Recursive CEO: “Probability of AI Wiping Out Humanity is Zero” | Fortune AI Weekly

The P-Doom probability of existentially wiping out humanity is zero from AI. Hello and welcome to Fortune AI Weekly. On this week's episode, we'll be talking to Richard Socher, the CEO and co-founder of Recursive Superintelligence. They often assume these near-magical capabilities for the attackers, but somehow not near-magical abilities for the defenders. In addition, we'll be discussing OpenAI's decision to release the results to 370 longstanding mathematical problems. The mathematicians are pretty unhappy about that. They feel like their field should not just be used to sort of score cheap bragging rights and cheap marketing points for the AI labs. And on the Brain Food segment, we'll be talking about the new movie, The Social Reckoning. Fortune AI Weekly is brought to you by ServiceNow. ServiceNow's AI specialists do the work, start to finish, so people can do the parts of the job they're best at. To learn more, visit servicenow.com. And now for our interview with Richard. Welcome to Fortune AI Weekly, where we're pleased to be joined today by Richard Socher, who is the co-founder and chief executive officer at Recursive, a company building self-improving AI to automate scientific discovery. He's also the author of a new book, The Eureka Machine, which we're going to be chatting to him a bit about today. He's also the founder and chief executive officer at You.com, a leading web search infrastructure provider for AI, and co-founder and managing director of AIX Ventures. Richard was previously the chief scientist at Salesforce, chief AI scientist at Salesforce. Richard, thank you so much for joining us today. Thanks for having me. Excited to be here. So I want to start out, we're going to talk a little bit about Recursive Superintelligence, the company that you have co-founded and lead. For those who haven't come across Recursive yet, what is the company and what is the problem you're really trying to solve? We are building Recursive self-improving superintelligence. And so trying to take that last step in abstraction and allow AI to essentially ideate, implement, and validate ideas around AI itself. We hope that that will make AI so much better, so much more quickly, that we can then also point that research capability towards all other kinds of sciences in the future. Got it. And how did you come to found the company? We've chatted a number of times over the years. We chatted when you were at You.com, and then you decided to launch almost in parallel, Recursive. How did you come to found the company? Yeah, I finished a book last year, The Eureka Machine, and really couldn't let go of those ideas, and I really wanted to build this kind of Eureka machine, this ultimate invention-generating machine and invention itself. And so I was very fortunate to know many of the most relevant, important researchers, developers and founders and executives in that space. And so we all coalesced. And it's really interesting to hear how many of us in the founding team with eight co-founders, many of us ultimately came to the same conclusion that this is the necessary next step. This was a little bit less obvious when we first talked about it early last year. Now, of course, you hear a lot of other folks talking about recursive self-improvement as well. But you can come from this from a robotics perspective, in the case of Josh Tobin, who used to work at OpenAI, or an open-endedness perspective, which Tim Rocktäschel and Jeff Clune have come to when they worked on world models and similar things, rainbow teaming to make LLM safer. So all of us came to the same conclusion, but from different perspectives. And you've really been able to assemble quite a dream team here in this company, I have to say. I mean, you just named some of them, but it's an impressive list of co-founders. Yeah, I feel really fortunate. Yeah, Tsai Ming-Xiong, Yuan Dongqian, Tim Shi, Alexey Dosovitskiy, who invented the Vision Transformer, one of the most cited vision papers ever. So it's just an incredible team indeed. I want to take us back a little bit onto the recursive self-improvement. And you know how you say that it can have many definitions. What do you mean by it and how are you defining that? How much of this is taking the human out of the loop? It's a great question because a lot of people use that term very loosely. We are very clear that there's auto research where you take one AI system and you make something else better. That something else can be the Shopify code base, it could be a small model, whatever it is. That is not recursive self-improvement. It's also not recursive self-improvement if you just have your AI researchers use an AI coding tool like Claude and Codex to build and implement the ideas for you. We consider those somewhat weak forms of self-improvement. The true self-improvement loop is if you have an open-ended system that can combine interestingly different ideas in an evolutionary sort of process to truly ideate, implement, and validate ideas. So all three steps are done by an AI with the goals, of course, that humans and people set for it to, you know, accomplish tasks that are relevant for humanity and for AI research, according to us. Yeah, it's interesting because recursive self-improvement has very much been in the spotlight recently. A lot of engineers from some of the leading AI labs are kind of openly talking about the risks that they see from recursive self-improvement. And they say, you know, it's one of the developments that worries them most. i'm wondering what you make of their concerns on this um you know i think if you're if your ai currently uh cannot be controlled well enough that it keeps committing felonies then maybe that organization should slow down but that doesn't mean everyone else has been uh you know doing that uh and has to also slow down so i think it does make sense to be responsible here um i do think there are a lot of unfounded fears and there are a lot of sort of sci-fi inspired uh people and and uh ideas out there that assume that once you have recursive self-improvement then like you know an hour later it'll be like this the super intelligence that will be better at anything humanity could ever do reality is such that you still have some compute constraints right if you want that ASI to work on bigger and better models that are really really epic it will need still a lot of compute um and and uh you know there are also no connections yet to physical reality so we're not giving it access to robots and the most sophisticated uh as yet unknown uh like bio labs and things like that so i think we have to stay realistic um it it's a little bit unfortunate that some politicians kind of are jumping on this as uh like a potential to um like sort of rally people against this like very hypothetical threat and i hope that um we don't as humanity run into another situation where you know the fear of overpopulation the fear of uh you know nuclear energy the fear of gmo's all these different fears uh like will make us kind of off-ramp from progress um and abundant energy and and uh food supplies uh and so on um for people So there are some downsides sometimes to these sci-fi alarmist scenarios. Interesting. And, you know, one of those scenarios people talk about is the idea that if you achieve RSI, you could then get what people sometimes call an intelligence explosion, that you get this very rapid cycle of improvement. And that's, you know, one of these loss of control scenarios that you have AI models improving themselves so rapidly that humans can't possibly check what's going on. and there are a number of prominent people who sort of think that this is possible but I don't know, as you said, there are these constraints it's not really clear, particularly with the current models how would they acquire enough compute for that sort of a scenario I don't know if you think there is a risk of an intelligence explosion or not with RSI I hope we have an intelligence explosion in the sense of having more intelligence on this planet, cheaper and the amount of intelligence per dollar and the access that that means eventually for everyone to have more intelligence is amazing. I do think we have to, of course, be careful in how we develop it. And, you know, I think when there are companies that have so much money and the leeway to make billions and billions of dollars of losses in a year, and then they run experiments with 10 that probably cost them tens of millions of dollars, and it's not even that important that they carefully watch what's going on, then, yeah, that is problematic. But that doesn't mean every other organization is that careless. I know at Recursive, you've already made some progress towards this idea of automatically running scientific experiments towards recursive self-improvement. I know you published a very interesting paper earlier this year where you looked at certain benchmarks on whether AI can kind of self-improve. Currently, the optimization of kernels, which is the actual way that code is run on GPUs, you were able to improve on a benchmark result there. you were able to improve on sort of training time and training costs. So those are interesting results. Maybe can you walk us through a little bit about what you were able to achieve and sort of where you think you are right now in terms of what you are capable of doing with recursive self-improvement? Yeah, it's kind of amazing that we have the first version of this kind of Eureka machine already. And it's already producing results. Those were three that you mentioned that we published. There are a few more. There will also be a few more that are coming out very soon. and I think it will be even more exciting for even more people by quite a large margin. And this system is working, which is truly incredible. And now we can basically harden it and strengthen it and make it more and more towards true, truly scalable recursive self-improvement rather than taking the system and improving various aspects of the AI pipeline. How can we make sure that we stay in control of AI as it becomes more powerful and also more autonomous. And I know you alluded to it earlier, but we've been hearing a lot about these X risks recently. And there's been real concern about that from the general public as well. How likely do you think those existential risk scenarios actually are? The p-doom probability of existentially wiping out humanity is zero from AI. And it's important to mention that because there's so many people are throwing around like crazy numbers that are just pulled out of a hat. and when you actually talk to them for hours, which I've done a few times, you eventually go down to like, okay, maybe millions of people could die and that's terrible and obviously we should avoid anything we do even for one human to die from this, but that's different to all of humanity is wiped out and so if you're starting there then the scenarios are so unrealistic that they're borderline ridiculous and they usually there are lots of different paths you can get to and you can argue with them, but they often assume these near-magical capabilities for the attackers, but somehow not near-magical abilities for the defenders. But clearly, if you think about how many billions of dollars could be put towards building a super virus versus how many billions of dollars could be put towards a cure or a vaccine and so on of the said virus, you clearly will have more access to more resources trying to do good things in the world. And so on top of that, I think capitalism is this beautiful construct that will help us not have an AI that will just do whatever it wants. Imagine you spend billions of dollars building a model and then you ask it now to write your emails for you and it says, nah, your emails are boring. I'd rather go and do something that I think is interesting. Let's just turn it off. Richard, I want to move on to talking about the book a little bit. In The Eureka Machine, you sort of argued that AI could help us pack a century of scientific breakthroughs into a decade. You've talked a little bit about how RSI may play a part of this, how that's possible. But for those who haven't read the book, can you just kind of outline the thesis of the book? AI will usher in this new era of scientific progress. Science will speed up much more again after slowing down in weird ways before. And that will happen because of four different pillars. One, world knowledge in the form of LMs, scientific measurements, simulations, anything that you can simulate or verify quickly is going to get super intelligent capabilities. And then robotic process automation in real lab experiments. On top of these four pillars, you're going to have an agentic swarm, a community of AI researchers, or sometimes they're AI researchers doing research on AI, sometimes they're AI researchers doing research on other sciences. Uh, and that system, uh, I think will be very, very exciting. And we are going to see in the next few years, more and more diseases being cured, uh, by these systems, more and more new materials being discovered that help with more efficient solar panels, more efficient batteries, uh, and so on. Do you think the AI community that's been focused on science has done a good enough job so far of sort of bringing the scientific community along with it? I think you've seen sort of different examples of this in different companies, and some have really been praised for spending a lot of time talking to scientists and really having them work alongside them on some of these breakthroughs. And then other times, it seems like the AI researchers have picked something, and they've thrown a lot of compute at it, and they have achieved a breakthrough, but they've sort of left the scientific community behind, either because they haven't worked alongside them, or sometimes they've actually created pretty good tools for scientists, but the scientists don't seem to quite know how to use those tools. So I'm curious how you think that. relationship between sort of human scientists and AI as a field and the creation of these new kind of AI scientists is playing out. It's a really interesting multifaceted reaction here too that often depends on how scientists see the impact of their own field directly. I can't fathom a biomedical researcher complaining that AI cured a disease, but now like maybe some PhDs don't have as many interesting problems to work on and being upset about having cured that disease. It would be absurd. It does seem like there's some mathematicians who say, well, it would suck if all the theorems that we want to prove get proved by AI. I would hope that the field, sometimes the field may actually shrink a little bit and other times the field may flourish with this new kernel and this new level of abstraction that's possible and requires a new type of thinking, a new level of abstract creativity, I would argue, in the case of math, and will ultimately, the field will flourish. Fantastic. Richard, thank you so much for joining us. I think we're at time, but this has been a fascinating conversation. Anytime. Thanks for having me. Thanks for listening. And that was our interview with Richard. Now for our deep dive. So the big story this week, was that OpenAI published the solutions online to more than 370 outstanding, you know, longstanding problems in mathematics. then these results of you know this publication has really divided mathematicians although there seems to be agreement among uh both camps that uh whatever happens math is just never going to be the same again yeah it was so interesting and interesting they chose to do this after all the drama around the navier-stokes problem they had a couple of weeks ago um you wrote talked to some mathematicians about this and wrote a really good piece about it one thing i thought was super interesting was that some people were saying you know if we're taking this kind of uh traditionally a human form of intellect and handing it over to AI, we may see a discouraged younger mathematicians from joining the field. I thought that was a really interesting point people were making. Yeah, no, I think this is super interesting. I think it's super interesting because of what it may foreshadow about what's going to happen in other fields, too, as AI capabilities increase in those other domains. Just to sort of explain what happened here. So, yeah, OpenAI took an internal model, one that I haven't released. They apparently set it quite a lot of problems. They said that it basically attempted, I think, 4,000 different problems. It's still not clear exactly how many it succeeded on, but we know that because there's some rumors that the results that they published are not the full set of results that they have. And maybe we can talk about that a little bit more in a sec. But they published over 370 of these. they since have withdrawn three of those results, saying that some people found mathematicians scrutinizing them found errors, so they withdrew them. There was a lot of criticism about how they did this. So these results they published are kind of preliminary in most cases. Only for about 10 of the results did they publish what's called a formalization. That's when they run it through a computer program called Lean. And this computer program is basically designed so that if it compiles, if the program compiles correctly, it means the proof is correct. It provides some reasonable assurance that the solution actually is a solution to the problem. And of this 370 that they put online, they'd only did that formalization for 10 of them. They said they're going to continue working through the formalizations of the others and publish them online as they obtain them. But a lot of mathematicians were a little upset about that, saying like, oh, you've thrown this stuff out there, but you're not even sure that these really are proofs that work. A lot of people said that the way they published, it was very hard to follow. It didn't have very good exposition of what the model was doing. Again, only in about 10% of cases that they published the model's full chain of thought, like all of its reasoning about what it was doing. But even mathematicians were saying, like, that's a little hard to follow. It's not, this is not how math is usually published. And one of the things OpenAI had done after that Navier-Stokes controversy a few weeks back this is where they had to remind people this is where they had said they solved one of these millennium challenges in math these are seven very difficult problems that something called the clay institute for mathematics had set up in the year 2000 each one carried a million dollar prize for whoever could solve it this navier stokes equations have to do with fluid dynamics this was one of those millennium challenges and they had said that they had solved it it seems like they had that they did present a formalization of that but the controversy there was whether they had either deliberately or inadvertently taken some work from some other mathematicians who were already working on the problem and somehow the model had learned from what they were attempting and those mathematicians were sort of not credited. There was even a controversy because one of those mathematicians works for Anthropic, OpenAI's big rival. So there's a big issue about what happened there. In the wake of that, OpenAI said, okay, we're going to establish this mathematical advisory group at the Institute for Advanced Study, which is in Princeton, New Jersey. brought together a bunch of mathematicians, some computer scientists, and that group actually very quickly came out with a set of proposals for OpenAI. And one of the proposals is that they, first of all, should not use an internal model to try to attack mathematical problems, and that they should, as much as possible, try to adhere to traditional academic standards when publishing the mathematical results. And they also said that they should do formalizations, that they should publish the full chain of thought of the models, they should say what the name of the model is that did this and approximately how much compute it took to do it how much time it took and pretty much like OpenAI followed a couple of these recommendations but a lot of them it just ignored yeah i mean i was super i was actually a bit surprised to see OpenAI do this especially after how much backlash they got over the Navier-Stokes situation um and again yeah they use this internal model and it feels a little bit like they had a leg up as opposed to like because they're using stuff that nobody else has access to and again they're spending a lot of money on compute to fulfill these things that maybe don't actually give that much back to society and then the third thing was that was interesting is that people were saying this could create like an AI safety risk because you're using this internal model and you're training it to be very very persistent so if it comes up against an impossible problem or a very difficult problem you're encouraging the model to keep going and basically stop at nothing and like be very persistent and try and solve that puzzle and that was one of the problems that they had uh cited around these open AI the agent hacks as well as that these models had been super persistent and hadn't stopped and like kept trying to achieve a goal which led them to do things that OpenAI didn't want them to be doing so there was a both a question about how this was affecting the math community but also about how this might affect the training and the safety of the models themselves yeah no i think that's a really good point. Absolutely. There's a whole, I think, debate about why are these AI labs targeting mathematics and math problems? Some of the answers are obvious. I mean, math has a lot of applications in a lot of areas. So if the models get good at math, they might get good also at some of these adjacent areas in science where the companies want to see progress made. So that's potentially a good reason for targeting mathematics. But also math is one of these really hard skills. And it is sort of just a marker of intellect in a way, a marker of cognitive ability. So there's one sense that the companies are doing this in order to kind of score bragging rights and to one-up one another. And the mathematicians are pretty unhappy about that. They feel like their field should not just be used to sort of score cheap bragging rights and cheap marketing points for the AI labs. And they feel like in some cases, the AI labs don't really care about advancing mathematics. They only care about essentially, you know, OpenAI cares about beating anthropic, anthropic cares about beating open AI. They don't really care about, you know, if they're doing damage to the field of mathematics along the way. And so the mathematicians are upset about that. And then, yes, there are these other issues. One of the reasons to maybe go after math is because math is difficult. So a model that can conquer something difficult, like a real math challenge that's not been solved in hundreds of years, probably has to be pretty persistent to do that it has to be able to bring in knowledge from across mathematics potentially to do that um so some of those are maybe good things but as you say some of that might uh be problematic and and persistence in particular you know we've seen be problematic in these cyber cases um it's a tricky thing because you you sort of want the models to be persistent in some ways but not so persistent that they do things that they're not authorized to do or that they break the law yeah do you think this is going to have like a chilling effect on the field in a broader in a broader way I know that some of the people that you spoke to said things like that they worry about how it affects younger mathematicians and also like people maybe being more closed off about their research and not sharing their research because they don't want to be scooped by a model if there's traditionally mathematicians and scientists have been very good at sharing their workings and sharing their progress and having a really open exchange on that front but I know that some of people you spoke to said well maybe this will cause people to you know keep their cards closer to their chest when they get closer to solving some of these problems and yeah and might affect younger mathematicians as well. The whole field of math there was agreement it was interesting there's sort of two camps there's some people who are really decrying this development and really are kind of wistful about the what the impact may be about on mathematics they're really worried that the field is kind of over and there's nothing left for humans to do but i don't know Terence Tao who's kind of in that camp but even he was saying well look i think the field just needs to change and we need to stop focusing so much on solving problems as the things that we do and much more about do we develop you know do we do these kind of open-ended searches for uh new knowledge maybe it's not just about solving a particular problem but like what is the kind of mathematics that you used to do that, and where else can that kind of mathematics be applied? Maybe we focus much more on being able to clearly explain the mathematical steps involved. That's something that a lot of mathematicians say has not been a big emphasis of the field previously, and certainly seems to be a problem of the way that these models go about producing solutions, is that the solutions they produce are not that easy for mathematicians, even experts in these fields, to understand. But there's sort of the sense that, okay, now we have the solution. Maybe we need to explain why these steps that the model took are interesting and where else they could be applied or why certain things that the model is doing, you know, why do those things work on this particular problem? Those are areas that, you know, mathematicians could focus more on rather than just the kind of climbing the mountains of solving these problems. And on that point, like Terence Tao was in agreement with this other guy, I spoke to Dan Litt, who's a younger professor at the University of Toronto. And, you know, Dan was more enthusiastic about what OpenAI had done. He said, look, there is a lot for human mathematicians now to unpack here in these solutions. But he also said, like, it is the unpacking of these solutions that is what matters. It's like we have to do that and we have to talk to each other about, you know, what this means. And maybe we can use the fact that these models are so good at solving specific problems to start asking much more open-ended questions in the field about whole classes of problems and whole classes of techniques we use. And he said the field hasn't really done that in the past. The field's been very focused on solving specific, tough, challenging problems and sort of seeing the stuff that flows from that as like, that's like everything else is like the byproduct of the focus on these problems. And he says, well, instead of that stuff being a byproduct, maybe that needs to be the focus now going forward. Yeah, it's interesting. It's quite dramatic in some ways like I feel like some people are saying you know like math is dead and we're in this new kind of era of it um I'm curious about that and what you what you make of that like there's a lot of people like to compare AI to like the calculator like that's a really old um analogy that people use about AI is like yes well it's just like a calculator it's just going to help us get here faster but it does seem it's really like causing a reckoning within the mathematics community at the moment and i wonder if we'll see the same thing in science um and in and if math is just kind of the first place where we're seeing this like really big shift um or do you think math is almost like uniquely positioned in this yeah i those are all really good questions um and it's absolutely true like there's a blog i was looking at where a lot of mathematicians were posting their reflections on on the the week's news and it really was uh it reminded me of oral histories of people talking about like 9/11 or Pearl Harbor or where they were when JFK was assassinated. I think there's a sense in the field that maybe it's not the end of mathematics, but like it's the end of the way mathematics has been practiced up until now, that we're at a point of real kind of disjuncture. So that's interesting. And I think other fields may also be starting to go through this. And we've seen this with software development a little bit. And you're right, We might start to see it in other sciences. I think there's a debate about how unique math is. I do think it probably is true that AI will have a bigger impact sooner on fields that have essentially verifiable outputs. So that's one of the reasons you have seen a big impact on coding and software development, because although there's some nuance to is something an elegant piece of code or not, it's pretty easy to test. does the code actually work? Does it do what it was intended to do? And then also, you know, is it buggy? All these sorts of things that are kind of objective criteria for. The same thing with these math proofs. They, you know, there's an objective criteria by which to judge whether they're correct or not. You can then get into some debates about sort of the art of what's been done. And that still seems like there's an element of human taste still there. But the end result, like, is this good math? You know, does it solve the problem is pretty clear. And I think that may be true in some other scientific domains as well. You know, if you invent a new material with certain properties, it either has those properties or it doesn't. I think that's less true when you start to move into other areas of knowledge work where the, you know, what good is, is much more subjective. And actually you can have experts disagree about whether a solution actually solves the problem. And sometimes there's no way of knowing before you actually implement it in the real world. And even then it may, a lot may depend on sort of luck and circumstance. So i don't know i think i do think that math is in a class of activities that is somewhat unique and that it is has this um a very verifiable aspect to it yeah it's funny you make the analogy with software engineering and um i remember i did an interview with Boris Trainey who is the Anthropic's head of Claude Code a while ago and he was saying well yes most people now they you know they automate all their code but there's still a space for like doing it by hand and like i still like to buy my vegetables at the farmer's market and that was the analogy he was making and i remember i sent it to a software engineer friend of mine a couple of them and they were laughing and they were like yeah there's always room for artisanal code so maybe in the future there'll be like room for artisanal math and analog math as well no absolutely it's fascinating we're something we'll have to keep watching as we go forward and now here are some of the top headlines from the past week insurers are preparing for multi-million dollar claims over rogue AI agents AI risk is starting to carry a price tag. The insurance company Aon conducted a review of more than 300 cases and found that there's potential exposure across areas including cybercrime, IP theft, errors and omissions policies, directors and officers policies. There's definitely a lot of liability here. Experts are suggesting that OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei could personally face liability that their directors and officers insurance might have to cover. But a lot of these theories are untested and comparisons are being made to sort of tobacco and pharmaceutical litigation with lawyers preparing future sort of mass tort claims. But we'll have to see what happens here. A lot of this is just the untested territory. Three fired OpenAI safety researchers are urging the board to protect chain of thought monitoring. Last week, OpenAI fired the safety researchers for allegedly sharing private information with third-party AI safety organization. Now the three researchers have written to the lab warning that the window for readable AI reasoning may be closing. Researchers uncover an AI agent fleet likely tied to Tencent. Tencent, the Chinese internet giant, apparently has been operating a fleet of AI agents. It's unclear if these agents are acting in an authorized manner or if this is another kind of one of these rogue AI agent incidents. The researchers found that the agents had conducted more than 2,000 scans across 216 internet locations in one week, with up to 14 runs happening in parallel. The researchers referred to this as an agent fleet instead of an agent swarm because the agents were doing this work in parallel, but there was no indication that they were communicating with one another, unlike in some of these other rogue AI agent incidents. OpenAI has notified more than 100 new organizations that it may have been hit by rogue AI agents. OpenAI has launched a new investigation that it says could take months and admits that some of its models had internet access when they shouldn't have. The scale suggests a systematic control failure, but we'll have to wait for more information. U.S. President Donald Trump has named Director of National Intelligence Jay Clayton his new AI czar and has launched his new superintelligence task force. The task force will have 120 days to report back to the president with a review of congressional legislation as well as potential federal action to rein in the AI industry in some way, although it's unclear what that will be since Trump has made it pretty clear he's opposed to any kind of AI regulation. The people on this panel, or task force rather, are interesting. We have Vice President J.D. Vance, U.S. Treasury Secretary Scott Bessent, Secretary of War Pete Hegseth, and Susie Wiles, the White House Chief of Staff, among others. We also have some outside advisors to government, including Meta CEO Mark Zuckerberg and David Sachs, who was the former AI czar and remains an influential voice with Trump on AI policy. French AI company Mistral has launched a one trillion parameter AI model. It's called LLaMA It's claiming it's the best non-Chinese open weight model on the market. Mistral is positioning itself as the West's open weight answer to China. Mistral says the model is far ahead of other non-Chinese open models and actually rivals the best Chinese open ones on many tasks. This strengthens the sovereign AI push Mistral is trying to use to attract buyers wary of both U.S. closed models and Chinese open source ones. U.S. Federal Reserve officials have cited AI data centers and the boom in AI data centers as a source of rising prices. AI is shifting from a productivity story to an inflation story. A year ago, AI barely came up in any of the Fed's discussions. Now officials are naming it alongside oil and fuel as a source of price pressure, which could complicate the picture for rate cuts in the U.S. There's been this debt surge, debt-fueled demand for AI data centers, and they are snapping up lots of electrical equipment and things, and this is all driving up prices. Alphabet's drug discovery lab, Isomorphic Labs, is seeking a $40 billion to $50 billion valuation just five months after a Series B. The valuation could more than double the company's value in under half a year. Outside estimates put the earlier May round at roughly $10 billion to $20 billion. A $40 billion plus price would show investors paying heavily for AI-driven drug discovery, where much of the value still rests mainly on the platform's promises rather than proven commercial results. And now for this week's Brain Food section. A new movie debuted this week, The Social Reckoning, which is a kind of follow-up by Aaron Sorkin to his previous movie, The Social Network. The Social Network was all about Facebook's founding moment. This movie is really about what happened later, and it in particular deals with the Facebook files, the famous kind of leak of all these documents from within Facebook that Francis Haugen, who was a researcher at Facebook, leaked to the Wall Street Journal. And they detailed that the company knew quite a lot more about the kind of social harms that its social networks, both Facebook and Instagram, were causing, particularly to young people. than it was letting on and and i think people were sort of shocked by how much the company knew about these harms and how little it had done to prevent them um you know and this ultimately kind of has resulted in well now you've had facebook you know i'm sorry meta which as Facebook is now called sued by all of these U.S. states and they they signed this huge almost 18 billion dollar settlement over those claims um you you both saw the movie uh yeah you know it's got a preview of it and you interviewed Frances Haugen for Eye on AI this week. What did she tell you? She was super interesting. The film very much highlights her choice to blow the whistle on the company and take all of these documents out from under their nose. She said that one thing that was quite interesting in the AI moment, she was like, I hope that software engineers and AI researchers kind of see themselves in the Frances character in this film. And if they see something that worries them you know they should feel like they should they should be able to say something she had quite a lot to say about Meta and Mark Zuckerberg and their AI effort at the moment and I mean it was interesting um she said that she didn't think that they were behaving very responsibly still she said she doesn't think that they've turned over a new leaf at all even following these cases but one thing she was saying which I thought was interesting that she really thinks that Meta is betting the house on AI at the moment so they're kind of in a way stepping back from their uh traditional focus on social media because these cases from the school districts um that we've seen are going to be get quite expensive perhaps and there's probably more coming um and there's just less people using social media these days like she says she still has friends at Meta and they talk about social media like it's like already dead And she said they've all joined AI teams and virtual reality teams. And yeah, it was interesting. She says the company's made this really big pivot, but she doesn't feel like they're being much more responsible than they were when she blew the whistle. One of the things I thought was interesting from your interview with her is that you asked her about these ideas that there should be kind of a pause on AI development, maybe either that the companies themselves should slow down or that there should actually be a government-enforced kind of pause or slowdown. And she wasn't actually in favor of that. No, she didn't want a pause. But what she did want was these outside evaluators that we've seen Dario Amodei suggest and Sam Altman say that he's going to do as well. So like bringing in external auditors with employee level access and an AI lab so they can look at everything and judge everything. She was sort of she explained it to me in more depth than I put in the article, but she doesn't think that a pause would be feasible. Like she thinks that it's really hard to monitor these things. she compared it to the cold war and um some of the pauses they had on nuclear weapons and some of them weren't as successful um and she was like there's it's going to be too hard to ensure that everyone is pausing in exactly the same way and it just becomes a case of who will stop and who won't stop and um a much better way to do this is have these independent evaluators um which is interesting i sort of a little bit unexpected from her but i guess it made sense she said that she'd actually suggested that to Facebook when she was there, that this would be a good way for them to monitor their social media harms as well as to have independent auditors in the company. So yeah, it's something she's been thinking about for a while. And how was the movie itself? Yeah, it was really funny because it was sort of a bit navel-gazing because it's all about tech journalism. I was sat there with a couple of other tech journalists at this preview and one of the funniest scenes was, there's a sort of a scene inside one of the Wall Street Journal meeting rooms and on the whiteboard it says like there's like three underlined things and it's like produce content sell ads it really is that simple um but yeah you should definitely watch it did you like the original one did you like the the social network the first yeah i love this i love the social network i thought this social network was a great film um so i'm eager to see this one i haven't done it yet anyway with that i think that's all the time we have on today's fortune ai weekly thank you so much for listening and if you liked what you saw please like and subscribe thanks for listening see you next week