Weekly Recap | Elon vs. Trump, Ukraine's Drone Attack, Cluely Update & OpenAI CRO

26:45Maybe we need a tinfoil hat. Who knows? But yeah, I mean, when you're in the world, it's like, hey, we're only down one percent. Let's go. The coup of the century. My question is, has there ever been a crash out of this magnitude ever in history? Well, you know, history when when Elon and Trump became friends, I actually probably world history equivalent. I feel like there's something in like that. Maybe didn't have it in the United States where, um, you know, they're crashing out. Crashing out used to mean calling up the New York Times and just ranting. Now you can just live posts like all your reactions and it's just all real time.

27:23This is like crash outs are actually intensified. You actually want to be long crash outs. Yes, definitely. The next. And he has their own social media platform that they own. So, you know, you got to be on both X and Truth Social to like stay on top of things. Yeah. Yeah. I actually did like a deep research report a while back on like, has the richest man in America ever been close with the U.S. president going back to like, you know, was Rockefeller particularly close? And because the narrative was like, oh, this is like so unprecedented. And in fact, it is unprecedented. Oh, really? Yeah. Yeah.

27:54I would have guessed that like Rockefeller was close. Me too. Me too. That's what I was going for was like, no, I imagine this is always it is always close. But no, I think because the president has become more powerful globally. Your your your your your point about, you know, mayor of America, dictator of the world, like it becomes increasingly valuable for the richest man to have a close alliance. And so it's become more. I don't know exactly how accurate that research was. It's totally possible that, like, behind the scenes, Rockefeller was really close to the president at the time and we just didn't write about it in the history books.

28:24But there certainly aren't very many anecdotes about the richest man in America going on. Yeah, so Pavel had a great... Ready for AP U.S. History 2050, you know, A, B, C, D, E, M. Yeah, so this is where Elon Musk called the president at the time a potential pedophile. Was it A, about Epstein Island, B, about a cave in the Philippines, C... What a mess. No, so Pavel had a good post. He was quoting the big bomb from Elon. He said, hypothetical question about the USA's power structure. Is the man with the most access to capital more or less powerful than the political head honcho? Purely hypothetical.

29:03It's a good question to ask. I mean, I think both like archetypes have grown both in absolute power, but also in relative power to the rest of the globe, basically since the Gilded Era. right if you think about like the president of the united states in 1925 i'd say pretty darn powerful but like there was clear like you know it was a you know sort of multipolar you know sort of world argentina was pretty darn rich at the time obviously europe was still you know sort of recovering from world war one but uk was generally you know sort of doing well like it was not you know it was clear there was a you know sort of huge you know outweighed effect and then if you look at probably the you know sort of biggest you know industries at the time you i don't think you could claim that even like standard oil at its peak, I'd have to go look at the exact numbers, but that like it had the size of budgets relative to like, you know, they're like us government in terms of, you know, sort of budgets, right.

29:53Versus I feel like now for the first time you both have, you know, sort of us president, extremely, extremely powerful. And then you have like, you know, sort of mag seven effectively, like the size of, you know, sort of, you know, state governments. And then also just more bureaucracy, more red tape. So like I, when I think about the 1920s, like Robert Barron's, it's like, it is the, it is the, you can just do things era. And so you want to build a railroad like, yeah, you might need to get like one rubber stamp, but it's not going to be 10 years and tons of lobbying and all this different stuff.

30:23So you can kind of just go, uh, you can just go, well, you know, it's bad when Kanye is saying bros, please know we love you both so much. It's just like the voice of reason is Kanye West. Thank you. We need to bring them together and, uh, you know, form a peace treaty. Nikita Beer just added his pronouns back to his bio. Let's go. You've got a rubber band. Elon's got a rubber band all the way back to extreme wokeism, straight back to super climate change. Wow. Somebody's resharing the picture of the Cybertruck blown up in front of the Trump Tower. It's just like this. This is in real life. It was foretold.

31:03But it was a question of when and what magnitude, not if. always bad if if vladimir putin is operating to negotiate between uh president trump and elon i think i think a lot of the world is is waiting for roy lee's take clearly and the clearly army they want that people have been asking him to get involved with geopolitics wow um i love the uh shiel moha put up a uh you know sort of meme about uh narenda the like uh prime minister of india um you know he basically uh copied and pasted the uh trump truth social post about negotiating peace between india and pakistan when it wasn't like actually fully negotiated and uh you know posting about you know uh you know negotiating a ceasefire between uh elon and trump funny thing is like truth social you can just read all of trump's posts without creating an account.

31:58It truly shows that like I would think that you would have to make an account to read them all But they just that it's not gated at all It's just this could be the biggest But you know, they clearly I don't think they care about monetization Bitcoin is actually falling alongside Falling Wow Bitcoin falling Boeing falling Tesla falling who's the biggest winner of the day. I think it's China China, yeah, China China Sean McGuire Bitcoin really sold off. It's down 3 % today at 101K. So still up, but rough. Winnie the Pooh just dipping his hands in that pot of honey, just snacking away, watching from the sidelines.

32:38Yeah. Let's see. Chinese stocks, U.S. stocks. I can't find anything. That's probably my commentary on the day, boys. Anyway, this is great. It was fantastic. Thanks for jumping on. Thanks for hopping on so quickly. Cheers. Next up, we have Roy from Clulee coming back for an update. He's hired 50 interns, I think, or something close to it. He said they're bringing every intern on. They're bringing every intern on. We got every intern coming in. Well, welcome to the studio, Roy. How are you doing? Oh, boom! Let's go. There they are. I think we're overpowering you. Can you hear us? Yeah, yeah, we can hear you.

33:17Fantastic. We're zoomed out all the way so we can see everybody. We got a small army here. This is incredible. How big is the team? Kick us off. How many you got at this point? The team is 11 full-time plus the interns. How many interns you got so far? Interns. Bro, we're closing in on 50, brother.

33:37That's amazing. Congratulations. What are they all doing? How do you manage everything? Is it purely social media? Is that what you want them to focus on? Growth? Yeah, yeah. Growth marketing. The only goal of the company is get 1 billion eyeballs onto Cluey. So we have unrestricted creative freedom and permission to do anything and everything. Just make the company go viral. Every single person you see behind you has over 100 ,000 followers on some social media platform. Wow. 200 ,000 plus. That's remarkable. Me too now. Yeah? There we go. You probably popped. What's working? what platforms have actually been driving the most growth?

34:18I mean, I'm sure you've run a lot of tests. What have you learned that you can share? Bro, Ben, take it away, bro. Let's hear it. UGC has been really good. We just hit 10 million views today. It's been eight days. Wow. There we go. Hoping to get 100 million views in the next month. What platforms specifically are the most fertile ground for targeting your specific customer? because you can imagine that there's a lot of folks who are AI curious on X, but then there's much broader, more viral audience, more general audience on platforms like TikTok, YouTube, Instagram, what's working and what is the next, next platform that you're going to be focused on?

34:56Yeah, well, we're trying to go viral on every platform regardless. But the main thing right now is Instagram Reels. Oh, Instagram Reels. Interesting. And what is the main value prop that you're hitting people with? Is it still the cheat on test thing or have you evolved at all? what still i can interviews yeah interviews okay and uh has there been it this was controversial when you launched it is it still controversial in the comments are you getting flamed has anyone big dunked on you and has that driven virality is that actually a net positive instagram is not like twitter like you could post the craziest shit on instagram and they will still not think it's controversial really so how to make it controversial like we have to engagement bait some other way.

35:38Like, it's teething tool is controversial on Twitter, but on Instagram, you could have, like, a white guy say the N-word ten times, and it's still not controversial. Like, you need crazy shit on Instagram. That's what we crack. Every single person here has, like, very great viral sense. And watch the reels that do go viral. You see there's, like, ways that we've engagement baited the videos, and this is what we'll keep doing to probably a billion views a month. How long does it take to figure out if an intern is cracked? Is it, like, an hour or two hours? How much time do you need? for me personally me personally probably like 10 minutes but for anybody watching probably would take like one one or two weeks there we go there we go uh how do you guys think about how do you guys think about product marketing obviously you're just going viral everywhere getting all this attention how do you make sure that it that it uh while he's shaking his head doesn't think about it's not about the product it's about the attention attention is all you need.

36:29You guys can make anything go viral. Yeah. Yeah, but how do you... Up the side of the street, you know, you make some UGC videos, make some Twitter posts, you know, you can sell anything. In 2025, product doesn't matter. I could jack off off the side of a building, sell some videos of it for$20 each, make$2 trillion. It's crazy. $2 trillion, that's intense. How do you guys think about Burn? Is it on your mind at all? I don't know if you saw the last tweet but as of literally like two days ago we're still we're still cash flow positive we're still fucking profitable we're still wow let's give it up to the property let's hear it uh it sounds uh so you're charging for the product and people are paying are they at all satisfied or do they feel uh like they got scammed bro like the product works you're either using this as a consumer and it's working because like like you're passing your interviews and or if it doesn't work you're not going to complain to me because i'm gonna go right to your employer and tell them yo guess complain about using the product like i'll get you blacklisted if you complain really um where how are you thinking about how do you how are you guys thinking about product evolutions what do you want to add to the product obviously uh you want to help people cheat on everything where where are you going to help people cheat next we don't care about like the product is going to be led by the virality of the content we have video ideas right now that we're going to try to push for different use cases we're going to see which ones go consistently the most viral you can It makes them go more viral.

37:52Then you can just build the technology after you have all the attention. So we'll figure out the exact use cases and exact niches we're going to quintuple down on once these guys get to work. What formats on Instagram Reels are the most modern in terms of consistently viral? You mentioned man on the street interviews. What do you do for a living? That's always been fertile ground. What about I see a lot of those mobile game ads that look like you're fighting down some sort of bridge. and then you go into the game, it's actually just a match three. Um, what, what, what are the different formats that you like to pull from every week?

38:29There's two new ones. And at any point, there's probably 10 to 20 viral trends that is happening. And these cycles so quick, you need to keep your finger on the pulse. These things will like expire immediately. You need to be on the ball. And like, like if I, if I told you right now, by the time people watch this on YouTube, like it would have all been expired. Well, they were live. So, so give us, give us the latest and greatest like what's going viral today well right now we got 10 million 10 million views using a snapchat format okay viral for like the last three years to be honest okay and and i i think that like we just have to get people who continuously scroll tiktok like six hours a day yeah but what's the what's the actual format that you use like describe the video what is the hook like break it down for me like you're explaining the like the art behind the viral format there's a caption it starts with the face usually a handsome dude or a pretty girl you're saying damn this interview starting with the interview is starting with the hard questions i should have been a cs major not a business major that's engaging because people are saying like bro like cs is way harder than business then it turns around the interviewer asks like like hey how are you doing why should we hire you and then this guy uses clue to generate a response but he reads it hella autistically like oh i revel in detail and and then that's that is like another conversation point like people are cooking on the guy because he uh he can't read properly the guy is like a doing a really dumb interview using that's great how are you guys using uh ai generated content internally i know a lot of these the videos that you guys are creating are just typical social media vertical video do you have an intern that's just generating uh basically copy and pasting, making a bunch of other...

40:09B-Roll, VO3, are any of these tools relevant? Anything clicking? Not yet. I think there's still like a 10 % left before they cross the uncanny valley. And the biggest thing is that people need to think your video is real. That is the difference between 100K views and 10 million views if people think it is real. Absolutely AI CEO is bearish on AI. What about... Google needs like 10 more Chinese researchers to like figure it out. And once they push out the latest update, then then vo3 will be there but right now we need real people yeah yeah uh well i mean what about just using ai as like stock footage replacement not not as the lead-in for the video not the entire video but just like sprinkled in to illustrate a point you know an establishing shot of like a building a helicopter pulling into a building like that that historically has been kind of something that you would reach to uh you know adobe stock video for vo3 feels like it's there, but are you not drawing on that at all yet?

41:10If there's a viral format when we need it, maybe we'll use it. But right now, like it's, it's really brain dead to go viral on Instagram. Yeah. Formats are not hard. You don't need a helicopter. You need a guy, a camera, a really shitty camera. I need a computer. I mean, what about like those, those kind of like AI mashups, like Harry Potter, Balenciaga or the, uh, the kangaroo with the plane ticket getting on the plane, like, like AI content can go viral when it's really, when it's like inspired almost by a human. It's not entirely AI generated, but it's using the tools effectively to create something that's like still catchy.

41:44Do you think you'll be using any of that anytime soon? Probably very soon. We're scaling up. Like what you see right now is probably about less than 1 % of what the size we will be by the end of the year. Like we are profitable. We're not trying to be profitable. We just keep making so much money. We can't help it. So we're really scaling this shit up. I'm not even trolling you 1000 and creators are going to be shipping out content. We're doing a complete internet takeover. Okay, so why in-house? Why do they even have to be employees? Couldn't you turn this into a multi-level marketing scheme or something, a pyramid scheme?

42:16We're going to do this. Oh, that's what you're going to do. Okay. MLM. MLM. I love it. Are you guys worried that you could be infiltrated by journalists? I'm sure they're circling the house right now. The hit pieces are going to come. We're doing a softball interview right now. I mean, the person that's brave enough to try to do a hit piece on the Cluley army. It's going to be. I bet they're dying, too. Look, more eyeballs is better. There's no company that ever died from a founder being too controversial. You got deal fucking infiltrating with genuine spies, and they're still doing fine, bro.

42:51Worker 17 guys, they're still kicking. Like, no company ever dies from being too controversial. You die because you don't make enough fucking money. Yeah, yeah, yeah, yeah. Speaking of making money, what's the pricing model right now? Are you doing anything on price discrimination? Is there a super high tier if you get a whale? What does a clearly whale look like? Can I spend$2 ,000 a month on this service? Yeah, you should add a tipping feature too. People should be able to tip you guys if they have a good experience. You should add really financialization, pay as you go, high interest rate loans.

43:21Just really push it. Make it sports gambling in there maybe. Just throw it all in. Yeah, I mean, it's$20 a month for a consumer,$100 a year. and our top line revenue is really being driven up by enterprise you're gonna have to talk to the sales team to get a custom pull but you know like there's a lot are you serious what whatever is that more on the sales side what who are the enterprise so you sell the sdrs you guys laugh because you think i can't sell enterprise because i'm no i i don't believe it i trust like these 20 these fortune 500 ceos like these are like 35 year old dudes who sit there squirt or laughing at my post yeah yeah yeah yeah no no it seems it seems legit it makes sense no i i believe it but but i mean you're not going even higher tier like what's the two thousand dollar a month clearly vision for for consumer there's a lot more we can do with more compute but right now we're like to be honest i didn't expect to grow this fast the engine team is quite small i'm like spending a lot of time trying to hire more more competent engineers we have a lot of back block tasks that we need to fill out especially for this last contract that we signed so we're full-time I'm focusing on the one big guy that we got right now.

44:26And after that, then we'll try and scale this up. But right now we're focused on the one big client that we signed. Yeah. Talk about your compensation strategy that people want to know. You said you can raise infinite capital and you're so confident. I believe you. But I'm curious to get some more insight there. Bro, I feel like it's so retarded to be a company. Sorry, am I allowed to say that? No, you're not allowed. No, this is a family-friendly show. It's very stupid to be a company. Try to race to the bottom to see how little you can pay your employees. Bro, if I'm making hella money, we're all making hella money.

45:01I'm trying to pay them more to see if, man, maybe tomorrow we'll start being cash flow negative. But I would like to pay these guys what they're worth. And the output is fucking insane. We did 10 million UGC views in, what, like eight days? You don't see this sort of traction in any company. And you don't see killers like this in any company. unless you're paying these motherfuckers like what they're worth bro like like like maxed out contracts maxed out contracts exactly exactly yeah uh what about devices i mean it seemed like this would be a natural fit for some sort of ai wearable or other platform um is there an app coming or are you interested in what's happening with johnny ive and open ai what's what's your take on the device world we're very interested in the hardware space we've got like a million things cooking on hardware we got people in the garage right now working on you don't you don't even know about bro like like like we're bringing manufacturing back to america and it all starts at the cluey garage let's go i'd love to see it nobody you know they doubted but you guys are re-industrializing america you guys really are the brain chips down there yeah brain chips brain chips that's the future there we go there we go the new neural link yeah I mean, I, you know, there's a world in the future where you guys actually just roll up Neuralink and OpenAI.

46:19For sure. Bring them under the Fluli umbrella. Yeah, definitely. It's possible. I'm excited to offer acquisitions for both of those companies. Yeah. It's in the roadmap. It's on the roadmap. All right. This has been a lot of fun. I'm excited for you guys. It is. And I have no doubt that you'll go from, you know, 10 million views a week, 10 million views a week to 100. And I'm excited to see you guys hit that billion view mark very soon. So keep it up. We are all very entertained and rooting for you. Shit, shit, shit, shit. I love the energy. Thanks, man. We appreciate you joining. Later, guys.

46:52Keep having fun. Bye. Next up, we have Kian from Nucleus coming on with a big announcement. Something like 10 years in the making, close to it, maybe seven years. We'll bring Kian in. Let's play some soundboard. How are you doing? Welcome to the show. That's a great intro. The tweets are flying. Oh, my God. You guys seen this? Yes. You seen this? Seeing this? Break it down for us. Explain what's happening. There's nothing like a launch day. I'm trying to figure out, guys, is this Gattaca or is it Theranos? Because people, they can't make up their mind. Oh, yeah. We're going to find out. They're trying to figure it out.

47:31They're trying to figure it out. What's going on? Let's give some context to the audience. Nucleus has launched Nucleus Embryo, the world's first genetic optimization software. Basically, parents can give their children the best start in life. They can pick their embryo based off of physical characteristics like eye color, IQ. They can go to disease risk like cancers or heart disease. Basically, we really believe parents can get all the information that exists about their embryos, and they can pick however they want. For me personally, you know, it's been 10 years in the making. The journalist actually covered it today in the Wall Street Journal.

48:01It was a journalist that covered my gene editing in a warehouse in Brooklyn 10 years ago. Yes. Wow. Overnight success. You know, it's a long time in genetics. Yeah, so break down the state of the art because like embryo screening exists. I think most parents in America, at least if they have the means, do some sort of screening while the embryo is growing. Is this purely for IVF? Is this just going a layer deeper? And then I want to talk about the regulatory and FDA component as well. Let's talk about it. So basically, if you go to an IVF clinic today, you're a couple. The vast, vast, vast majority of clinics.

48:38The first thing I should understand is that the IVF process is principally controlled today by clinicians or doctors. Honestly, couples don't have as much liberty in our perspective as they should. It's their baby. It's their embryos. They should have the right to that information, and they should have to pick off any vertical. However, today in the clinic, what generally happens is people test embryos for very rare and severe genetic conditions. For example, like chromosomal abnormality, like Down syndrome, for example, or even a condition like cystic fibrosis or Tay-Sachs or PKU. right these are conditions that are very rare um that maybe someone might have a carrier for cystic fibrosis but again it's it's pretty rare um then there are conditions that we've all heard about heard about things like breast cancer things like coronary artery disease the things that actually kill the vast majority of people today right chronic conditions kill the vast majority of people today those conditions are just not tested for in the clinic even though we have very good science actually that can make those predictions how do we know this as a as a dna company as well that's what we do right we build models that predicts disease and the way test those models in adults so we go from adults to embryos it's actually because we can basically well validate these models to show that they work in both the embryonic context and in the adult context and so what we're really doing is we're going from okay instead of just looking for really severe like down syndrome cystic fibrosis why not do breast cancer why not do heart disease why not do colorectal cancer right why not do schizophrenia why do parkinson's but then why stop there and this is really the important thing because ultimately you know if you think about diseases and traits the extreme version of any trait is actually a disease right height is a good example of this one extreme end is like you know John for example he's like mark then the other end is like me dwarfism right okay you know so you know IQ is like example this one end is like you know autism the other end it can actually be some sort of you know cognitive basically challenge that people have and so when you think about it what do you start realizing that people have drawn a line in the sand saying you can get you know rare diseases you can't get common diseases but then they really said you can't can't get any traits like height even though the best predictor we have today actually in the world the best polygenic predictor is for height so as a company we've kind of completely reimagined this and said wait a second what's going on here you should have access to the entire stack rare diseases we do cystic fibrosis common diseases like breast cancer and also traits all the way up to something like IQ.

50:51Yeah. So I mean, uh, that test, are you just giving people the data? Because I imagine that once you get into particular recommendations, that's more of what I would expect a licensed doctor to need to do. Yeah. My sense is that they, you can allow people to get the data from their doctor and then, and then feed it into nucleus. Is that correct? So that's, that is correct. And actually we have a couple, there was like 10 announcements today. You know how we do We like to do 10 announcements in one day. We are actually very, very excited to announce a huge partnership with Genomic Prediction. Genomic Prediction is actually the oldest embryo testing company that exists.

51:28They've done genome-wide tests in embryos for almost a decade at this point. And I think they've done over 120 ,000 couples for PGTA, which is a specific kind of test. And so we're actually partnering with them. So we make it very easy for Genomic Prediction customers to request their files and actually port it over to Nucleus. But really, this isn't just for Genomic Prediction customers. Anyone who's undergoing IVF can go to their clinic and say, I want my embryos data. You can take that data, you can upload it to Nucleus, and then all of a sudden, you know, the application of DNA makes this technology universally accessible.

52:00Now, how much of the benefit is actual algorithmic analysis bringing in other data points to contextualize the data versus just better UI and better hydration of existing text? because we had a friend on the show who was talking about getting some medical results from a doctor. The doctor's office was closed. It took two days until the doctor was going to be able to interpret the results. He was able to just take a photo, upload it to ChatGPT and say, hey, is this really, really bad? Should I be panicking? Because it seems somewhat out of the range. And ChatGPT was able to say, hey, you still got to talk to the doctor, but this isn't the craziest thing I've ever seen.

52:43This isn't way out of distribution. And so that's almost like a pure UI layer, but extremely valuable. I know it might not be like the right narrative for some people that it's like not as innovative. But I think that like all that matters at the end of the day is giving people benefits. It's always both. It's always both. Fundamentally, technology, just for technology's sake, it's not siliconized about. It's siliconized about making something that people want, okay, that people can actually use. Exactly. So if you think about the nucleus innovation, it's two-pronged, okay? One is in the informatics, right?

53:11You know, I've been doing this for five years. Sure. I almost, I would argue to myself that I probably spend too much time developing the science, right? Because science, in a nutshell, isn't actually very useful. You need to expand access to it. So on that point, we do multiple different kinds of analyses that make it such that we can actually provide the most comprehensive analyses that exist today. But moreover, and this is really, I think, a key point to your point, John, is people understand them. People can see them. I mean, you can pull up the platform. I'm not sure if you guys have shown it already, but it's very easy to sort, compare your embryos.

53:42You can actually name your embryos, you can stack rank your embryos, you can understand what the score means. We lead with overall risk or we tell you, for example, instead of saying you're in the 99 percent off for genetic risk for a condition, which, you know, what does it actually mean? We say, hey, you have a 5 percent chance or the like of, let's say, schizophrenia or some other condition. In other words, by leaving overall risk, people have a much greater intuitive understanding of the results we're communicating to them. We have genetic counselors on hand. So this really is a... What are we showing here?

54:04Are we showing something? Are we showing the... Yeah, yeah, yeah. We pulled up here. That's another thing you're showing, actually. That's another thing. That's a fun one. That's an Easter egg. That's an Easter egg. That's the kind of approach that we're taking here. And I think consumers are responding to it, right? People want to have access to their data. The clinician, the doctor shouldn't decide what embryo to implant. You should. Okay, so talk to me about what requires FDA approval. Obviously, new medical devices. Like if you were developing a machine to take in an embryo and sequence the DNA, I would expect that the FDA would want an approval for that medical device.

54:37But if you are taking data and just showing it to a customer in a different UI, that feels like probably a very light FDA process. And there's probably a continuum in the middle where once you're making a recommendation, they have rules around that, right? We as a company do not tell you which embryo to implant. Sure. You know, basically parents, the couple has complete agency to decide how they want to use the information to implant their embryo. Moreover, let's be clear. Height, right? I mean, can a height analysis be a medical device? It doesn't even make sense, right? IQ, height, there's traits, for example.

55:12We all, you know, traits are something that I don't think necessarily belongs in even the kind of infrastructure thing about medical care, right? These are things that go beyond medical care. These are things like, you know, that people just kind of intuitively know and that there are DNA tests done every single day due to see for these analyses because they're not disease analyses, right? So we do both diseases and traits, to be clear. My point is many of these innovations, you have to wonder, should the government say if someone can or cannot pick their embryo based off height? That doesn't seem right to me.

55:38I think it should be in the complete liberty of the individual to decide that. Yeah, but I mean, we're a democratic country. And so if a huge swath of the population says that the FDA should review that type of test or that type of analysis. Ice analysis? It could happen. I mean, the FDA reviews all sorts of different stuff. And so I guess the question shifts to like, do you expect a change from FDA on the way these analysis tools are regulated? I think right now the most important thing is just putting these high quality rigorous scientific results in people's hands and then helping them basically have healthier children helping them give their child the best start in life yeah um you know i i think that's generally speaking that you know people should have more liberty more choice in medicine i think the broader longevity trend actually touches on that point as well yeah um so that's what we're excited to do at nucleus yeah i mean the fact that you're partnering with a company on on actual on the actual like medical device side like they are doing the sequencing the embryos like that really takes it out of the Theranos question entirely in my mind.

56:48I think you feel like you should be beating the drum there a little bit more It's like like we didn't say we got some new device But we ship that's the difference I know it's like Go use it That's the evidence Okay I Love the visual of John and his wife selecting between embryos and it's like six ten or seven two tough choice well if we go with the 610 he has you know potentially fly commercial once in his life we can actually play this game right now okay here we're going to play a game right now I'm going to put in the chat pickyourembryo.com okay everyone listen to this pickyourembryo.com I'm going to go to it oh my god here we go little easter egg here okay let's see what's more important to you John intelligence or muscle strength come on oh absolutely muscle strength let's go we're the future is bodybuilding yeah john would john would take a he would he would happily have a five two son if he had you know top 0.01 percent bodybuilding exactly yeah okay so which one is your lifespan or height uh come on lifespan lifespan let's go let's go let's go maybe low depression you got to be golden retriever mode you got to be uh you need low depression you need low depression let's go low ocd i don't mind bouncing around a bunch okay what's more risk taking anxiety uh let's go high risk taking there we go okay okay we're analyzing this is some generative stuff going on this is great i got nadia too enduring athlete let's go physically strong cautious built to last yeah this is great is this driving a lot of uh a lot of attention a lot of downloads is going viral yet this seems like something that's designed to be charitable dropped it right technology brothers we got you the exclusive let's go let's go there you go out there you know they can pick your embryo people say what's it like maybe not doing ivf yeah no problem only only nine percent of people choose nadia okay well we're we're contrarian we like that here yeah that's fine it's great oh well well uh congratulations on the news congratulations on the launch um yeah the pace is wild last thing what's going on with uh have you seen these just blood billboards oh yeah they're all over la so so there is there's someone who's running a campaign right now justice for elizabeth holmes claiming that theranos was not the scam people think it was and there's a there's a documentary coming out and there's billboards all over la for just blood like it's just blood it's not that big of a deal and john just be clear there's an exclusive on technology where there's next week about from this person right they're going to tell their story next week just to make sure you invite them already.

59:28I hope we're we're we are we are toying with the idea that someone reached out to kind of connect us. We're thinking about doing it, but we're not we're not 100 percent sure that would be appropriate for the based on the website. I don't know if I don't know if it's appropriate. Yeah, it doesn't look like it was designed with Figma, so I don't know. We can't quite do it. It's a little bit. It's definitely doesn't use linear. Yeah, but they they claim that that Elizabeth Holmes has been proven innocent. And so it's a bold claim. we like we like to see people making bold claims by what uh jury is my question yeah the jury of someone who knows html kian uh always a great time energy is fantastic electric electric thank you for coming on firing us up congratulations on the launch we will talk to you soon talk to twitter for sure okay we'll see you there bye guys bye we have someone from open ai here We're going to stick to technology and business, but welcome to the show, Mark Chen.

1:00:24Good to see you. Good to see you guys. Awkward day, but I'm excited to talk about deep research. I am excited to talk about AI products. Would you mind introducing yourself and kind of explaining what you do because OpenAI is such a large company now and there's so many different organizations. I'd love to know how you interact with the product and the research side and anything else you can give to contextualize this conversation. Yeah, absolutely. So first off, thanks for having me on. You know, I'm Mark. I am the chief research officer at OpenAI. So in practice, what that means is I work with our chief scientist, Jakob, and we set the vision for the research org.

1:01:00We set the pace. We hold the research org accountable for execution. And ultimately, we really just want to deliver these capabilities to everyone. That's amazing. In terms of research, I feel like a lot of what happens in the research side is actually gated by compute. Is that a different team? Because what if the researchers ask for a$500 billion data center? That feels like maybe a bigger task. It is useful for us to factor the problem of research and also kind of building up the capacity to do that research. So we have a different team. Greg leads that, which really thinks holistically about data center bring up and how to get the most compute for us.

1:01:39And of course, when it comes to allocating that compute for research, Jakob and myself do that. That's great. And so what can you share that's top of mind right now on the research side? There's been this discussion of pre-training scaling wall, potentially the importance of reinforcement learning, reasoning. There's so many different areas to go into. What's actually driving the most conversations internally right now? Yeah, absolutely. So I think really it's a really exciting time to do research. I would say versus two or three years ago, I think people were trying to build this very big scaling machine.

1:02:21And really, the reasoning paradigm changed a lot of that, right? You know, like reasoning is really taking off and it really opens this new playing ground, right? It's like there are a lot of kind of known unknowns and also unknown unknowns that, you know, we're all trying to figure out. It kind of feels like GPT-2 era, right? Where there's so many different hyperparameters you're trying to figure out. And then I think also, like you mentioned, pre-training, that's not to be forgotten either. Today, we're in a very different regime of pre-training than we used to be. Today, we can't treat data as this infinite resource.

1:02:55Yeah. I think a lot of academic studies, you know, they've always kind of treated, you know, you have some kind of finite compute, but infinite data. I don't think there's much study of, you know, like, you know, finite data and infinite compute. And I think, you know, that also leads to a very rich playground for research. Do we need kind of a revision to the bitter lesson? Is that a refutation of the bitter lesson or do we just need to rethink what the definition of scaling laws looks like? No, I don't think of anything as a refutation of the bidder. Really, our company is grounded in we want simple ideas at scale.

1:03:35I think RL is an embodiment of that. I think pre-training is an embodiment of that. And really, at every single scale, we face some kind of difficulty of this form. It's just like you've got to find some innovation that gets you past the next bottleneck. And this doesn't feel fundamentally very different from that. Mm-hmm. What's most important right now on the actual compute side? We heard from NVIDIA earnings that we didn't get a ton of guidance on the shift from training to inference usage of NVIDIA GPUs, but it feels like it must be coming. It feels like this inference wave is happening. are those even the right buckets to be thinking about tracking metrics in terms of the story of artificial intelligence?

1:04:24Because, yeah, I mean, it's like if the reasoning tokens are inference tokens, but they're what lead to higher intelligent, more intelligent models, like it's almost back in the training bucket again. What bucket should we be thinking about? Or how firmly are we in the the, the, the, the applied AI era versus the research era? Well, I think research is here to stay. And it's for all the reasons I mentioned above, right. It's such a, like a rich time to be doing research, but I do think, you know, inference is going to be increasingly important as well, right. It's such a core part of RL that you're doing rollouts.

1:05:05And I think, you know, we see 2025 as this year of agents, right. We think of it as a year where models are going to do a lot more autonomous work. You can let them kind of be unsupervised for much longer periods of time. And that is just going to put big demands on inference. When you think about kind of our overall vision, we lay it out as a series of steps and levels on the way to AGI. And I think the pinnacle, really that last level, is organizational AI. You can imagine a bunch of AIs all interacting. And yeah, I think that's just going to put huge demands on inference. On that organizational question, I remember reading AI 2027, and one of the things that they proposed was that the AIs would actually literally be talking to each other in Slack.

1:05:55Does that seem like the way you imagine agents playing out, using the same tools as humans instead of one agent says i'm gonna go talk with teams and talk with slack i'm gonna do a little negotiating but maybe it just happens super super fast 24 7 or or is there like a new machine language that emerges yeah um i mean i think one thing that's really helped us so far in ai development is uh to come in with some priors for um you know how humans do things and that's actually um you know if you bake those priors and they typically are great starting points so I could imagine maybe you start with something that's Slack-like and give it enough flexibility that it can develop beyond that and really figure out the way that's most effective for it to communicate.

1:06:44One important thing, though, is we want interpretability too. I think it's very helpful for us today that what the agents do is easy for us to read and interpret. I don't think you want that to go away as well. I think there's a lot of benefits just even from a pure debug the whole system perspective, or just let the models speak in a way that is familiar with us. And you can also imagine we might want to plug in to the system too, right? So whatever interfaces we're familiar with, we would ideally like our model to be familiar with as well. I think it's also pretty compatible with, we hit a big milestone.

1:07:25We got, I think, 3 million paying

1:07:38business users for a while. I was hoping you would drop a number. Yeah. Congratulations. That's actually huge. That's amazing. But I think one big part of that is, you know, we have connectors now, right? Yeah. We're kind of coming into, you know, like G drives. And I think, yeah, you can imagine, you know, like Slack integrations, things like that. I think we just want the models to be familiar with the ways we communicate and get information. Yeah. Can you talk about benchmarking? It feels like we're potentially. Yeah. Do you think about benchmarks at all? Oh, yeah, a lot. I mean, but I think it's a difficult time for benchmarks, right?

1:08:20I think we used to be in this world where you have these human written benchmarks for other humans. And I think we all have these norms for what are good benchmarks. We've all taken the SAT. We all have a good conception of what it means to get whatever score on that. But I think the problem is the models are already at the point where for even the hardest human written benchmarks for other humans, it's really near saturated or saturated. I think one clear example here is the Amy, probably the hardest auto-gradable human math eval, at least in the US. And yeah, the models are consistently getting like 90 plus percent on these.

1:09:03And so what that means is I think there's kind of two different things that people are doing. They're developing kind of model-based benchmarks. They're not kind of things that we would give to an ordinary human. Things like humanities last exam, things like, you know, Epic AI that are really, really at the frontier of what people can do. And I think the hard thing is it's not grounded in intuition, right? Like, you know, you don't have a lot of people who have taken these exams. So it makes it harder to kind of calibrate on whether this is a good exam or not. One of the exciting things that's on the flip side of that is I really do think we're at the era where models are going to start innovating, right?

1:09:48Because I think once you've passed the last kind of like the hardest human rating exams, that's kind of at the edge of innovation. And I think you already see that with the models, right? Like they're helping to write parts of papers. Um, and, and I think the other kind of way that, uh, people have shifted is, you know, there's these, you know, ultra frontier evals, but they're also people kind of just indexing on real world impact, right? You look at your revenue, kind of the value you deliver to users. Um, and I think that's ultimately what we care about. Can you, can you, uh, bring that back to interpretability research, like with these super, super hard, uh, math evals, for example, uh, if, are we doing the right research to understand if the thought process mirrors, not just, not just one shotting the answer, Oh, you, you, you, you memorized it or you magically got it correct, but you actually took the correct path.

1:10:43Kind of like, you know, you're graded for your work, not just the answer if you're in grade school. Um, and, and, you know, Dario said that, uh, interpret, interpret, interpretability research will actually contribute to capabilities and even give a decisive lead. Do you agree with that? What's your reaction to that concept of interpretability research being very important? Yeah, I mean, we care a lot about it here at OpenAI as well. So one thing that we care a lot about is interpreting how the model reasons, right? Because I think we've had a very kind of specific and strong view on this in that we don't want to apply optimization pressure to how the model thinks so that it can be faithful in the way it thinks and to expose that to us, you know, without any kind of incentives to cater to what the user wants, right?

1:11:31I think it's actually very important to have that unfiltered view because, you know, oftentimes, like if the model isn't sure, you don't want to hide that fact, right? Just for it to kind of please the user. And sometimes it really isn't sure, right? And so we've really done a lot of work to try to promote this norm of chain of thought faithfulness and interpretability. And I think it gives you a lot of sense into what the model's thinking and, you know, what are the pitfalls that it can go off into if it's not reasoning correctly. That's such an important point, because if you have somebody on your team and they come to you and they say, hey, you know, I think this is the right answer, but we should probably verify it.

1:12:13It's like, it's still valuable. Totally. It puts you on the right path. If somebody comes to you 100 % confidence, this is the truth and something wrong. It's like, well, like trust is just destroyed. Yeah, totally. Don't you guys feel like, you know, safety felt a lot more theoretical a couple years back, right? But like today, you know, like the things that people are talking about a couple years, like scalable oversight, really having the model be able to tell you like and convince you that the work it did was right. It feels so much more relevant right now. Just because the capabilities are so strong.

1:12:42Yeah. I mean, just personally, I've completely flipped from being like, uh, oh, the safety research is not that valuable, uh, because I'm not that worried about getting paperclips. It just seems like a very low likelihood that that's kind of like the bad ending, like immediately in this foom and all this crazy, uh, the gray goose scenarios were just so abstract and sci-fi. It just felt like economics will, will, will, will, will fall into place and there will be, uh, like a, like a cold, like a nuclear ending, which is like, we didn't build nuclear plants and we just stopped everything because we humans seem to be good at that.

1:13:13But now that we're actually seeing things. Yeah, it's crazy how fast it's been. Right. Like, oh, yeah, I think my, my, my personal story is, it's like, you know, what, what got me into, uh, AI was AlphaGo, right? Like just watching it get to that level of capability. Yeah. And you were kind of like, it was such an optimistic and also kind of a little bit of a sobering message, right? When you saw Lisa, it'll get beat. Um, and I just remember, you know, like we, we saw the coding models, you know, when we first launched like, I think very OG codex, you know, with GitHub Copilot, it was maybe like under, you know, a thousand ELO on Code Forces.

1:13:48And I still remember the meeting where I walked into where the team showed my score and they're like, hey, the model's better than you. You come full circle and it's like, wow, like I put decades of my life into this and, you know, the capabilities are there. So like, if, you know, I'm kind of at the top of my field in this thing and it's better than me. Like what can it do? Yeah. Yeah. That's amazing. Uh, I have so many more questions on alpha go. Are there, uh, are there lessons from scaling how scaling played out there that you can, that we can abstract abstract into the rest of AI research?

1:14:23What I mean is, uh, as I remember it, the alpha go training run was not a hundred K H two hundreds. Uh, but what would happen if we actually did an AlphaGo style training run? I mean, it would be an economic money pit, right? Like they've had no economic value to do, but let's just say some benevolent trillionaire decides I'm going to spend a billion dollars on a training run to beat AlphaGo and go even bigger. Is Go at some point solved? Would we see kind of diminishing scaling curves? Could we throw extra R out? Could we port back everything that we've doing in just general AGI research and just continue fighting it out in the world of Go?

1:15:08Or does that end and does that teach us anything? Yeah, yeah. Honestly, I feel like if you really are curious about these mysteries, join our team. That's the thing I want to say. So yeah, I mean, really like kind of the central problem of today is RL scaling, right? When you look at AlphaGo, right, it's a narrow domain, right? And I think in some sense that limits the amount of compute you can pump into it. But even kind of small toy domains, they can teach you a lot about how you scale RL. Like what are the axes where it's most productive to pump scale in? I think a lot of scaling research just looks like that, whether it's on RL or pre-training.

1:15:44So you identify a lot of, you know, different variables under which you can scale. And like where is kind of where you get the best kind of like marginal impact for pumping scale there. I think that's a very open question for RL right now. And I think what you mentioned as well, it's just like, you know, going from narrow to broad, right? Does that give you a lever to pump a lot more scale in as well? I think when you look at our reasoning models today, they're a lot more broad based than, you know, just being able to kind of an expert system on Go. So yeah, I really do think that there are so many levers to scale.

1:16:21What about Move 37? That was such an iconic moment in that AlphaGo LisaDoll match. They placed Move 37. It's very unconventional. Everyone thinks it's a blunder. It turns out not to be. It turns out to be critical. It turns out to be innovation. Do you think we are, we're certainly post-Turing test in language models. We're probably post-Turing test in image generation, but it feels like we're pre-Move 37 in text generation in the sense that there hasn't been like a fully AI generated book that everyone is just, oh, it's the new Harry Potter. Everyone has to read it. It's amazing. And it's fully generated or this image.

1:17:03the images they do go viral but they go viral because they're ai move 37 in the context of go did not go viral because it was ai it felt like it was actual innovation so uh is that the right frame does that make any sense um i think it's not the wrong frame so i think some some quick thoughts on on that um i i think kind of um when you have something that's you know very measurable like win or lose, right? Something like go. Yeah, it's like very easy for us to kind of just judge, right? Like did the model do something right here? And I think the more fuzzy you get, you know, it is just harder, right?

1:17:41Like when it comes to, is this the next Harry Potter, right? Like, you know, it's not a universally loved book, I think fairly universal, but you know, there's some haters. And yeah, I think it is just kind of hard when it comes to these human subjective things where it's really hard to put down in words like what makes you like Harry Potter, right? And so I think those are always going to lag a little bit, but I think we're developing more and more techniques to attack kind of these more open-ended domains. And I don't know, I wouldn't say that we're not at an innovative stage today. So I think my My biggest touch with this was when we had the models compete on the IY last year.

1:18:26So I like the International Olympics for computer science. Basically the top four kids from each country go and compete. And these are really, really tough problems. Basically selected so that they require some innovative insight to solve. Right. I think and we did see the model come up with solutions, even to some very ad hoc problems. And and so I think there was a lot of surprise for me there. Right. I was completely off base about which problems the model would be able to solve the most. Right. I think like I kind of categorize their six problems, some of them as more kind of like, oh, this standard, a little bit more standard.

1:19:13This is a little bit more out of the box. I was like it's not going to be able to solve this more out of the box one, but it did. And I think I really does speak to kind of these models have the capacity to do so, especially trained with our own. Now, now, now put that in context of what's going on with Arc AGI. Obviously, OpenAI has made incredible progress there, but it just when I do the problems, it seems easy. And when I look at the IOI sample problems, I think this would be a 20 year process for me to figure out how to achieve that. And I can do the ArcGIS on my phone. Is this the spiky intelligence concept?

1:19:50Is this something that a small tweak in algorithmic design, just one shots ArcGIS or, or is there something else going on there that we should be aware of? Yeah. I mean, I think part of this is the beauty of ArcGIS as well, right? Like I think I'm not sure if there's another human intuitive simpler benchmark, which is for the models. I think really that's one of the things they really optimize for on that benchmark. I do think when it comes to models though, there's just a little bit of a perception gap as well. Models aren't used to this native, just screen type input. I think there's a lot we can bridge there.

1:20:30Actually, even O4 Mini, it's a state-of-the-art multimodal model in many ways, including visual reasoning. And I think, you know, you're starting to kind of build up the capacity for the models to take images, manipulate and reason about them, generate new images, write code on images. And I think it's just been kind of under-focused. But I think when I talk to researchers in the field, they all see this as a part of intelligence, too. And we're going to continue to focus there. Yeah. Is is is RKGI kind of in the if we're dropping a buzzword on it is like program synthesis. Is there a world where I know that I know the tokens like the images, we see them as renderings of squares and different colors.

1:21:15but uh the when they're fed into the llm they're typically uh just a stream of of numbers effectively is there a world where actually adding a screenshot is what's important like visual reasoning yeah yeah so i think i think that could be important it's just like kind of uh you know whenever it comes to like textual representation of grids um models today just don't really do that well right And I think it's just kind of because humans don't really ever write down textural representations of groups. We have a chessboard, like no one really kind of just like types it out in a grid. And so the models are kind of like under trained a little bit on on what that looks like and what that means.

1:22:00So, you know, I think with more reasoning, we'll just bridge the gap. I think with better visual perception, we'll just bridge that gap. Yeah. How are you thinking about the role of non-lab researchers in the ecosystem today? I'm sure you try to recruit some of the best ones, but the ones that don't join your team. Tell us about the one that got away. Yeah, the one that got away. Yeah, no, I mean, I think it's still actually a fairly good time for specific domains, right, to be doing research. And, you know, I think the style is just very different. And you do feel the pull of non-lab researchers into labs because I think they feel like a lot of the burning problems in the field are at scale, right?

1:22:45That's great, yeah. And that's kind of one of the unfortunate things too, right? Like when you look at reasoning, you just don't see that happen at small scale, right? There's like a certain scale at which it starts becoming signal bearing. And that requires you to have resources, right? But I do think, you know, a lot of the really good work that I've seen, you know, there's experimental architectures. I think a lot of good work is happening in the academic world there, like a lot of study in optimization, um, a lot of study in kind of like GANS, you know, um, there's certain fields where you see a lot of fruitful research that, that happens in academia.

1:23:23Yeah. That makes a lot of sense. How about, uh, consumer agents? How are you thinking about them? Uh, you talked earlier about sort of B2B adoption and that's all very exciting, but how much do you and the research org think about breakout consumer agent products? Yeah, that's a fantastic question. I think, um, We think about it a lot. I think that's the short answer. We really do think this year we're trying to focus on how we can move to the agentic world. And when I think about consumer agents, I think ChatGPT proved that people got it. People get conversational models. But when it comes to consumer agents, we have a couple of theses that we've tried out in the world.

1:24:09I think one is deep research, right? I think this is something that can do five to 30 minutes of work autonomously, come back to you, and really kind of synthesizes information. It goes out there, gathers, collects, and kind of compresses the information in a form that's useful. A little bit of pushback there. Like I can see that as a consumer product when someone like Aiden is like, I want new towels. And he uses deep research to like figure out like what is the best towel across every dimension. But when I think of deep research, yes, it has applications with students, but it's often. Some of them might just be the paradigm because I feel like consumers being like, give me a deep research report on this country and where to travel and things like that.

1:24:53We keep using this flight example, but I haven't actually tried to book a flight with deep research. It's totally possible that it could go and pull all the different flight routes and calculate all the different Delays and all the different all the different parameters of if I fly to this airport and park or I can use Valet here or something like that. Yeah. Yeah And I guess like when I think of agents, it's it's deep research is like, you know curating Information in which you can take action on but it's like at what point is action a part of that sort of loop Right where you can not only curate a list of flights that you want but then actually go out and have agency.

1:25:31I think one of our explorations in that space is operator, right? It's where you kind of just feed in raw pixels from your laptop into or from some virtual machine into the model. And it produces either a click or some keyboard actions, right? And so there it's taking action. And I think the trouble is, you know, it you don't ever want to mess up when you're taking action. Yeah. I think the cost of that is super high. You only have to get it wrong once to lose trust in a user. And so we want to make sure that that feels super robust before we get to the point where we're like, hey, look, here's a tool.

1:26:13That's so different than deep research because you can wind up on some news article and read one sentence that it gets a fact wrong or the commas in the wrong place and the numbers off. And but that's just the expectation for just text and analysis. And if you delegated that, yeah, you're going to expect a few errors here and there. Oh, that's actually a different company name or that's the that's an old data point. There's new data, but very different if I book a flight and book the wrong flight and I can wind up in Chicago instead of New York. Exactly. And I think the reason why we care so much about reasoning is because I think that's the path that we get reliable agents through.

1:26:49Sure. You know, we've talked about like reasoning, helping safety, but reasoning is also helping reliability. Right. It's like you imagine like what makes a model so good at a math problem? It's like it's banging its head against it. It's trying a different approach and then it's like adapting based on what it failed at last time. And I think that's the same kind of behavior you want your your agents to have. It's like, yeah, things like adapts and keeps going into it. And that's that's the humans do this every day. You're booking a flight. You keep hitting an error. It's not which or which form you missed, right?

1:27:21And you're just sort of banging your head against the computer and eventually it says, okay, you're booked, right? So I think that's a great call out. Yeah. I mean, there's so many more questions we could go into, but I'm interested in the scaling of RL and kind of the balancing act between pre-training RL and inference, just the amount of energy that goes into getting a result when you distribute it over the entire user base. How is that changing? And I guess, are we post like really big, really big runs? Is this going to be something that's like continually happening online? Or it feels like we're moving away from the era of like, oh, some big development, some big run happened, and now we're grouping the fruits of it versus a more iterative process.

1:28:12I mean, I don't see why it has to be so right. I think if you find the right levers, you can really pump a lot of compute into RL as well as pre-training. I think it is a delicate balance, though, between all of these different parts of the machine. And when I look at my role with Jakob, it's just kind of like figure out how this balance should be allocated, where the promising nuggets are arising from and resourcing those. Yeah, it's kind of, in some sense, I feel like part of my job as a portfolio manager. Yeah. That's a lot of fun. Well, thank you so much for joining. This is a fantastic conversation.

1:28:49We'd love to have you back and go deeper. Great hanging, Mark. We'll talk to you soon. Yeah, peace. Have a good one. Next up, we have Shalto Douglas from Anthropic coming on this show. I'm getting so many. Jordy is giving us the update on that. I'm just getting a lot of messages saying, no one cares about AI. Talk about the drama on the timeline. Well, we do care about AI. We care a lot about AI. But it is a mess out there. Wow. Yeah, the end of the Trump-Elon era. I don't know. Maybe we have to get some people on to talk about it tomorrow or something. I'm going to do it today. Anyway, we have Shalto from Anthropic in the studio.

1:29:30How are you doing? What's going on? Good to see you guys. Hopefully, you're staying out of the chaos on the topic. Don't open the time. Don't open. We'll do you a favor. Sweet child. Just everything. Stay focused on the application. Stay focused on the next training run. We really humanity really cannot afford for any researchers to open X. What a hilarious day. Anyway, I mean, yeah, how are you doing? What is new in your world? What are you focused on mostly day to day? And maybe it's just a way of an intro. Yeah. So at the moment, focused really on scaling RL. And that is the theme of what's happening this year.

1:30:12And we're still seeing these huge gains where you go, you know, 10x compute increase in RL. We're still getting like very distinct linear gains on the basis of that. And because RL wasn't really scaled anywhere close to how much pre-training was scaled at the end of last year. We have like basically a gamut of like riches over the course of this year. So where are we in that RL scaling story? Because I remember some of the rough numbers around like GPT-2, GPT-3, we were getting up into like, it cost$100 million. It's going to cost a billion dollars. Like just rough order of magnitude, not even from Anthropik, just generally.

1:30:48Like what is a big RL run cost or how many are we talking 10K H200s or 100K? Like are we going to throw the same resources at it? And if so, how soon? Yeah. So I think in Dyer's essay at the beginning of the year, he said that a lot of runs were only like a million dollars back in like December. Thinking of like DeepSeq v3 and this kind of stuff like R1, which means that with us like at least two OOMs just to get to the scale of GPT-4. And GPT-4 was two years ago. Yeah. Right. RL is also perhaps a bit more naively paralyzable and scalable than pre-training. You know, pre-training, you need everything in one big data center, ideally, or you need like some clever tricks.

1:31:25RL, you could like in theory, like what the Prime Intellect folks are doing, scale it all over the world. out of it. And so you're held back far less than you are in Retreat. Sure. So everyone and their mother has a billion dollars now. There are hundreds of thousands of GPUs getting pumped all over the place. I feel like we're not GPU poor as a society. Maybe some companies need to justify it in different ways, but it sounds like there's some sort of like reward hacking problem that we're working through in terms of scaling RL. What are all of the problems that we're working through to actually go deploy the capital cannon at this problem?

1:32:07Yes. So, I mean, think about what you're asking the model to do in RL is you're asking it to achieve some goal at at any cost, basically. Yeah. And this comes with a whole host of like behaviors, which you may not intend. In software engineering, this is really easy. I'd like to it might try and hack unit tests or whatever. But in much more longer horizon, real world tasks, you might ask it to say, go make money on the internet. And it might come up with all kinds of fun and interesting ways to do that, unless you find ways to guide it into following the principles that you want it to obey, basically.

1:32:38Or to align it with your idea of what's sort of best for humanity. And so it's actually, it's a pretty intensive process. It's a lot of work to find out and hunt down all the ways these models are hacking through the rewards and patch all of that. And yeah. Yeah. How are we going to see scaling in the number of rewards that we're RLing against, if that makes sense? I would imagine that at a certain point, we unless we come up with like kind of like the the the Genesis prompt, go forth and brief fruitful or something and multiply the you could imagine training runs on just knocking down one one problem after another.

1:33:22And is that is that kind of the path that we're going down? I very much think so. There's this idea in which, like, you know, the sort of world becomes an RL environment machine in some respect, because there's just so much leverage to making these models better and better at all the things we care about. And so I think we're going to be training on on just everything in the world. Got it. And then does that lead to more model fragmentation, models that are good at programming versus writing versus poetry versus image generation? Or does this all feed back into one model? Does the idea of the consumer needing to pick a model disappear?

1:33:59Are we in a temporary period for that paradigm? I think the main reason that we've seen that so far is because people are trying to make the best of the capital. We are all still GPU poor in many ways. And people are focusing those GPUs on the sort of like spectrum of wars that they think is most important. And I'm a bit of a big model guy. I really do think that similar to how we saw with large pre-trained models before, where small fine-tuned models had gains over the GPT-2 era, but then were obsoleted by GPT-4 being generally good at everything. I think, to be honest, you're going to see this generalization and learning across all kinds of things that means you benefit from having large single models rather than specialization or area fine-tuned models.

1:34:46Can you talk a little bit about the transition from or any differences between RLHF and just other RL paradigms? Yes. So RLHF, you're trying to maximize a pretty like lossy signal. Things like pairwise, like what do humans prefer? And I don't know if you've ever tried to do this, like judge two language model responses. I get prompted for that all the time. Right. And I'm always like, I don't want to read both of those. I'll just click the one on the left. Exactly. Exactly. And I click one of the random ones sometimes. Yeah. Or I click like the one that just looks bigger or I'll read the first two sentences.

1:35:21But yeah, I'm not giving straight. I'm not doing my job as a human reinforcer. Exactly. Human preferences are easy to hack. Yeah, totally. Environments in the world are much truer if you can find them. So something like, did you get your math question right? Is a very real and a true reward. Does the code compile, right? Does the code compile, exactly. Did you make a scientific discovery? We've got very little rewards right now, but pretty quickly over the next year or two, you're going to start to see much more meaningful and long horizon rewards. You're going to see models bribing the Nobel committee to win Nobel prize.

1:35:57There's reward hacking. But that's the other thing we want to prevent, right? Exactly. Yeah. Yeah. That's the real nightmare scenario. What about, like, there's so many different problems that we run into that feel like it's just really, really hard to design any type of eval. My kind of benchmark that I use whenever a new model drops is just tell me a joke. They're always bad. Or even the latest VO3 video that went viral was somebody said like, stand up comedy joke. And it was kind of a funny joke, but it was literally the top result for joke Reddit on Google. And then it clearly just took that joke and then instantiated it in a video that looked amazing.

1:36:44Um, but it wasn't original in any way. And so, uh, we were joking about like the RLHF loop for that is like, you have an endless cycle of comedians running AI generated materials and then, and then, you know, uh, speak, uh, microphones and all the comedy clubs to feedback what's getting laughs. I mean, honestly, that would work pretty well. Yeah. If any comedians want to hook us up with an RL loop, I mean. Yeah, yeah. But I mean, for some of those less, like, as you go down the curve, it feels like each one gets harder and harder to actually tighten the loop. We see this with, like, longevity research, where it's like, okay, it takes 100 years to know if you extended a human life.

1:37:25Like, yes, you could create a feedback loop around that, but every change is going to be hundreds of years. And so even if you're on the cycle, it's irrelevant for us in the context that we talk about AI. So talk to me about like, are you running into those problems or will there be like another approach that kind of works around those? So there are a lot of situations where you can get around this by just running much faster than real time. Like let's say the process of building like a giant app, like building Twitter, right? It's something that would take human months. But if you got fast enough and good enough AIs, you could do that in several hours.

1:37:56Sure. Like parallelize heaps of AI agents. They're all building things. And so you can get a faster reward signal in that way. In domains that are less well specified, like humor, I agree, it's really, really hard. And this is like why I think in some respects, like creativity is like at the top end of the spectrum, like true creativity is much, much harder to replicate than the sort of like analytical scientific style reasoning. And that will just take more time. You know what? The models actually are pretty good at making jokes about being an AI. This feels weirdly fresh. Like everything else is kind of a weird copy of something like it, like it just, it feels like it's derivative.

1:38:31Basically it's trying to infer what humor is and it doesn't really understand it, but jokes about being an AI are quite funny. Yeah. I, I, I think this also might be, I don't know if it was directly reward hacking, but I noticed that, uh, one of the new models dropped and a bunch of people were posting these like 4chan, like be me memes. And, and, and they were, it seemed like they were kind of hacking the humor by being hyper-specific about an individual that they could find information on online. And so you're laughing at the fact that it's like, oh, wow, that is like something that I've posted about.

1:39:00It's making a reference, but it's not really that funny to me. It's other than it's just like, wow, they really did its research. Like it really knows Tyler Cowen intimately, which is cool, but I didn't find it hilarious. Yeah, yeah, yeah. Very interesting. Let's talk about some sort of deep research product projects and products. Um, we were talking to will brown and he was saying like AGI is here with some of the bigger models, but the, but the time that AGI can feel consistent, it diverges. And so you could be working with someone who's, you know, a hundred IQ, but the, but they will stay consistent for years as an employee, or they'll, they'll keep, you know, living their life.

1:39:44Whereas a lot of these super smart models are working really well. And then after a few minutes of work, the agents kind of diverge and kind of go into odd paradigms. And it feels very not human. It feels like they're hyper intelligent one way and then extremely stupid in others. What's going on there? What is the path to extending that? Is that more like having more better planning and better, uh, better like dividing up the task or, or will this just kind of naturally happen through the RL and scale? Yeah. So there's that jaggedness, right? Which is what you're seeing is how we call it. And I think that is largely a consequence of the fact that maybe something like deep research, it's probably been RL to be really good at producing a report.

1:40:26Yeah. But it's never been RL on the like act of producing valuable information for a company over a week or a month, or like making sure the stock price goes up in like a quarter or something like this. Right. Like it doesn't have any conception of how that feeds into the broader story at play. It can kind of infer because it's got a bit of world knowledge from the base model and this kind of stuff. There's never actually been trained to do that in the same way humans have. So to extend that, you need to put them in much longer running, much like your long horizon things. And so deep research needs to become, you know, like deep operate a company for a week kind of thing.

1:41:00Sure. Is that the right path? Like it feels like the road might be there's a like the longest running LLM query used to be just like a few seconds, maybe a few minutes. And I remember when when some of the reasoning models came out people were almost trying to like stunt on it by saying like oh I asked it a hard question. I thought for five minutes now deep research is doing 20 minutes pretty much every time Is the path two hours two days or are we gonna see more? efficiency gains such that we just get the 20 minute model, the 20 minute results in two minutes and then two seconds. Yeah. So this is somewhere where like inference in many respects and prioritization becomes really important.

1:41:41So both like how fast is your inference? It literally affects the speed at which you can think and the speed at which you can like do these experiments. Also, how easily you can paralyze becomes really important. Like can you dispatch a team of sub agents to go and do deep research and like compile like sub reports for you so that you can do everything in parallel? These kinds of like, it's both like there's an infrastructure question here that feeds up from the hardware and the chips and this kind of stuff to like designing better chips for better inference and all this. And an RL question of like, how well can you parallelize and all this.

1:42:16So I think we just need to compress the timelines, compress the timeframes basically. Yeah. So if I'm, if I'm like an extremely big model and I'm running an agentic process, like how, how much am I hankering for like a middle sized model on a chip or like baked down into silicon that just runs super fast? Because it feels like that's probably coming. We saw that with the Bitcoin project progression from CPU to GPU to FPGA to ASIC. Do you think we're, we're at a good enough point where we can even be discussing that? because every time I see the latest mid-journey, I'm like, this is good enough.

1:42:53I just want it in two seconds instead of 20. But then a new model comes out. I'm like, oh, I'm glad I didn't get stuck on that path. But yeah, how far away are we from, okay, it's actually good enough to bake down into silicon? Well, there's a question here of baking it down to silicon versus designing a chip, which is very suited for the architecture that you care about, right? And baking it down to silicon, unsure. I think that's a bet you could take. but it's a risky one because the pace of progress is just so fast nowadays. And I really only expect it to accelerate. But designing things that make a lot of sense for the transformers or architectures of the future should make a lot of sense.

1:43:35That's a big gap, though. Transformers or architectures of the future. If we diverge, there's a lot of companies that are banking on the transformers sticking around. What is your view on transformer architecture sticking around for the next couple of years? I mean, look, they stuck around for five years, so they might stick around for a little while. But there's different, you think about architectures in terms of this balance of memory bandwidth and flops, right? One of the big differences we've seen here is Gemini recently had actually a diffusion model that they released. I was about to ask you.

1:44:02The other day, right? So diffusion is inherently extremely flops intensive process, whereas normal language model decoding is extremely memory bandwidth intensive. You're designing two very different chips depending on which bet you think makes sense. And if you think you can make something that does flops like four times faster than diffusion and like four times cheaper than your other is good, diffusion makes more sense. So there's like this is dance basically between the chip providers and the architecture, both trying to build for each other, but also like build for the next paradigm. Yeah, it's risky.

1:44:29Do you I don't know how much you how much you've played with image generation, but do you have any idea of what's going on with images and chat GPT? It feels like there's some diffusion in there. There's some tokenization maybe some transformer stuff in there It almost feels like the text is so good that there's like an extra layer on top Almost and that it's it's almost like reinventing Photoshop and and I guess the the broader question is like it feels like an ensemble of models Maybe the discussion around around just Agents and text based LLM interactions shouldn't necessarily be transformer versus diffusion, but maybe how will these play together?

1:45:07Is that a reasonable path to go down? Well, I think pretty clearly there's some kind of like rich information channel between even if there are multiple models there, there's like, it's, it's conditioning somehow on the, on the other model, because we've seen before, like, let's say when, uh, you know, models use mid journey to produce images, it's never quite perfect. It can't perfectly replicate what went in as an input. It can't perfectly like adjust things. Uh, so there's a link somehow, whether that's the same model producing tokens plus diffusion. I don't know. Um, like, yeah, can't comment on what open I was doing there.

1:45:38Yeah. Yeah. Um, are, are there any other kind of like super wild card long shot, uh, research efforts that are maybe happening even, even in academia where, I mean, this was the big thing with, uh, what was his name? Gary, uh, he was talking about, I forget what it was called. Symbol manipulation was a big one. And I feel like, you know, you can never count anyone out because it might come from behind and be relevant in some ways. But are there any other research areas that you think are like purely in the theory domain right now that are worth looking into or tracking that, you know, low probability, but high upside if they work.

1:46:21Hmm. That's how fun. This is a tough one. But we'll say it's not a symbolic thing. Please. It's crazy how similar transformers are to systems that manipulate symbols. Sure. Like what they're doing is they're taking a symbol and they're like converting it into a vector and then they're manipulating and moving stuff like information around across them. Sure. Like this, this whole like a debate that all transforms cannot represent symbols and they cannot do this. It's it's yeah, it's not real. So Gary Mark is underrated or overrated? I guess. Yeah, yeah. But, but I mean, if you, if you twist the, if you twist it so much, you wind up with saying like, well, really like the, the transformer fits within that paradigm.

1:47:04And so maybe it's, you know, it's, you know, it like the rhetoric around it being a different path was maybe false the whole time. something like that. But as I remember that debate, it was really the idea of compute scaling versus almost like feature engineering scaling and will the progress scale with human hours or GPUs essentially. And that has a very different economic equation. And it feels like there's been there's been some rumblings about maybe with a data wall we'll shift back to being human labor bound but do you think that there's any chance that that's relevant in the future or is it just algorithmic progress married with bigger and bigger data centers in the future so i'm pretty bit a lesson built uh in the sense that i do think removing as many of our biases and our like clever ideas from the models is really important just like freeing them up to learn now obviously there's like there is clever structure that we put into these models such that they're able to learn in this extremely general way.

1:48:10And, uh, but I am more convinced that we will be compute bound then we will be like human researcher, uh, out human research, our bound on, on this kind of thing. Like we're not going to be feature, feature engineering and this kind of stuff. Sure. We're going to be trying to devise incredibly flexible learning systems. Yeah, that makes sense. Um, On the scaling topic, part of my worry is that the ooms get so big that they turn into these mega projects that at a certain point you're bound by the laws of physics because you have to move the sand into silicon chips and you have to dig up the silicon.

1:48:50And it is so bad. Yeah, there's only so much sand. And like the math gets really, really crazy just for the amount of energy required to to move everything around to make the the big thing. Where are you on on how much scale we need to reach AGI? How whether or not we will see like the laws of physics start acting as a drag on progress, because it certainly feels exponential. We're feeling the exponentials, but a lot of these turn into sigmoids, right? So I think we've got, what, like two or three more OOMs before it gets really hard. Leopold has this nice table at the end of his situational awareness.

1:49:29I think like 2028 or something is when under really aggressive timelines that you get to 20 % of U.S. energy production. It's pretty hard to go exponentially beyond 20 % of U.S. energy production. Now, I think that's enough. Every indication I'm seeing says that's enough. Now, there might be some complex data engineering, raw engineering, this kind of stuff that goes into, there's still a lot of algorithmic progress left to go. But I think that with those extra OOMs, we get to basically a model that is capable of assisting us in doing research in software engineering. Yeah, which is the beginning of the self reinforcement.

1:50:09Interesting. Is that just a coincidence? Like this feels like one of those things, this feels like one of those things where like the moon is the exact same size as the sun in the sky. It's like, oh, it just happens that AGI happens within this time. Have you unpacked that anymore? Because it feels convenient. Not to, you know, I know less about this. There's a lot of weird conveniences or like weird. It's a good sci-fi story, let's say. Totally. We've got Taiwan in between China and the US and it produces the most valuable material in the world. It's locked between the two. Incredible plot. Yeah.

1:50:39Incredible plot. Yeah. Really bad for the people that don't believe in simulation theory. It really feels It feels like this is scripted. It's fascinating. Talk to me more about getting to an ML engineer in AI and kind of that reinforcement. I imagine that you're using AI code gen tools today and Anthropic is broadly and everyone is. But what are you looking for and what are the what's the shape of the spiky intelligence? Where do they fall flat and what are you looking to kind of knock down in the interim before you get something that's just like go yeah so i mean we definitely use them the other night i like i was a bit tired of us to do something just sat watching it in front of me working for half an hour it was great it was truly a weird experience particularly when you look back a year ago and we're still copy pasting stuff between a chat window and you know a code file yeah um uh what i like meters evals for this kind of stuff so they have a bunch of evals where they measure like the ability to write a kernel the ability to run a small experiment and improve a loss and they have these nice progress curves versus humans and I think this is maybe the most accurate reflection of like what it will take for it to really help us at doing progress and there's a mix here like where they're not so great at the moment is like large-scale distributed systems engineering right like debugging stuff across heaps and heaps of accelerators and like the way the feedback loops are slow and you actually like if your feedback loop is like an hour then it's you spending the time on doing something.

1:52:11Yep. Feedback is 15 minutes. And for context there, the hour long feedback loop is just because you have to actually compile and run the code across everything. You need to spin up all your machines or you need to run it for a while to see if something's going to happen. At that point in time, you're still cheaper than the chips. So it's better that you do it. But for things like your kernel engineering or for like, you know, actually even just understanding these systems incredibly helpful. Like one thing I regularly do at the moment is in parts of the code base in like languages that I'm unfamiliar with or stuff like this, I'll just ask it to rewrite the entire file, but with comments on every line, game changing.

1:52:50Wow. It's like - Comments on every line. Yeah, or just hunt through like thousands of files and explain how everything interacts to me, draw diagrams, this kind of stuff. It's really, yeah. Yeah, how important is a bigger context window in that example you gave That feels like something that's important. And yet I just naively like Google's the one that has the million token context window. I imagine that all the other frontier labs could catch up, but it seems like it hasn't been as much of a priority as maybe like the PR around it sounds like. Is that important? Should we be driving that up to like a trillion token window?

1:53:24Is that just going to happen naturally? There's a nice plot in the Gemini 1.5 paper where they show the loss over tokens as a function of context length and they show that the loss goes down quite steeply actually as you put more and more and more like of a code base in the context you get better and better and better predicting the rest yeah that makes sense the context length it's a cost um you know the way transformers work is that uh there's you know you have like this this memory that is proportional the kb cache is proportional to how much context you've got and so you can only fit so many of those into like your various chips and this kind of stuff uh and so longer context actually just costs more because you're taking up more of the chip and you're sort of like you could have otherwise been doing other requests basically.

1:54:05So bringing it back to the custom silicon is that a unique advantage of the TPU is is that something that Google has has thought about and then wound up to put themselves in this advantaged position or is it a durable advantage even? Yeah so TPUs are good in many respects partially because you can connect hundreds or thousands of them really easily across really great networking whereas only recently has that been true for GPUs. With NVLink? Yeah, with NVLink and the NVL72 stuff. Okay. So it used to be eight GPUs in a pod, and then you connect them over a worse interconnect. And now you can do 72, and then it breaks down.

1:54:40With Google TPUs, you can do 4 ,000, 8 ,000 over a really high bandwidth interconnect in one pod. And so that is helpful for things like just general scaling in many respects. I think it's doable across any chip platform, but it is an example of somewhere that being fully vertically integrated is a is easy to benefit yeah that makes sense uh talk to me about arc agi why is it so hard it seems so easy it does seem easy doesn't it uh that's it well it certainly seems like more more evaluatable than tell me a funny joke right yeah yeah and i mean i think if you are old on arc agi then it would you probably get superhuman at it pretty fast um but i think we're all trying not to rl on it so that it functions as like an interesting held out.

1:55:24Sure. Okay. Is that just an informal agreement between all the labs? Yeah. We try and have a sense of honor between us. That's good. Sense of honor. That's amazing. How many people on earth do you think are getting the full potential out of the publicly available bottles? Because we're now at a point where we have, you know, billion plus people are using AI almost daily. And yet I have to admit, my sense would be it's maybe like 10 ,000, 20 ,000 people on the entire planet are getting that sort of full potential, but I'm curious what your assessment would be. Yeah, I completely agree. I mean, I think that even I don't get the full potential out of these models often.

1:56:03And I think as we shift from you're asking questions and it's giving you sensible answers to you're asking you to go do things for you that might take hours at a time and you can really like paralyze and spin, that we're going to hit like yet another inflection point where even less people are like really effectively using these things because this basically doesn't require you to like, it's like StarCraft or Dota. Like it's going to be like your APM of like managing all these agents and that's going to be a process. Yeah. I think StarCraft is such a good example. You think you're just absolutely crushing it and then you realize like there's an entire area of the map you're just getting destroyed on.

1:56:40It's such a good comp. That's great. Anything else, Jordy? I think that's it on my side. I mean, I would like this to be an evolving conversation. Yeah, this is fantastic. We'd love to have you back and keep chatting. Absolutely. It was really fun. Love to go back on with us. Yeah, we'll talk to you soon. Cheers, Shelta. Have a good one.