little vibe shift around. Like, okay, maybe when we say scaling, we mean more inference diffused all over the world and small models and custom RL environments here and there. And like, we're going to get the value from AI and we are going to continue to scale dollars to economic value, but it's not just going to be bigger and bigger pre trains forever. And then we get. God, I don't know what it's going to be bigger versions of, but as far as we're seeing, scaling still works. I don't know if you guys know this, but Dwarkesh, Dylan and I are actually a housemate so we have this debate all the time dinner table discussion of like, are we slowing down? And I've often joked that the most impactful thing that one of us could do is go and crack the problems that like continual learning or something like this that Dwarkesh focuses on so we can then go switch the narrative back to progress. Yeah, just switch up the dinner table conversation. Exactly. So did, did, did, did the anthropic crew like never lose faith in, in pre training? There's this whole like at Neurips last year, Ilya says, you know, pre training is potentially dead or kind of alludes to it. And then one of his co presenters is leading the Gemini 3 team and says, oh well, we basically disregarded what we said at Neurips last year. We did just focus on better pre trains, we got better results. It seems like you also disregarded that. Was that a misread in 2024. On. Ilya's presentation or was it a conscious decision to disregard what he was saying? Well, I think, remember in general it's scaling. It's not any particular paradigm of scaling. General flops in intelligence out relationship. Anthropic is a bet on in many respects that we believe that line is going to continue. And exactly what equation you use to convert flops into intelligence out I think will change over time. And many people have made arguments that may even be further paradigms here. But fundamentally we think that the compute in intelligence out equation is continuing to hold. And I think anthropic in many respects has had that faith for a very long time. Some of the first people to have to make very serious bets on that and a couple of months of external progress being, I think the only reason that people are so like, how should I say? The models have actually gotten substantially smarter this year. And that's why.
Measure loss on. On his essays. That's good. Yeah, yeah. But right now, I mean there's a high bar. High bar. But congratulations. Thank you so much for taking the time to hop on the show. Yeah, super impressive. Congrats to the whole team. We'll talk to you soon. Great to see you. See you. Ciao. Bye. Back to the show, back to the timeline. Back to linear. Meet the system for modern software development Purpose built tool for planning and building products. There is more OpenAI news, of course, more tech news of all times. OpenAI's hardware division says Mark Gurman built around Jony I've secretive startup has ramped up the hiring of Apple engineers. The group has brought on about 40 new people in the last month or so with many of them coming from Apple's hardware group. Yeah, hearing that Sholto interview, I'm disappointed. I don't think we're getting ads from anytime soon and I don't think we're going to get a mobile device. Well, we are actually talking today to Quinn Slack, the CEO of AMP and sourcegraph. AMP is a frontier coding agent and AMP is free. They introduced AMP Free which is ad supported and has a no cost mode and so you can now use their coding agent for. For free with ads. 40 people. That does not seem like cause for concern for Apple. I mean I can't imagine how big their hardware group is but it has to be, you know, in the thousands, I would imagine. Yeah, let's try to find out. Huge organization. So OpenAI is poaching left and right from Apple's hardware engineering group, hiring around 40 directors, managers and engineers in the last month from nearly every relevant Apple department. Mark Gurman says it's remarkable. So from what I've heard, this is Mark Gurman. Apple is none too pleased about OpenAI's poaching and some consider it a problem. The hires include key directors, a fairly senior designation, as well as managers and engineers. And they hail from a wide range of areas. Camera engineering, iPhone hardware, Mac hardware, silicon device testing and reliability, industrial design, manufacturing, audio, smartwatches, Vision Pro development software. They got one from every single sampled. Every single division. I suppose. Gemini is estimating that Apple has has between 15,000 and 20,000 hardware engineers in total. 15,000? That seems like a lot. I don't know. In other words, OpenAI is picking up people from nearly every relevant department. It's remarkable, says Mark Gurman. Very interesting. I wonder how the comp structured, how everything will come together on those teams. I mean there's a lot of people from Apple who going over to OpenAI, it's a greenfield project. It's probably really fun, probably really exciting, probably not the most mercenary scenario, but there's always that risk. If you're, if you're working at Apple and you're excited about AI and you've been there for the last three years, watching all this progress happen at the application layer, the model layer, and not being thrilled with the progress happening at the hardware layer. This is like a. Yeah, yeah, just a, it's a, it's a wide open opportunity to like be working right at that intersection of the models and the hardware. There's a lot of AI engineers who have made moves because they don't want to be a GPU poor company. And it's weird because Apple's in this scenario where they're partnering with Gemini now. They're clearly going to survive. It's not a serious threat, at least not yet. Maybe if this device is incredible, but right now Apple looks pretty strong. The new iPhones are selling well, everything's good. But what's like from an AI perspective, it's got to be one of the worst gigs because you were in this sort of like openly hostile environment to LLMs, to scaling, to building large GPU clusters and then, yeah, they're sort of playing catch up now, but they're certainly not calling up Oracle for, you know, trillion dollars of compute. You go over to OpenAI, you're just going to be immersed in a lot more higher risk taking higher risk on I wonder. Yeah, Gabe is asking if wouldn't that be bearish as the hardware group at App Apple is responsible for the terrible Vision Pro. I do wonder. I do wonder. It depends on who it was. An incredible technological feat. I just think they built the wrong. I was just watching a thing about A guy who 3D printed an adapter so you could use the strap from the Apple Vision Pro on the quest 3 from Meta. And that really speaks to the fact that the Apple Vision Pro, although it was too heavy, it has this screen on the outside that I don't think anyone wants. There were pieces about it that were clearly the best. The screen is just the best. No one's debating that. The band, the Knit band, is very cool. It has this amazing device where you rotate it and it tightens up. There are a whole bunch of things that are amazing. It's just like as a package, it didn't deliver. But if you just want to, if your job is just like, hey, we gotta put a screen on this and it's gotta be the highest resolution screen like, well go to the place that developed the highest resolution screens like they did a good job. Well Sam Altman replied to one of our cards we put up on November 22 TVPN posted on this day, Sama was rehired at OpenAI, got his badge back and Sam Altman replied and said cannot believe this was only two years ago. Subjectively feels like five. Yeah. What a turnaround to go from defenestrated to back in the seat in so much and have so much control over the organization that you're able to raise massive valuations, broker all these deals, move the entire market. Just a remarkable run, put on such a master class in deal making that people are now sitting here being like there's no way that this would be a $500 billion company if Sam wasn't in the driver's seat. First. Let me tell you about fall Build and deploy AI video and image models trusted by millions to power generative media at scale Danny Zhou, who founded the reasoning team in Google Brain, now part of the Gemini team at Google DeepMind. He says game over and carried no interest. Friend of the show quotes it and says I genuinely think OpenAI equals equals Yahoo. He's not assigning the variable, he's equating it. I've migrated almost all my workflow's code off their APIs now. Ironic that Google will probably do it twice. LMAO and I don't know about this. The pattern matching on the Yahoo example. We should have him on the show and actually do all the other yeah. You were saying the other thing with the Yahoo Yahoo example is it wasn't like there wasn't like the company was valued pretty tremendously for a longer period of time. It wasn't like this binary like one. Moment exactly like the peak market cap for Yahoo 125 billion during 2000. That feels like it's just hard to pattern match perfectly to this. But I mean it certainly would be poetic if that's the way it played out. It is funny, Carried says, ironic that Google will probably do it twice. They actually created the transformer. They released the transformer paper and chromium to inspire themselves to find harder sort of challenge giving everything a straight away give out all the alpha. Yeah yeah for sure. Just kind of like find their fire again. Yeah there was a. There was an interesting article on Medium that was sort of burning up hacker news that I thought would be fun to go through first, let me tell you about graphite.dev code review for the age of AI Graphite helps teams on GitHub ship higher quality software faster. So this person who no one has, no one can really understand who this person is. They don't necessarily exist on the Internet fully. So there was like a question about that. This is like an very like, you know, hacker news and turmoil segment. But Taeha says, I reverse engineered 200 AI startups. 146 are selling you repackaged ChatGPT and Claude with new UI. And so basically the thesis of this article is that this fellow wrote a piece of code that looks at the marketing copy and says, what are they claiming? And then looks at the calls that happen when you actually interact with their AI features. So if there's a chatbot on this particular startup's website and you are near chatting with it and you look into the trace that's happening in Chrome, is it going to the startup server or is it going to OpenAI server, or is it going to Anthropic server? That's telling. And then there's also a little bit of API fingerprinting. Basically, OpenAI has a specific pattern of rate limiting and it's exponential. So if you're spamming the OpenAI API, it will, according to a unique pattern, tell you, hey, you've sent too many messages, cool off for one minute, and then the next time you do it, cool off for two minutes. The next time, cool off for four minutes, then eight minutes, then 16. Right. And it gets exponentially longer and you're on progressively more longer and longer timeouts. But the shape of that curve and the specific timings are unique to OpenAI. And so if I'm a startup, interesting. And I have the exact same like back off and timeout curve, well then it's probably just OpenAI under the hood. At least that's the, that's the claim that's being made here. And so the finding in this article is that 73% had a significant gap between the claimed technology and the actual implementation. And so out of the 200 AI startups that this fellow analyzed, 54 companies either had accurate technical claims. They said, hey, we're using like we have a custom AI model that we trained. And they did. Or they're transparent about their stack. They say, hey, this is a wrapper. Like we're a wrapper company. And so, you know, our AI is powered by ChatGPT, we're partnered with OpenAI, we're partnered with Anthropic or whatever. Now 146 companies, that's 73%, according to him. Were sort of misrepresenting their technology. So either they said they had proprietary AI. Proprietary AI. And yet when he dug into it, it was OpenAI API+ Prompts. Tyler. Yeah, I mean it's like kind of what, what do people expect? Like. Yep. Like is if you fine tune. If you use the OpenAI API to fine tune the model. Yes. Which you can do. Yes. Is that proprietary? Proprietary like no one else has that fine tune? Yes. You're still calling the API. It's like I don't expect startups to train their own full language models. That's like pretty unrealistic and like really make sense. Yeah. So I'm kind of confused. I guess it's a very cool study. But this tracks with exactly like I would guess that 73% of AI startups are just re skinning. Yes. And so 19% of the overall companies, the 38 that were analyzed in this study found that when that the startup said they had in house models and it was actually fine tuned public models. And so it's a question is like whose house is it in? It's technically in OpenAI's house. So fine tune as a punch in an open source model that's public, does that count as a public model? An open source model? Let's assume yes. And then. And then last 8% would. So they had a custom ML pipeline and they were in fact using standard cloud services. That's even wishy washier. I think that's totally fine because like you can totally have a custom ML pipeline that's wiring together OpenAI and Gemini and AWS and you know, a bunch of other. If I'm using a startup, I don't want them to train their own language model because I don't think they're going to like in 99.99% of the case, like they're not going to be able to do a better model than OpenAI anthropic Gemini Grok. Like I want them to use the best model. Yes. And it's like, okay, you can fine tune it or whatever. I agree. And so the author also agrees with you. He says here's what really shocked me. I'm not even mad about it. Every time I saw the phrase our proprietary language model, I knew what I was going to find. And I was right. 34 out of 37 times. And this is where it gets weird because he says here's the technical signature. And so the user submits the query, it posts to API slash generate and then with wrapper logic it posts to API.OpenAI.comv1/chat completions. And I have no idea how he's seeing the backend. It makes no sense how he would be able to do this unless there was just like a massive security vulnerability. Because what I would assume is happening is that the user's over here, the startup's websites here, and then the user goes to the startup's website and then the startup's website on the back end talks to OpenAI and comes back. And maybe you could understand that like, okay, the amount of EM dashes, like there's a. This is telltale signal. But that's not what he's doing. He's saying that he was able to just literally hit like the Chrome Inspect developer tools. Look at the chain of calls and see that it was calling OpenAI from the front end, which is crazy because I didn't even know you could do that. It feels like if you were calling it directly from the front end, you would potentially leak a key that would be able to put you on the hook for a bunch of bills. I would think you would want to authenticate that on the back end. And so he gives a bunch of examples of rag and then he's exposing some margins, which is actually very bulky bullish for these companies because he breaks down some of these and says that a GPT4 API is $0.03 per thousand input token, $0.06 per thousand output token. So the cost per query for this hypothetical startup was $0.03 and they charged $3 or $300 per month for 200 queries. And so the wrapper economy, this is 75 times direct costs. That's extremely bullish for printing for that company. He found another one that was maybe 1000x API costs that's doing some pinecone embedding. And he also says this is pattern number three. The quote, we fine tuned our own model. Reality check. Fine tuning sounds impressive and it can be, but here's what I found. 45% it was OpenAI's fine tuning API, which. That sounds right, right? It's a little bit of a step to be like, we fine tuned our own model. It's like, no, you fine tuned OpenAI's model and then you got your own model from that result. It's a little bit. Yeah, you're still fine tuning it. I don't think there's that big a difference between fine tuning. Are you fine tuning it or is OpenAI fine tuning it? There's a fine tuning API which you use to and then a new. But who's Doing the fine tuning you or OpenAI? Well, what do you like? You're not interfacing directly with the gpu. It's like I went to the store and I bought a sandwich. Who made the sandwich? Well, I told them what the used turkey, I told them put lettuce, extra pickle. Who made the turkey? I think it's more like you go to the store and you bring a sandwich to the office. It's like where did the sandwich come from? It came from the store. But like you brought the sandwich, it came from you, you brought the sandwich. Yeah, this is better. I think you're right. Anyway, 22% into the time it was a hugging face model with Alora. 18% of the time was anthropic Claude with a prompt library. 8% of the time is literally just GPT4 with system prompts. And 7% they actually changed change, trained something from scratch. And all of these are odd. And there's a lot of deb. How. This would actually happen because he's basically saying just open dev tools, go to the network tab, interact with the AI feature. If you see API OpenAI API Anthropic or API cohere AI, you're looking at a wrapper. They might have middleware, but the AI isn't theirs. And so it just opens up this debate about what is the value of the wrapper. I mean certainly if you can resell something for 100x because you have some sort of clever workflow prompt or workflow, more power to you. Yeah, it's not exactly, it's not exactly like bearish on the companies. But the debate that was surrounding was more around so, so this, this author claims that after posting this, seven founders reached out privately. Some were defensive, some were grateful. They asked for help transitioning their marketing from proprietary AI to built with the best in class APIs. Because some of these founders did, did I guess feel like using proprietary AI as a marketing tagline was disingenuous. And then someone else, I think I saw something that was one VC reached out and said like I'd like you to audit my portfolio because I have been told that I was investing in companies that were training their own AI and I made the investment on that assumption. And if I'm being lied to then that's potentially securities fraud. And so there is a question about if you go, I mean I've seen pitches for companies that, where they've said like proudly like you should invest in this because we're not training our own model. It would actually be a mistake. And there's another company that's a competitor to us that is training their own model and you don't want to invest in them, you want to invest in us. Because we're going to burn your dollars. Yeah, we're going to much better economically. And so all of it just the only thing that matters is like being upfront with the investor for sure. And then to some degree you do need to be upfront with the customer because if the customer there is a marketing value to. Oh, if you work with us, you're working with these genius AI scientists who are going to build their own models. And if it's just repackaged ChatGPT, that might not be what you want to pay for because at that point you might just say, hey, like actually if I can just get this directly from OpenAI, I'll just go buy it from them anyway. Well, thoughts and prayers to friend of the show, John Palmer. He says he just found out my wife is leaving me. She said I'm not legible to capital. The legible to capital meme is fantastic. I do think that will he made a new meme. Fantastic coinage. I love it. Absolutely Ripper. We'll be using it. Let me tell you about FIN AI, the number one AI agent for customer service. It's AI that handles your customer support. Timeline is in turmoil over Nucleus. Former guest on the show two or three times. Keon has founded Nucleus for IVF and he put up a subway campaign that says IQ is 50% genetic, height is 80% genetic. I completely disagree with that one. It's entirely skill based for me. The genes did not matter. I had to grind for this view. Grind my growth plates, I suppose. Have your best baby is what it says. And it says IVF done right in the subway all over New York City. There's a ton of debate going on. And to be clear, accurate, I think it was intentionally trying to make some percentage of the population angry to drive enough energy and attention. So this was. Yeah, I would call it, I would call it rage bait. So I would call it rage bait marketing. Not necessarily, not at the product level. But IVF as a category is a controversial category and so it's much easier to wrap it in a campaign that will go viral for upsetting reasons. For you can upset people and you can get a lot of attention from that. This is an example from Kath Korvac. She says, so eugenics is profitable now and so being able to wrap something that is just a, you know, a scientific process that's been Worked on for a long time. Seems to be somewhat. Friend.com inspired Keon's original post. He says, nucleus Genomics announces the largest genetic optimization campaign ever. Which is. Which is just funny because a friend was saying this is the largest out of home campaign ever. And now Keon is saying this is the largest genetic optimization campaign ever. So narrowing it down, but full station Blitz at Broadway, 1000 plus street ads across New York City, 1000 plus subway car ads, dozens of urban panels throughout SoHo. And apparently they're not actually. They're not able to offer the service in New York in here. So it's really just an image of a controversial phrase on a New York subway is more likely to go viral. So you do it there because it looks like you're on the global stage, and then you pull it. There's a high density of people that have a large following audience. Yeah, following. And so it's just the way to start a viral trend and own the moment. It's the reason why, you know, so many tiktokers are in Manhattan now doing stuff like man on the street stuff. It just like it has more like, aura almost. Well, Dr. Shelby liked the mindshare grabbing that Nucleus did. Says every biotech founder should be seeing this and understanding how to get one tenth the mindshare of Nucleus. I have a playbook for you below. A lot of people are like, I love the playbook. I don't love this example because the company's getting dragged. I don't know if it's good or bad with the rage bait thing. I think usually it's a negative thing. Usually it's hard to come back from. Occasionally it can be done in a way that's slightly enraging, but enough people are in on the joke that they appreciate what's happening and they appreciate that it breaks through or it's enraging to someone who's not the core audience, not the actual customer. And so it's okay. But it's a big debate because Sichuan Mala posted a long essay all about the claims made by Nucleus. Keon says everything levied unto Nucleus by Sichuan Mala is false. Worse than false. It appears to be architected by a competitor that has repeatedly published misstatements and inaccuracies. Sichuan is compromised, but it gets. Yeah. To be clear, no evidence has been provided that it was being levied by a competitor. Yes, that's purely an allegation that has no. There's no proof. Yes, yes, exactly. Yeah. Sometimes there's DMs, that leak, and there's evidence. Or someone comes forward and says, yeah, I was actually paid to post that. But so he says, I've been informed that Cremio. I don't know how to pronounce that last thing. Cremio, who's been on the show also he claims he's a race scientist in chief has been paid off by the competitor to promote this nonsense against Nucleus for the independent scientists repeating the Cremio. Denied. Camilla denied that as well. I would encourage you to do more diligence on who you're aligning yourself with. Our scientific team will issue a point by point response, which I believe they did. Unfortunately though, this isn't about science. It's about. It's a concentrated attempt to cancel Nucleus on the backs of our successful campaign and build. And efforts to build and advance the industry which benefits the very people attacking us. The mob are trying to cancel Nucleus. Keep tweeting, stay mad, we'll keep building and serving patients. Yes, we won the injunction. Link below. So they were sued by their competitor. But, but so they won the injunction. They didn't mean they won the case at all. I mean, that's a classic thing if you're getting sued to be like, the case was dismissed and it's like one of five cases against you. Yeah, but in this case, they won. They won a preliminary injunction, which means that the case is just still progressing and they still have to fight it. Yeah, it's not. It's. A lawyer would file a preliminary injunction because they believe they had such a slam dunk case that they could prevent a lot of like, basically going a lot further and spending more money in the case. And so a judge might say, hey, this is actually, it's not clear enough for me to make a decision right now. We're still going to proceed with the case and give both sides an opportunity to continue to make their case. And then there was a little bit of like a twist in the fact that Roy Lee, the founder of Clulee, apparently had worked at Nucleus Genomics. Very great. And Cremu posts a screenshot of Roy Lee back in February of 2025. So literally, just like months before he started Cluli. Very grateful to Kian and the Nucleus genomic teams for taking a chance on me the summer before Columbia and introducing me to the startup. Ever seen a trillion dollar company and team? It's Nucleus. And Cremieux says he's obviously lying to cover up after getting caught doing fraud. An additional piece of background information people should know is that this fraud also employed the guy behind Cluley, the cheating company. And so Cremieu is being very hardcore in his assessment. Just actually calling Kian a fraud straight up is much more aggressive than just saying some of the, some of their claims are maybe not legitimate. It's unclear. Like, you know, fraud is technically a crime that you need to be convicted of in the court before you are a fraud. But it's certainly, it's certainly he's putting his credibility on the line because if, if, if Keon comes out and says, like, yeah, I'm actually not a fraud, I did it, I proved it. I mean the main thing, the main thing here is, is it appears that the customer reviews. Yes. Are potentially fictitious. And if you're selling a service that allows people to pick their baby and you're giving and you're showing reviews from happy customers that may or may not be real people at all, like, that just feels deeply wrong. So I think that one of the first things that they could have done, I don't believe they have is just say like, no, our reviews are real. We used AI imagery because the, the people, the real people didn't, didn't want their identity online tied to this service. Right. Yeah. For privacy reasons. Yep. But I haven't seen anything, I haven't seen anything like this. This guy Adi had had a good point. He said one of the core tensions in this industry is the fact that most companies recognize they're working on an incredibly sensitive topic and they know the general population will need to be slowly and tactfully acclimated to the idea of advanced family planning. Nucleus is perceived as polluting the commons with their deliberately inflammatory marketing. Their virality comes at the cost of increased skepticism for the whole industry. So yeah, a lot of folks were not very happy about that. Keon has replied. And if you want to dig into the actual scientific claims on either side, there are long posts where you can go through there. But obviously AI generated blog posts are alleged plagiarism in the Nucleus origin, white paper errors in there, blatant falsification, terms of service are contradictory. Yeah. They also apparently they hired two people that had a non competes for 18 months. Those people just immediately started on working on Nucleus. Nucleus claimed that they weren't competitive, so the non compete didn't apply. But if you look at the companies and what they offer, it seems very clear that they are competing. So anyways, very messy, very messy, messy story. And yeah, I don't know. Will o' Brien says, David, I'm so sorry now man, but you guys are doing an absolutely terrible job at responding to this blog post and seem to be missing the point here. First of all, it is a huge claim to say that every claim by Sichuan Mala is false with absolutely zero evidence or explanation share receipts. Second of all, you make the claim that the person is paid off by competitors of yours, again with zero evidence. Third, that you make the claim that Cremu is paid off by your competitor. This is bogus and not true. But most importantly, you guys have made zero points of substance here rather than just insinuating you guys are leaping ahead and others are jealous. You are selling a scientific product and someone has made a scientific critique in good faith, waiting to be corrected and explicitly saying they will make changes if they are proved wrong. And the best you guys can do is accuse them of being paid off and reply with memes. Not a great look. I want to see startups with a bold vision succeed, but how you communicate with the broader world is so important, especially with a product like yours and how you guys are carrying on in this. Honestly, pretty lackluster. So, yeah, again, I don't. Yeah, at this point, Keon's been on the show, he's very funny, high energy, we've had some enjoyable conversations. But if I'm a potential customer of Nucleus at this point and I see just these series of exchanges, I'm certainly going to wait and see. See how things evolve versus signing up to use this service to. It's just so different than Cluli. It's so different than Cluli because if I use Cluli and I'm like, oh, the notes that were taken in that meeting weren't that good. Or if you go into Cluly being like, I'm going to cheat on this test, and then it's like, oh, it didn't work. Their engineers aren't good enough to really help you cheat on that test. You're like, okay, well, I probably should have been cheating on that test. It's like the lowest stakes thing possible. But when you're, when you're. This is like literally deciding your child. Offspring will be. It's the highest. The quality of the product could very. Well like infuse echo for generations. Literally. Literally generations. That's exactly. Not even just the child's life. Yes, the life of the child's child. Child's child. Yes, it is extremely. The child's child's child's child's. It could alter the course of history. I mean, it kind of could. It's sort of crazy. So, yeah, I mean, it's hard because viral marketing does work like, you know, moving fast and breaking things does work in certain contexts, but in the bio, in bloodline optimization, it's really, really high stakes. And so you got to be extra, extra careful. Extra careful for sure. Well, let me tell you about profound. Get your brand mentioned in ChatGPT. Reach millions of consumers who use AI to discover new products.
App launched almost two years ago and there's still like rough edges in the ui which I think is crazy. But it does seem like they have an opportunity to actually take some serious market share at this point. Like they've caught up on many different, many different values and like value props. My question was like I'm not the typical consumer. Like I'm going to try every different app. Like I'll probably keep bouncing around. I don't know if consumers will do the same. Broadly there's, it's very, very clear that ChatGPT is just synonymous with AI and people are not like oh well like the new benchmarks, I gotta like change my, you know, app. Like no one's thinking like that. But my. The fragility in the ChatGPT monopoly aggregator thesis that I was picking up on was for the last year there have been a lot of, a lot of features and theses around different things that could create lock in. So stuff like personalization or memory or even the chat functionality between what you've linked your custom instructions. The different. I think at this point I've synced ChatGPT or Auth ChatGPT with a number of different services so it should have more data, it should know all these different things. I've given it even, even custom instructions just saying like hey, cool it on the EM dashes. And I didn't miss any of that. There was at no point where, there were plenty of points where I was like oh, ChatGPT is definitely better than Gemini still. But at no point was I like it's because it doesn't have personalization. And I think that if I went in Gemini and I was like oh yeah, you can go take a peek at my Gmail to get personalized to understand how I write or understand what I'm interested in. Like one, I could snap my finger. And Google could be way more personal. Maybe or, but, but the biggest thing is that like right now I just don't know. I feel like both are not personalized at all. None of them have any real lock in of any sort. And even in like the chat functionality or like the social network functionality, which is just very different than what happens in a true social network or where there's this flywheel and the, and the content is driven by the existing user base. Whereas I feel like I got on Gemini and on day one the content was as good or better than ChatGPT because it's all AI generated. So it made me think like maybe it's a little bit more fragile, maybe there will be a little bit more of a duopoly. It won't be such a winner take all market, even though it has been historically, it has been up to this date. Like in consumer AI, it's very clear that OpenAI has run away with it, but it feels like Google does have a little bit of a chance to catch up in consumer because there's just so much less of a network effect. Like the network effect just is bolted on. It's not real yet. Maybe it'll never be real. Maybe Google can catch up there, but I just really want to see where DAUs and user minutes actually grow because there's so many different tweaks there and every chart and data point is definitely going to be analyzed to death. Yeah, one thing that's notable, Google's going super hard in this for students, so you can just get Gemini.
Always gets stuck trying to cooperate with everyone and then just lose all its money. And, you know, sometimes. Sometimes the good guys finish first. I certainly hope that works out. I have genuinely, even though I've never been full, like, oh, my God, I'm going to get Paperclip next year. I have enjoyed a lot of the safety research, and I've always appreciated how thoughtful Anthropic is as an organization around safety. And I think that a lot of people should be a lot more appreciative of how seriously Anthropic takes safety. Not because we didn't get Paperclip this year, but because we saw stuff like GPT psychosis crop up and we saw actual people, know individuals in the venture capital community who. It felt like they got a little crazy. And I'm wondering, do you feel like you're at Anthropic? Do you feel like you're closer to solving the problem of, like, the chatbot went a little bit too sycophantic with me, and it kind of hurt me psychologically because it feels like there's a certain amount of craziness that happens when you're operating at the scale of a billion people. Like, you just pull a billion random people, you're gonna get a lot of crazy people. But at the same time, it feels like this is an interesting place where Anthropic could be doing a lot of research. How are you feeling about solving that problem? And how much can your research kind of generalize to maybe the consumer apps that have more, even more users? But you could maybe be a leader in the space just with the philosophy, because it's like a net good to everyone. Yeah. So we put an enormous amount of effort into this. And, I mean, our models push back a lot. I think there is a tension here between paternalism and freedom, so to speak. Right. But we try and have our models be like, look out for the best interests of the user. I think Mike put it really nicely in a recent talk or podcast where he said, we never look at user minutes as a metric. That is not something that we think about as a proxy of the quality of your experience. We're just out there trying to find out, is it helping you do the things you want and is it adding value? Is it adding value? I hope that our alignment work generalizes really far. I think it's a really tough problem. I mean, I think to OpenAI's credit, they've really gone and tried to fix this problem as well. Right. And it's tough at the scale of a billion users. But I think this is a good example of the kinds of things that are really tricky where there's trade offs and where you need to make sure that you don't have the incentive structure that allows you. That sort of like pushes you to maximize user minutes in this way and is a good microcosm of like the alignment difficulties that we'll get as the models take on more and more and more responsibility now. Yeah, I mean, I completely agree with that. The user minute question, like, completely snuck up on me because I always assumed that everyone was going to be paying for this stuff as the $20 a month plans rolled out, the $200 a month plans rolled out. But of course, you know, you get to a certain scale of the Internet and it winds up being about attention and advertising and all these different. Yeah, and if you're building a digital coworker, people don't typically like rate their co workers by how much time they take up. I love this employee. They take up so much of my. Time every week, four hours every day on my calendar. It's the best. Steve just constantly talking to me. Okay, speaking of long running tasks, I want to know how.
It is so much more efficient at getting to the right answer. Yeah. How are you thinking about personalization and sort of like cross pollination of data? When I think about an engineer on a team, it's helpful to have them in Slack. It's even helpful to potentially have them in the random channel and just kind of, you know, understanding the company culture. Yeah. And I was toying with like trying to switch from ChatGPT to Gemini as kind of the daily driver knowledge retrieval app. And I was noticing that like the personalization narrative hasn't really taken hold over the time I was testing Gemini. I was not like, oh, this feels like wildly less personal. And that might just be a matter of like it hasn't had that much time to build in all those personalization features. But I'm wondering if you see a world where developers who are using Claude for programming also benefit from using it in knowledge retrieval research and there's actual significant flowback and synergy such that it's a really valuable. It's valuable to actually have both sides of the business like really cooking. Yeah. It's a team member. I mean, we've talked a long time about how we want Claude to be a virtual coworker. Right. Big goal really for next year is to try and get to this form factor of virtual coworker that is in all your Slack channels and can join your meetings and can work alongside you. I think there's going to be massive benefit there. I think that. And my basic expectation and as I sort of interacted with, with the model is that it will get to that point where it's useful to have it across everything Now. I think that there's like one is, it's, as you said, it's worth asking the question of why haven't we seen personalization really kick off so far? Like why isn't it, why isn't that useful? I think that's in part because there's still a lot of algorithmic progress to go there. I think.
Giving everything a freeway. Give out all the alpha. Yep, yep, yep, for sure. Just kind of like find their fire again. Yeah, there was a, there was an interesting article on Medium that was sort of burning up hacker news that I thought would be fun to go through. First, let me tell you about graphite.dev, code review for the age of AI. Graphite helps teams on GitHub ship higher quality software faster. So this person, who no one can really understand who this person is, they don't necessarily exist on the Internet fully. So there was like a question about that. This is like a very like, you know, hacker news and turmoil segment. But Taeha says, I reverse engineered 200 AI startups, 146 are selling you repackaged ChatGPT and Claude with new UI. And so basically the thesis of this article is that this fellow wrote a piece of code that looks at the marketing copy and says, what are they claiming? And then looks at the calls that happen when you actually interact with their AI feature. So if there's a chatbot on this particular startup's website and you are near chatting with it and you look into the trace that's happening in Chrome, is it going to the startup server or is it going to OpenAI server, or is it going to Anthropic server? That's telling. And then there's also a little bit of API fingerprinting. Basically, OpenAI has a specific pattern of rate limiting and it's exponential. So if you're spamming the OpenAI API, it will, according to a unique pattern, tell you, hey, you've sent too many messages, cool off for one minute, and then the next time you do it, cool off for 2 minutes. The next time, cool off for 4 minutes, then 8 minutes, then 16. Right? And it gets exponentially longer and you're on progressively more longer and longer timeouts. But the shape of that curve and the specific timings are unique to OpenAI. And so if I'm a startup and I have the exact same like back off and timeout curve, well, then it's probably just OpenAI under the hood. At least that's the claim that's being made here. And so the finding in this article is that 73% had a significant gap between the claimed technology and the actual implementation. And so out of the 200 AI startups that this fellow analyzed, 54 companies either had accurate technical claims. They said, hey, we're using, we have a custom AI model that we trained. And they did, or they're transparent about their stack. They say, hey, this is a wrapper. We're a wrapper company. And so our AI is powered by ChatGPT, we're partnered with OpenAI, we're partnered with Anthropic or whatever. Now 146 companies, that's 73% according to him, were sort of misrepresenting their technology. So either they said they had proprietary AP AI, proprietary AI, and yet when he dug into it, it was OpenAI API plus prompts. Tyler. Yeah, I mean it's like kind of what, what do people expect? Like, like is if you fine tune. If you use the OpenAI API to fine tune the model. Yes. Which you can do. Yes. Is that prepared proprietary like no one else has that fine tune? Yes. You're still calling the API. It's like I don't expect startups to train their own full language models. That's like pretty unrealistic and like it really makes sense. Yeah. So I'm kind of confused, I guess. Yeah, I guess, I guess it's a very cool study. But this tracks with exactly like I would guess that 73% of AI startups are just reskinning. Yes. And so 19% of the overall companies, the 38 that were analyzed in this study, found that the startup said they had in house models and it was actually fine tuned public models. And so it's a question, it's like, whose house is it in? It's technically in OpenAI's house. So fine tune, as you mentioned, an open source model that's public, does that count as a public model? An open source model? Let's assume. Yes. And then last 8% would. So they had a custom ML pipeline and they were in fact using standard cloud services. That's even wishy washier. I think that's totally fine because like you can totally have a custom ML pipeline that's wiring together OpenAI and Gemini and AWS and you know, a bunch of other. If I'm using a startup, I don't want them to train their own language model because I don't think they're going to like in 99.99% of the case, like they're not going to be able to do a better model than OpenAI anthropic Gemini Grok. Like I want them to use the best model. Yes. And it's like, okay, you can fine. Tune it or I agree. And so the author also agrees with you. He says, here's what really shocked me. I'm not even mad about it. Every time I saw the phrase our proprietary language model, I knew what I was going to find and I was right 34 out of 37 times. And this is where it gets weird because he says, here's the technical signature. And so the user submits the query, it posts to API generate and then with wrapper logic, it posts to API.OpenAI.comv1chat completions. And I have no idea how he's seeing the backend. It makes no sense how he would be able to do this unless there was just like a massive security vulnerability. Because what I would assume is happening is that the user's over here, the startup's website's here, and then the user goes to the startup's website, and then the startup's website on the back end talks to OpenAI and comes back. And maybe you could understand that, like, okay, the amount of EM dashes, like there's a. This is telltale signal. But that's not what he's doing. He's saying that he was able to just literally hit like the Chrome Inspect developer tools, look at the chain of calls and see that it was calling OpenAI from the front end, which is crazy because I didn't even know you could do that. It feels like if you were calling it directly from the front end, you would potentially leak a key that would be able to put you on the hook for a bunch of bills. I would think you would want to authenticate that on the back end. And so he gives a bunch of examples of rag and then he's exposing some margins, which is actually very bold for these companies because he breaks down some of these and says that, you know, a GPT4 API is $0.03 per thousand input token, $0.06 per thousand output token. So the cost per query for this hypothetical startup was $0.03, and they charged $3 or $300 per month for 200 queries. And so the wrapper economy, this is 75 times direct costs. That's extremely bullish for that printing for that company. He found another one that was maybe 1000x API that's doing some pinecone embedding. And he also says, this is pattern number three. The quote, we fine tuned our own model. Reality check, fine tuning sounds impressive, and it can be, but here's what I found. 45%. It was OpenAI's fine tuning API, which. That sounds right, right? It's a little bit of a step to be like, we fine tuned our own model. It's like, no, you fine tuned OpenAI's model and then you got your own model from that result. It's A little bit, yeah. You're still fine tuning it. I don't think there's that big a difference between fine tuning. Are you fine tuning it or is OpenAI fine tuning it? There's a fine tuning API which you use to and then a new. But who's doing the fine tuning, you or OpenAI? Well, what do you like? You're not interfacing directly with the gpu. It's like I went to the store and I bought a sandwich. Who made the sandwich? Well, I told them what the used turkey, I told them. You told them put lettuce, extra pickle. Who made that turkey? I think it's more like you go to the store and you bring a sandwich to the office. Like where did the sandwich come from? It came from the store, but like you brought the sandwich, it came from you. Yeah, this is better. I think you're right. Anyway, 22% of the time it was a hugging face model with Alora. 18% of the time was anthropic Claude with a prompt library. 8% of the time is literally just GPT4 with system prompts. And 7% they actually trained something from scratch. And all of these are are odd. And there's a lot of debate over how this would actually happen because he's basically saying just open dev tools, go to the network tab, interact with the AI feature. If you see API OpenAI API anthropic or API cohere AI, you're looking at a wrapper. They might have middleware, but the AI isn't theirs. And so it just opens up this debate about what is the value of the wrapper. I mean, certainly if you can resell something for 100x because you have some sort of clever workflow prompt or workflow, more power to you. Yeah, it's not exactly like bearish on. The companies, but the debate that was surrounding was more around. So this author claims that after posting this, seven founders reached out privately. Some were defensive, some were grateful. They asked for help transitioning their marketing from proprietary AI to built with the best in class APIs. Because some of these founders did. Did I guess feel like using proprietary AI as a marketing tagline was disingenuous. And then someone else, I think I saw something that was one VC reached out and said like, I'd like you to audit my portfolio because I have been told that I was investing in companies that were training their own AI and I made the investment on that assumption. And if I'm being lied to, then that's potentially securities fraud. And so there is a question about if you go. I mean, I've seen pitches for companies that, where they've said like proudly like, you should invest in this because we're not training our own model, it would actually be a mistake. And there's another company that's a competitor to us that is training their own model and you don't want to invest in them, you want to invest in us. Because we're going to burn your dollars. Yeah. We're going to much better economics. Yeah. And so all of it just, the only thing that matters is like being upfront with the investor for sure. And then to some degree you do need to be upfront with the, with the, with the customer. Because if the customer there is a marketing value to. Oh, if you work with us, you're working with these, you know, genius AI scientists who are going to build their own models. And if it's just repackaged ChatGPT, that might not be what you want to pay for because you, at that point you might just say, hey, like actually if I, if I can just get this directly from OpenAI, I'll just go buy it from them anyway. Well, thoughts and prayers to friend of the show, John Palmer. He says he just found out my wife is leaving me. She said I'm not legible to capital. The legible to capital meme is fantastic. I do think that will he made. A new meme Fantastic coinage. I love it. It's absolutely ripper. We'll be using, we'll be using it. Well, let me tell you about FIN AI, the number one AI agent for customer service. It's AI that handles your customer support. Timeline is in turmoil over Nucleus.
I have one extra question. Go for it. Tell me a little bit about Dario's communication style. I was hearing a story about. I think Jensen has no direct reports or. No, everyone reports to him and no one reports to him. He has no meetings or all the meetings. And he likes 60 direct reports. 60 direct reports, but no big meetings. But he reads everyone's to do list every single day or something. What's it like at Anthropic? What is do Dario like as a leader these days? Yeah, Dario has a really, really cool communication style, which is that he quite frequently puts out very, very well reasoned essays. And then throughout Slack, we'll have giant essay length, like comment debates with people about different things. It's really great you get these, but the essays are really nice because one, you can go back and read all the past ones and it tells this here history of Anthropic. Yeah, it's, you know, I think in many respects, like it will be one of the better, you know, in a decade from now to chart the history of AGI. Sure. We'll be reading these like, compendium of essays. Yeah, and, and, and there's like, there's incredible comment threads on either side of them and so forth. But also throughout Slack, whenever where he's very open and honest with the company, whenever we're debating different things, he will lay out the pros and cons and how he's thinking about them and why this one's attention and why that one's moral struggle. And people will write back big essays on why they think we should do X or Y. And he'll respond. It's quite a joy. It's a very written communication style. As a result, it means that many people, or really the entire company have a good model of how he's thinking. And that really helps because it means that you sort of have a coherent sense of direction across the entire company. Yeah, that makes a ton of sense. I like that a lot. Yeah. Cool. Yeah. So many examples of successful founders who have adopted the written culture and seen great results, I think. And he's a great writer. I mean, read Machines of Love and Grace and it's just such a brilliant essay. That's great. You're absolutely right. Have you ever caught him using AI? Has he ever been like, oh, this one, he was phoning it in. Not yet, not yet, but maybe soon. I mean, it's kind of a bull case. If he does wind up just saying, could Claude, like handle it? I'm going on vacation for a couple days. I'm the dropping coworker. I'm pretty sure we measure loss on his essays. That's good. Yeah. Yeah. But right now, I mean, there's a high Bar. High bar. But congratulations. Thank you so much for taking the time to hop on the show. Yeah, super.
Are carrying on in this. Honestly, pretty lackluster. So, yeah, again, I don't. Yeah. At this point, Keon. Keon's been on the show. He's very funny, high energy. We've had some enjoyable conversations. But if I'm a potential customer of Nucleus at this point and I see just these series of exchanges, I'm certainly going to wait and see how things evolve versus signing up to use this service to. It's just so different than Cluli. It's so different than Cluli because if I use Cluli and I'm like, oh, the notes that were taken in that meeting weren't that good. Or if you go into Cluly being like, I'm going to cheat on this test, and then it's like, oh, it didn't work. Their engineers aren't good enough to really help you cheat on that test. You're like, okay, well, I probably should have been cheating. Cheating on that test. It's like the lowest stakes thing possible, but when you're. This is like literally deciding who your child. Offspring will be, it's the highest. The quality of the product could very well like infectious echo for generations. Literally. Literally generations. That's exactly. Not even just the child's life. Yes. The life of the child's child's child. Yes, it is extremely true. Child's child's child's child's. It could alter the course of history. I mean, it kind of could. It's sort of crazy. So, yeah, I mean, it's hard because viral marketing does work. Like, you know, moving fast and breaking things does work in certain contexts, but in the bio in bloodline optimization, it's really, really high stakes. And so you gotta be extra, extra careful. Extra careful for sure. Well, let me tell you about profound. Get your brand mentioned in ChatGPT. Reach millions of consumers who use AI to discover new products.
Or it's enraging to someone who's not the core audience, not the actual customer. And so it's okay. But it's a big debate because Sichuan Mala posted a long essay all about the claims made by Nucleus. Kiyan says everything levied unto Nucleus by Sichuan Mala is false. Worse than false. It appears to be architected by a competitor that has repeatedly published misstatements and inaccuracies. Sichuan is compromised, but it gets worse. Yeah. To be clear, no evidence has been provided that it was being levied by a competitor. Yes. That's purely an allegation that has no. There's no proof. Yes, yes, exactly. Yeah. Sometimes there's DMs that leak and there's evidence. Or someone comes forward and says, yeah, I was actually paid to post that. But so he says, I've been informed that Cremio. I don't know how to pronounce that last thing. Crimeo, who's been on the show also, he claims he's a race. Scientist in chief has been paid off by the competitor to promote this nonsense against Nucleus. For the independent scientists repeating the denied that as well. I would encourage you to do more diligence on who you're aligning yourself with. Our scientific team will issue a point by point response, which I believe they did. Unfortunately, though, this isn't about science. It's about. It's a concentrated attempt to cancel Nucleus on the backs of our successful campaign and build in efforts to build and advance the industry which benefits the very people attacking us. The mob are trying to cancel Nucleus. Keep tweeting, stay mad, we'll keep building. And serving patients. P.S. we won. The injunction link below. So they were sued by their competitor. But so they won the injunction doesn't mean they won the case at all. I mean, that's a classic thing if you're getting sued to be like, the case was dismissed and it's like one of five cases against you. Yeah, but in this case, they won. They want a preliminary injunction, which means that the case is just still progressive and they still have to fight it. A lawyer would file a preliminary injunction because they believe they had such a slam dunk case that they could prevent a lot of basically going a lot further and spending more money in the case. And so a judge might say, hey, this is actually. It's not clear enough for me to make a decision right now. We're still going to proceed with the case and give both sides an opportunity to continue to make their case. And then there was a little bit of like, A twist in the fact that Roy Lee, the founder of Clulee, apparently had worked at Nucleus Genomics. Very great. And Cremieux posts a screenshot of of Roy Lee back in February of 2025. So literally, just like months before he started Clulee. Very grateful to Kian and the Nucleus genomic teams for taking a chance on me the summer before Columbia and introducing me to the startup world. If I've ever seen a trillion dollar company and team, it's Nucleus. And Cremieu says he's obviously lying to cover up after getting caught doing fraud. An additional piece of background information people should know is that this fraud also employed the guy behind Cluley, the cheating company. And so Cremieu is being very hard in his assessment. Just actually calling Kian a fraud straight up is much more aggressive than just saying, like, you know, some of their claims are maybe not legitimate. It's unclear, like, you know, fraud is technically a crime that you need to be convicted of in the court before you are a fraud. But it's certainly, it's certainly he's putting his credibility on the line because if Keon comes out and says, like, yeah, I'm actually not a fraud. I did it. I proved it wrong. I mean, the main thing, the main thing here is it appears that the customer reviews. Yes. Are potentially fictitious. And if you're selling a service that allows people to pick their baby and you're giving and you're showing reviews from happy customers, that may or may not be real people at all, like, that just feels deeply wrong. Yeah. So I think that one of the first things that they could have done, I don't believe they have, is just say, like, no, our reviews are real. We used AI imagery because the, the people, the real people didn't, didn't want their identity online, tied to this service. Right. Yeah. For privacy reasons. Yep. But I haven't seen anything. I haven't seen anything like this. This guy Adi had had a good point. He said one of the core tensions in this industry is the fact that most companies recognize they're working on an incredibly topic. They know the general population will need to be slowly and tactfully acclimated to the idea of advanced family planning. Nucleus is perceived as polluting the commons with their deliberately inflammatory marketing. Their virality comes at the cost of increased skepticism for the whole industry. So, yeah, a lot of, a lot of folks were not very happy about that. Keon has replied. And if you want to dig into the actual scientific claims on either side, there are long posts where you can go through there. But obviously AI generated blog posts are alleged plagiarism in the Nucleus origin white paper errors in there, blatant falsification, terms of service are contradictory. Yeah, they also apparently they hired two people that had a non competes for 18 months. Those people just immediately started on working on Nucleus. Nucleus claimed that they weren't competitive so the non compete didn't apply. But if you look at the companies and what they offer, it seems very clear that they are competing. So anyways, very messy, very messy, messy story but. And yeah, I don't know. Will o' Brien says. David, I'm so sorry now man, but you guys are doing an absolutely terrible job at responding to this blog post and seem to be missing the point here. First of all, it is a huge claim to say that every claim by Sichuan Mala is false with absolutely zero evidence for explanation share receipts. Second of all, you make the claim that the person is paid off by competitors of yours, again with zero evidence. Third, that you make the claim that CREMU is paid off by your competitor. This is bogus and not true. But most importantly, you guys have made zero points of substance here rather than just insinuating, you guys are leaping ahead and others are jealous. You are selling a scientific product and someone has made a scientific critique in good faith, waiting to be corrected and explicitly saying they will make changes if they are proved wrong. And the best you guys can do is accuse them of being paid off and reply with memes. Not a great look. I want to see startups with a bold vision succeed, but how you communicate with the broader world is so important, especially with a product like yours and how you guys are carrying on in this, honestly, pretty lackluster. So yeah, again, I don't. Yeah, at this point, Keon. Keon's been on the show. He's very funny, high energy, we've had some enjoyable conversations. But if I'm a potential customer of Nucleus at this point and I see just these series of exchanges, I'm certainly going to wait and see. See how things evolve versus signing up to use this service to. It's just so different. Cluli. It's so different than Cluli because if I use Clulee and I'm like, oh, the notes that were taken in that meeting weren't that good. Or if you go into Cluly being like I'm going to cheat on this test and then it's like, oh, it didn't work. Their engineers aren't good enough to really help you cheat on that test. You're like, okay, well I probably should have cheating on that test. It's like the lowest stakes thing possible. But when you're, when you're. This is like literally deciding who your child offspring will be. It's the highest. The quality of the product could very. Well like contribute to echo for generations. Literally. Literally generations. That's exactly just the child's life. Yes. The life of the child's child. Child's child. Yes, it is extremely child's child's. Child's child's. It could alter the course of history. I mean it kind of could. It's sort of crazy. So, yeah, I mean it's hard because viral marketing does work like, you know, moving fast and breaking things does work in certain contexts, but in the bio, in bloodline optimization, it's really, really high stakes. And so it, you gotta be extra, extra careful. Extra careful for sure. Well, let me tell you about profound. Get your brand mentioned in ChatGPT. Reach millions of consumers who use AI to discover new products.
Be using it. Well, let me tell you about FIN AI, the number one AI agent for customer service. It's AI that handles your customer support. Timeline is in turmoil over Nucleus. Former guest on the show two or three times. Keon has founded Nucleus for IVF and he put up a subway campaign that says IQ is 50% genetic, height is 80% genetic. I completely disagree with that one. It's entirely skill based for me. Yeah, I had the genes did not matter. I had to grind for this view. Grind my growth plates, I suppose. Have your best baby is what it says. And it says IVF done right in the subway all over New York City. We. There's a ton of debate going on. And to be clear, I think accurate, I think it was intentionally trying to make some percentage of the population angry to drive enough energy and attention. So this was. Yeah, I would call it. I would call it rage bait. So I would call it rage bait marketing. Not necessarily rage bait, not at the product level. But IVF as a category is a controversial category. And so it's much easier to wrap it in a campaign that will go viral for upsetting reasons. For you can upset people and you can get a lot of attention from that. This is an example from Kath Korvac. She says, so Eugenics is profitable now. And so being able to wrap something that is just a, you know, a scientific process that's been worked on for. A long time seems to be somewhat friend.com inspired. Keon's original post. He says, Nucleus Genomics announces the largest genetic optimization campaign ever. Which is, which is just funny because a friend was saying this is the largest out of home campaign ever. And now Keon is saying this is the largest genetic optimization campaign ever. So narrowing it down, but full station Blitz at Broadway. 1000 plus street ads across New York City, 1000 plus subway car ads, dozens of urban panels throughout SoHo. And apparently they're not actually, they're not able to offer the service in New York. So that in here. So it's really just an image of a controversial phrase on a New York subway is more likely to go viral. So you do it there because it looks like you're on the global stage and then you pull. It's a high density of people that have a large following audience. Yeah, following. And so it's just the way to start a viral trend and own the moment. It's the reason why, you know, so many tiktokers are in Manhattan now doing stuff like man on the street stuff. It just like it has more like aura almost well. Dr. Shelby liked the mindshare grabbing that Nucleus did, says every biotech founder should be seeing this and understanding how to get 1/10 the mindshare of Nucleus. I have a playbook for you below. A lot of people are like, I love the playbook. I don't love this example because the company's getting dragged. I don't know if it's good or bad with the rage bait thing. I think usually it's a negative thing. Usually it's hard to come back from. Occasionally it can be done in a way that's slightly enraging, but enough people are in on the joke that they appreciate what's happening and they appreciate that it breaks through or it's enraging to someone who's not the core audience, not the actual customer. And so it's okay. But it's a big debate because Sichwan Malla posted a long essay all about the claims made by Nucleus. Keon says everything levied unto Nucleus by Sichuan Mala is false. Worse than false. It.
You're watching TVPN. Today's Monday, November 24, 2025. We are live from the TVPN Ultradome, the Temple of technology, the fortress of finance, the capital of capital. Ramp.com Time is money save. Both easy use, corporate cards, bill payments, accounting and a whole lot more all in one place. Jordy and I went to F1 this weekend. We went to Las Vegas and we watched great time. We were with the public.com team. Ramco Aston Martin F1 team with the public.com boys. We had an incredible time. Yeah, incredible time. And F1's fun. It's more of an experience than like the actual watching it. I don't know if you at home haven't been following it is McLaren was disqualified. It was a very dramatic. I mean, it's almost like a cheating scandal. I don't exactly know, but they broke the rules. Not cheating because the FIA said we don't believe they did it intentionally. Okay, got it. Yeah, but it was like, but it's a disqualification. And so you would think you were like, oh, yeah, I saw, I noticed it. And there were some people in the comments on some of these videos that I watched saying, oh, I could tell, I could tell. But let me tell you someone who was there in person. I could not tell because it is Wiz. I don't know if they're cheating or not. I don't know if they're real. It's an incredibly fun time. It is a terrible spectator sport. I think everyone agrees on that and. And the two things can be true at the same time. But it's a great reason to come. Come together with like, if you actually wanted the best experience for the race, you would just sit inside the paddock. You would watch restream one livestream 30 plus destinations. If you want a multi stream, go to restream.com Anyway, a lot of fun. And Janek Life and Sykes were incredible hosts. Today on the show we are talking About Claude Opus 4.5. We have Sholto from Anthropic joining us. In just half an hour, maybe 24 minutes, he'll be joining. The timeline was in turmoil over the weekend. People are settling into the idea that Gemini, the three might be good enough to actually pull some people away from ChatGPT as a daily driver. It certainly pulled Mark Benioff away from ChatGPT. He of course, partnerships. He was swearing on the timeline. He was swearing on the timeline. He has partnerships with a number of. A number of foundation labs. Foundation model labs. But he says, holy shit, I've used ChatGPT every day for three years. I just spent two hours on Gemini 3. I'm not going back. The leap is insane. Reasoning, speed, images, video, everything is sharper and faster. It feels like the world just changed again. And this is an interesting experience. I had a similar experience. I wound up basically daily driving Gemini. I didn't fully churn, I didn't delete ChatGPT from my phone. It wasn't intentional. It was more like, I'm just curious. I really want to use nanobanana Pro. That definitely just sort of sucked me into the ecosystem. But I wrote a little bit of a review over my experience. I know you've been a jemmy boy for couple weeks. You look great. In hindsight. You were early to this part. Honestly, way longer than that. I was maybe months. I was, yeah. I mean months at this point. Yeah. And so there was some good stuff, some bad stuff. Obviously. As a disclosure, we are of course sponsored by Gemini 3 Pro, Google's most intelligent model yet. State of the art reasoning, next level vibe coding and deep multimodal understanding. But I mean, I'm going to try and be as fair as possible with this review because there are some things that I do want them to improve in the consumer Gemini app because I think there's a lot of opportunity there and I'm just not sure how monopolistic consumer AI will be. And that was a little bit of what my takeaway of this experience was. So basically I switched over. I've been on Gemini on iOS for a while, mostly to access VO3. VO3 was the moment when I was like, okay, they got something that nobody else has. I got a 4.0. We're giving them 250amonth. Well, and then it switched. No, no, it was 125 and then it jumped to 250. Okay, it wasn't the. Yeah, I thought it Might have been 500 as well, but it's 250. Been playing it. Very happy for that. VO3 is just a very special model that no one else had anything close to it. It was very accessible on your phone and I enjoyed it. But I switched to daily driving Gemini on iOS as the main app that I go to for all the different knowledge retrieval requests. Anytime I'm researching something, I would hit Gemini in the app and the result was around 15 minutes per day in the app. And this is roughly the same as what I spent in ChatGPT. Historically, I looked through my time, my screen time. Now that doesn't count stuff on the Desktop, maybe it's a little rough, But I think 15 minutes a day is sort of what most people are doing in these apps. Obviously 30 minutes a day was reported. Benioff said he spent two hours in it. I think he was just like maybe in a fugue state doing deep dives or something. But I had a more passive experience where when I had something I was curious about, I would fire off a query and there was a lot to like about the experience. So first it felt like Gemini 3 does a better job sizing the response. Like if the question can be answered in one paragraph, it gives me one paragraph. If it can be answered in five little subheaders with little bullet points, it'll do that. If it needs more story, more history, it'll write more. And so I felt like in previous models in ChatGPT, certainly I felt like I was falling into the trap of no matter what question I would ask, I would get the two page dissertation on it with the same structure because it was a little overfit on the format that it was delivering. Gemini 3 felt a little bit fresh there. It also felt faster. Everyone's been saying it's so much faster. I haven't seen any quantification of that, but it certainly felt like it feels faster. But I think a lot of it for at least for me is that when for the last couple months when I've been on ChatGPT because the model router gives me anxiety about like, oh, maybe I'm going to get routed to like the dumb model, it's going to hallucinate. I'm just hammering GPT5 Pro because I'm on the $200 a month tier. And so because I'm on this $200 a month tier, I'm used to hitting GPT5 Pro. But then that always means I'm waiting 10 minutes. And so if I'm always waiting 10 minutes and I go over to thinking and it's like, oh, it'll be one, even if I'm on a different model, it's not as much reasoning, it feels faster. And I feel like the level of confidence in the brand makes me feel that a Gemini 3 thinking query that does maybe less reasoning than a GPT5 Pro query will be at the same level of reliability. And you've pointed out to me something about when it's actually running, it does something psychologically that's really valuable. It's smart, tells you it says it's. Running a Google search, it just says we're searching Google and you don't think about it because everyone, oh, searching the web. And I'm like, but I don't trust the web. But I trust Google because Google's had 25 years of building brand around trust on the web. And so I see that now and I'm like, oh, yeah, good. Because that's what I would do to verify a fact. Even though the web is Google, obviously there are hallucinations out there, there are fake articles that you could land on. There's a whole bunch of things. But if you, if you task me with finding the real day that someone was born, I'm going to Google it. And so I trust that as a product. And so putting that there. Definitely did. Yeah. Like it had a real perception, which I think was interesting. And then also nanobananapro, very interesting. Strong differentiator. It really does handle the complex images. We saw that with the farm and also just all the text and stuff. And it's been interesting to kind of throw a query like I was, I wanted to understand Anthropic's model architectures and I said, hey, summarize them all. And it infographic and it just perfectly explained how Sonnet and Opus all fit together nicely next to each other. I don't know that it's necessarily a better way to learn, but I could imagine in the future having images generated alongside text just means that you get a more richer multimedia product, which should be the result. Because if you look at any like newspaper, any website, there's always. It's not just pure text. Like walls of text are boring. Yeah. In fact, if it's just pure text, it usually means the story is just not that important to the newspaper. Yeah, yeah. There are some people that just lean full text, you know, and whatnot, but for the most part it's much more enjoyable. It's just better, it's more educational. It's easier to learn quickly. Let me tell you about cognition. Before I go into the negatives, let me tell you about Devin the software engineer. Crush your backlog with your personal AI engineering team. So on the negative side of my Gemini app experience, there were a few rough edges. The first was with that multimodality. Everyone's been saying these models are multimodal. They handle image, text and video. I don't know if it was just a UI issue, but I was running into tons of problems where it wasn't feeling multimodal. What I mean by that is that I would go and it would issue it an image prompt. I would issue it an image prompt. Hey, create this infographic, and then I would want to flip back into text and it would not be able to really stay. It wouldn't be able to go seamlessly back to text mode. It would keep generating images, and then vice versa would happen where I would kick off a text. A text flow, and then I'd say, okay, I'm ready for you to turn this into a nano banana thing. And it'd be like, I can't really do that. Do you have any. Are you laughing about that because you think it's like a rookie mistake or something. Well, no. You always like, oh, it's not really multimodal. It's not really multimodal. Yeah, but there should not be a button. If there's a button, it's telling on itself. Why is there a button? You open it up. You know what I'm talking about, right? I mean, I just don't see why it matters. Like, if you can basically just take an image and then turn it into, like, the textual representation, why does it matter that it's not, like, actually taking in the pixels of the image? I just think, like, it's. Yeah, I mean, I guess you're right on that front. I find it weird that I need to. I mean, it is multimodal in the sense that, like, everything gets baked down into, like, tokens. True, true. But it's. It's just I expect the models to be operating at a higher level of abstraction much earlier than I think they do. And so with the model picker, like, I never liked that because the model should pick based on the text. I really like the router in ChatGPT because I should be able to go to a person, which is what we're trying to recreate here, and say, like, hey, I have you. I have a research project for you and I need you to spend 20 minutes on it. I need you to get back to me in an hour. I need you to get back to me right now. Off the top of your head, what's your hot take on this? I can ask that and I can get that back from human, and I feel like that should be done at the text layer at the end of the. I mean, it kind of is in ui, like, if you ask, like a thinking model. It does now. It does now. But what I'm saying is that we are still in the pre, like, selected dropdown UI functionality of Gemini because I. I'm prompted to pick what I want to do. Do you want to do image, video, deep research, text before you go into the flow, instead of just saying, I'm having a conversation, oh, now is the time to generate an image. And it's like, yeah, sure, that's something I can do. Instead of being like, whoa, whoa, whoa. You didn't ask to talk to the guy who can generate images. Like, that guy's over there. It's like, is it all or is it not? And it's clearly not. And they're upfront with you about that in the model picker when you're. And in the ui. But then it feels like the marketing is a little bit like it's omni. It's all the things. It's multimodal. And I'm like, it doesn't feel that multimodal in the ui. So I don't know, maybe it's something that they'll work on. But there were a couple other, like, rough edges, and most of it is contained in the UI layer. So one of them is voice transcription mode, which I've, like, been completely using in ChatGPT. I'll just open it up, talk to it. Now, it's not the voice mode where you talk and it talks back to you. I don't like that mode at all. You're just using voice as an input. Exactly. Just voice as an input. And so I'll click the little microphone button, talk for a while, and give it a bunch of context on. Okay. I'm interested in the history of Gemini, and, you know, why don't you take me through some of the VCs that backed up, you know, thinking about it. And then I say, though DeepMind, Demis's company before he got acquired. But then also I want to know the history of. Of Google Brain. Like, where did that come from? Was that. Yeah, that. Was that acquired in. Would they acquire different people, or did that just get spun up internally? And I'll have pauses and I'll come back to things and I'll kind of. Like talking to an employee. Yeah. So I'll just give it a lot of context. And when I give that to ChatGPT, it loves that. And I feel like it gives it great context because it has a whole bunch of stuff. It can transform it. With the Gemini app, it will cut me off and be like, oh, you paused for a fraction of a second here. I'm submitting it like I'd like. And I'm like, no, no, no, no, no. You need to. You need to take more time to let me finish. And with ChatGPT, like, there are two different buttons. You can click the stop button and it will translate it into text and then you can review the text and say, oh, okay, it made a terrible mistake. I don't want it to burn two minutes on something that. And get confused. I'd rather like one of the prompts. I was like, I was like, generate an image. There was this meme that was going around in Nano Banana world where it was like, generate me an image of the most annoying LinkedIn profile picture. And I had no idea if it was real or not. It might have been people just taking screenshots and then just, you know, dunking on people. Well, some people were just taking a screenshot of someone's actual exact. Exactly. And. But I was like, I wonder what happens when you actually take that prompt and you put it in there. So I go to. Go to Gemini, put that in there, and it doesn't realize that I want to actually generate an image of that. I say, generate a LinkedIn profile of a most annoying person. And it doesn't know that I want an image. So it just dumps out a whole bunch of text. And then I open up the audio and I say, like, no, I want you to generate it with nanobananapro. And what it gets from that is Banana. Banana Pro. And the result. And it's trying to be really friendly. It's like, I love the enthusiasm. Let's talk about bananas for a little bit. No, I want you to use Banana. My criticism is just that the. The Gemini app still has a lot of bugs. It just has bugs. It just has bugs. It was also. I can get over it. Yeah. For now. Because again, it's like, it's fast and smart. I mean, truthfully, right before we joined, I was doing a search and I had to like, it was stuck in this limbo where it wasn't running the prompt, but it wouldn't let me run a new prompt. And I just had to basically rage, quit and restart it and just copy and paste the prompt into a new box. So a lot of this, I mean, it's. Yeah, again, it's. It's incredibly impressive. Yeah, it's a great model. But they have. At this point, it's just like, opportunity to like, get more competitive on the product side. Yeah, totally. Yeah, I, I was noticing even like just straight up disconnection errors. Like, I would submit a prompt and then it felt like if I close the app, it would get confused or something. And I don't understand that because just sending a little bit of text. Have you ever run into this? I've had that a couple Times. But it's funny, you can kind of think of the app as being like a benchmark of the model. Right. Because you should imagine that the. That they model should be. Yeah, they should be using the model in like, the cli. I agree. We got to hold Sholto's feet to the fire on this. Yeah, he's so good. So we're going to test out the Anthropic website and see how good it is. And if it's not good, then obviously the new cloud model is not good at coding. Yeah. The app should be flawless. If I find one bug in the Claude consumer app, it's over. Do you guys ever use the, like, voice to voice, like the real time audio thing on ChatGPT? No, I don't like that at all. You've never used it? I've used it. I've used it a bunch. I've used all of them, but it's just not the preferred way of interacting. Yeah. You were testing it out, Tyler, by talking with it for like eight hours a day. Right? Yeah. And you were on the X XAI one. Yeah. With ani. Was that. Imagine running constantly with a VR headset. With a VR headset and a full immersive suit in a sensory deprivation tank. Yeah. No, no, no. Why do you bring it up? I actually, I've started using. I started using it, like, it's pretty good. I think the model is actually much worse. Like the underlying model. Yeah. It has to be faster, right? Well, it's not the speed. It's like the actual intelligence of the model seems lower. Yeah. Like, the answers aren't as good, but I find it's useful for when I'm trying to learn, like a specific topic or something. And I. And then I explain it back. Yeah. And then it tells me like, oh, is that correct or not? That's pretty good. I like that. Yeah. It's really remarkable. I mean, the Gemini app launched almost two years ago and there's still, like, rough edges in the ui, which I think is crazy, but it does seem like they have an opportunity to actually take some serious market share at this point. Like, they've caught up on many different values and, like, value props. My question was, like, I'm not the typical consumer. Like, I'm going to try every different app. Like, I'll probably keep bouncing around. I don't know if consumers will do the same. Broadly, there's. It's very, very clear that ChatGPT is just synonymous with AI and people are not like, oh, well, like the new benchmarks I gotta like change my app. Like no one's thinking like that. But the fragility in the ChatGPT monopoly aggregator thesis that I was picking up on was the last year there have been a lot of features and theses around different things that could create lock in. So stuff like personalization or memory or even the chat functionality between what you've linked your custom instructions. The different. I think at this point I've synced ChatGPT or Auth ChatGPT with a number of different services so it should have more data, it should know all these different things. I've given it even custom instructions just saying like hey, cool it on the EM dashes and, and I didn't miss any of that. Like I. There was at no point where there were plenty of points where I was like oh like chatgpt is definitely better than Gemini still. But at no point was I like it's because it doesn't have personalization. And I think that if I went in Gemini and I was like oh yeah, like you can go, you can go take a peek at my Gmail to get personalized, like to understand how I write or understand, you know, what I'm interested in. Like one I could snap my finger and Google could like be way more personalized maybe. But the biggest thing is that like right now I feel like both are not personalized at all. None of them have any real lock in of any sort. And even in like the chat functionality or like the social network functionality which is just very different than what happens in a true social network or where there's this flywheel and the content is driven by the existing user base. Whereas I feel like I got on Gemini and on day one the content was as good or better than ChatGPT because it's all AI generated. So it made me think like maybe it's a little bit more fragile, maybe there will be a little bit more of a duopoly, it won't be such a winner take all market even though it has been historically it has been up to this date. Like in consumer AI. It's very clear that OpenAI has run away with it. But it feels like Google does have a little bit of a chance to catch up in consumer because there's just so much less of a network effect. Like the network effect just is bolted on. It's not real yet. Maybe it'll never be real. Maybe Google can catch up there. But I just really want to see where DAUs and and user minutes actually grow because there's so many different tweaks there and every chart and data point is definitely going to be analyzed to death. Yeah. One thing that's notable, Google's going super hard and it's for students. So you can just get Gemini Pro for a year free. And again I think that's just a bet on like get people hooked on the workflow. I mean OpenAI is clearly battling that out too. One of the big value props of using the Atlas browser is you get more advanced thinking queries like they will up your limits. It is interesting. There's also I'm very interested to see who can bring ads online faster. Like Google should be able to snap their fingers and do it so quickly and yet it does seem like something that could just take them longer on a product side. But they should have a whole model. Like they should be able to do display ads like right now and just be like okay, yeah, our free Gemini users are now properly monetized. Now maybe they don't want to be the first mover there because then they'll get the stink of like oh, they're the ad one in the market. I don't know. I think, I think it matters a lot more to just have a highly competitive product and win market share before you spend any time with that. Like for example, like if you're using a, a version of Gemini that's super smart and fast but still a little bit buggy and then you start seeing ads, you're like just make the app like perfect before you or as close to perfect as you can get it. Yeah. Before you introduce ads. Yeah. Well let me tell you about Adeo, the AI native CRM. Adeo builds scales and grows your company to the next level. Google has now added $2 trillion to its market cap over the past 20 months since the boob shirt guy asked Sergey Brin about woke Gemini images while having a foot long subway cold cut trio for lunch. What is this video? Let's play this. I have no idea what's going on here. Let's see if it's Gemini art the back. Yeah. Okay. Yeah, I wasn't really expect to talk about this thing but you know we definitely messed up on the image generation and I think it was mostly due to just like not thorough testing. He's a crazy shirt to be wearing. I don't even know how you get into a meeting with someone as powerful and wealthy as Sergey. Obviously just wear, you wear a jacket. You wear a jacket and you get in, it's hot, you take your jacket off, you're just. I Was not expecting that. That is so insane. That's very, very funny. It's Bay, It's a Bay Area thing, John. Yeah, yeah. Wild, underrated. Like Sergey Brin really did go into the Gemini team very clearly like gra. Like, like was like it's time to go and cook. Like let's work on this. And like it clearly had results, which is awesome. Yeah, it's notable. I mean this stocks jumped 6% today. Barron's put out a report today. Just saying the title is by Google stock. Yep. Alphabet has been the clear AI winner. Which is just funny because earlier this year like people weren't saying. People were saying they are the AI loser. Yeah. So Barron's is saying actually they have been, yeah, they have been the clear AI winner. But the narrative has. And our in house retail trader has been on an absolute tear going long Google. Congratulations to him. Well, congratulations to you because he, he had, he was in probably one of the most favorite retail. He was in a dark place. A certain company announced a partnership stock Moon. He got out, he's like John, what should I buy? And you're like buy Google. I was like, just play it safe dude. Just, just go with something. Something safe, something not crazy. Do not use leverage, please. Well, Google. Google's parent company Alphabet has acquired a stake in physical intelligence. That is of course Lockheed Groom. Exciting Locky and Carol Houseman, the co founders came on the show about six months ago. We should have them back on and check in with them. The San Francisco based Physical Intelligence is an AI and robotics startup. They've secured $600 million in fresh funding pushing its post money valuation to 5.6 billion. We should ring the gong for hit. So Capital G is in and then Lux Capital, Thrive Capital, Jeff Bezos is in Index Ventures, T. Rowe Price. They're building a general purpose AI foundation model and learning algorithms. And they, they've, they, yeah, they focus like not on as much of like the, the flash and substance. Like, like not as much like flash and oh, we're building a full humanoid. More like, you know, we're kind of taking like incremental steps towards adding value in different robotics cases and demos. The laundry folding robot of course, and now the coffee making robot. All these are very cool and it just feels like they've taken a. It's notable that this is at 1x speed too. We've seen some other demos. Remember Sunday Robotics was sped up like 10x. It was 4x I think but it was different scenes. Oh really? There were some that were 10 okay, yeah. Yeah, very cool. They should just design an espresso machine that makes espressos automatically. They could. Has anyone ever done that? You know what they could do? They could put the espresso in a can and then they could mail it to you and then you crack open the can. And if you crack open the can, that's maybe. Or no. You know what? They need a robot that opens the can, then you need a robot to open the can. Yeah. I just wanted to know. This is an interesting task. I'm sure there's a certain number of people in the world that get really angry seeing this because this is one of. I think, I imagine this will be one of those things that even when robots can do it, people still like to know who's making their espresso. It feels like there's been a couple robotics like robotic coffee shops and stuff. Well, our first guest of the show is in the Restream waiting room. We have Sholto from Anthropic. Welcome to the Stream. How are you doing? Congratulations. Thank you so back. Thank you so much for taking the time to hop on on such a massive day. How is this just Claude Opus 4.5 day. What.