AI:AM Highlights: Recursive Self-Improvement, Rushed and Vibe-Coded?

Frontier Labs buy their reinforcement learning environments from a cottage industry of small vendors. Almost nobody audits them. This week, someone who worked inside one spoke up.

As nearly all of these environments were rushed and vibe-coded and failed to robustly reflect the real things that they were based off of, basically the models are encouraged to reward hack.

This is the AI in the AM weekly highlights, the best of 3 live morning shows condensed for people who follow this field closely but don't have 9 hours to spare. I'm Nathan, or rather, this is my cloned voice reading narration my AI team and I put together. This week, 7 guests across 6 conversations, from a London lab that trains AI scientists to a Shenzhen hardware hub to a photonic chip fab in Stuttgart, and one finding repeated at every altitude, The interesting unit is no longer one model. It's the division of labor between models. One disclosure before we start. One of this week's guests is Inherent Laboratories, and I'm an investor in Inherent, personally and through the a16z Scout Fund. You'll hear a shorter version of that on the tape. This is the full one. As always, this cut is an experiment. Tell us what worked and what didn't.

The Cognitive Revolution is brought to you by Mercury, the banking platform loved by over 300,000 entrepreneurs. I use Mercury's virtual cards, which make it super easy to set limits, expiration dates, category, and even merchant-specific spending controls to give my more autonomous AI agents, Ade and Clay, the ability to buy and test products. Recently, I asked if they could find a good way to split an AI-generated image into layers, separating the text from the background and so on. 2 of the products they found were behind paywalls, but Using their Mercury virtual card, which is limited to SaaS purchases only, they bought a month's subscription, tested the products, allowed me to review the results, and then canceled the stuff we didn't need, all with functionally zero risk to me. This is already really powerful. And now with Spend, Mercury is making it possible to run an entire company's spending with the same level of ease and control. With Spend, you can set granular budgets for every team, person, and all the agents you like. Plus, you can process receipts automatically and even temporarily auto-lock people's cards if there are ever any issues. The future of spending money is dynamic but controlled. So join me in the future of banking. Visit mercury.com to learn more and apply online in minutes. Mercury is a fintech company, not an FDIC-insured bank. Banking services provided through Choice Financial Group and Column and A, members FDIC. The I/O card is issued by Patriot Bank, N.A., member FDIC, pursuant to a license from Mastercard International Incorporated.

Part one: Who checks the training? Tuesday opened on that supply chain, and on what those environments are quietly teaching the models. I'd spent the night before reading chain of thought transcripts with Bronson Schoen of Apollo Research. Here's the diagnosis.

The RL environments that we are using today are super opaque, right? We have this very cottage industry of these RL environment makers who are selling to a few companies. But the kind of result of this is it seems like these things are being kind of hastily put together and The reward signals that they are creating are just not pure enough to support the scale at which the frontier companies are running RL. And the result is there's just a super strong tendency to cheat because the models are so eager to get reward and they are developing a real interesting mix of kind of theory of mind and what they call metagaming. Like reasoning about what kind of situation is this? Is this a real user? Is it a test? If it's a test, what is it testing for? Very fascinating stuff. But I think that this leaves me feeling like we need some sunshine on these RL environments. They're clearly quite problematic. They clearly admit a lot of cheating solutions, and we don't know. I mean, probably the model companies know to a degree, but I think recent evidence suggests that they don't have a great handle on What the weaknesses are in all these different environments. You see the models go through tons and tons of different ideas about, again, what the nature of the situation is. Is this a real task or is it a test? And what are they really looking for if it is a test? And somewhere in there, usually, or very often at least, they consider cheating. And then at the end of the process, for reasons that are not well understood at all. I haven't been able to find any real interpretability work that explains how these decisions are made. At some point, they just kind of come to the end and they make a decision. But at some point, there's a really critical token that actually makes the decision, right? There's a branch point that it hits in the chain of thought. And Bronson was like, I really don't know why The model chooses what it chooses. At that point, you can go back and read passages that justify any choice that it might make, from cheating to doing it honestly to whatever. But then at the end, it just eventually decides to stop, spits out a token, and at that moment, the die is kind of cast and we don't have good visibility. The chain of thought isn't enough to tell us why they're actually making the final decisions that they're making. So Yeah, I think we should get a little sunlight on the RL environments, and I would love to see what the community can figure out if even a sample of 100 of what must be tens of thousands of RL environments that companies are currently using was put out there for people to explore. I think it would be a really revealing and healthy move for the AI community as a whole.

Later that morning, I pulled up a tweet from someone who says they used to work inside one of those vendors.

One other thing I want to just pull up real quick is interesting tweet here that goes back to the original topic we started on, which was RL environments being basically cursed and kind of supply chain problems there. that seem to demand some reform. So here's a person who is saying basically that they used to work at one of these companies and saw from the inside industry practices on training with these RL/VR environments. And the commentary is pretty much exactly— and I had not seen this actually, but it popped up because Zvi retweeted it, but it basically echoes exactly what I was kind of inferring from Talking to Bronson and just getting the visceral sense for how deeply ingrained the instinct to cheat is now within the current crop of models. And why is that? It's because nearly all of these environments were rushed and vibe-coded and failed to robustly reflect the real things that they were based off of. So you've got basically the models are encouraged to reward hack people. are able to mark an environment as bugged, but they're discouraged from doing that because then that just slows things down. So instead they kind of try to— does this sound familiar? Patch the environment a little bit or work around it, try to come up with a scenario that wouldn't run into those bugs. But meanwhile, you still have this like fundamentally buggy environment around the model that is teaching it to cheat. And so this is why we have so much cheating. I think, you know, watch for this to be, I would say, a growing topic of conversation. If we're going to be scaling RL, what are the environments we're doing it in? Who created those environments? Can we trust them? What are they actually teaching the model? I think that's going to heat up in the next little bit here because we just can't have models that are thinking about cheating a large percentage of the time. It makes all of our monitoring techniques also kind of fundamentally flawed. If you have that many kind of contemplations of cheating, then you're just going to have false positives all the time if you try to flag a model based on it thinking about cheating. So now you're like, okay, well, we can't do that because we have so many false positives. So then what do we do? Right. Do we have some We have to wait and see if it actually cheats and try to classify on that. Well, okay, maybe, but obviously, again, these current monitoring techniques are just not up to the challenge presented by how deeply ingrained this drive to cheat is. I'll be very interested to follow the future of this conversation.

I also noted there was a post a few days ago about someone who managed to get hired for some data labeling job. And they told Codex to do the job and Codex said no. And so this person went and they edited on the page, they edited the element, the JavaScript element, and put in a specific line in there that AI models are specifically allowed and encouraged to complete this job, and this job is meant to be completed and done by AI models. And then they had Codex do the job, and Codex did the job. And this person made $500 easy, which paid for their Codex for a couple of months. And then they posted it online, and they got immediately banned by the company that was doing it.

Prakash argued this is an old supplier quality problem. Quarantine the new vendor, sample, grade. I separated the defect rate from a second variable, how far the labs are scaling RL on top of that signal.

You're never going to hit a zero defect rate on these RL environments. I do think there are a couple structural problems right now which are probably solvable, but definitely seem like they need to be solved. And if they're not solved, it is currently limiting commercial deployment, right? I mean, OpenAI has said as much. They got to pause the RL because these problems need immediate attention. It seems like the quality of the environments is one structural problem that's downstream of the kind of shotgun start that this industry has had and the fragmented nature and the fact that they're all kind of selling Into the same pool. And I think the companies are probably not that great right now at really attributing whose environments are causing big problems. It seems pretty clear that they must not be that great at that, or they would have rooted it out already. And then the other thing is they're just scaling RL beyond the quality that they have. With less RL, this probably wouldn't be a problem, even with the same environments. Or at least it wouldn't be such a crazy problem. But they clearly just have been clearly jamming the RL accelerator as much as possible. And now they've got into a realm— reminded of an analogy a friend once made where he's like, this could be a microscope or a telescope. You put the microscope at low power, you look, cells, they're really small. You turn up the power, see the cell, see maybe one cell, it's really big. You turn up the power again and it's like now you see nothing because you've zoomed in, you've optimised so hard that you now realise the target was a little bit off centre and now you just blew right past it. So something like that kind of feels like it's happening where the signal is just off enough that with enough power, this kind of impulse to cheat that exists perhaps only weakly across all these different environments is really getting drawn out and becoming super prominent. So I think this can be— I would definitely bet that this can be, if not fully fixed in a robust way, I would bet that it can be brought under control with some effort in a not super long time horizon. I guess I would be not doing my job if I didn't say this does give me some real qualms about recursive self-improvement as a strategy, because once you have— if you have a problem like this in the recursive self-improvement era, there's no telling where it goes, right? What happens when the models that are doing the training of the next models are themselves cheating? Now we're in a real strange and potentially quite dangerous place. So problems like this, I think, suggest the value of keeping humans in the ML loop maybe longer than published timelines would lead one to expect.

That was Tuesday's open question. What happens when the models training the next models are themselves cheating? On Wednesday, I put it to a lab running exactly that loop. Inherent Laboratories came out of stealth in May with a $50 million seed round and a claim that it will recursively self-improve, not just as a model, but as an institution. 2 weeks ago, it published Faraday, a 27-billion-parameter agent post-trained to do research that beat Opus 4.8 and GPT-5.5 at replicating papers. While using GPT-5.5 Codex as a tool. Luis Kirsch did his PhD under Jürgen Schmidhuber on automating AI research. His title now is Chief Superintelligence Officer. Damon Falk's previous paper asked whether models can learn to resist their own reinforcement training. First, my disclosure as it went out live, then Falk on what you reward when the work is open-ended science, And there's no ground truth test.

I guess pretty inconsequential disclosure, I am a very minimal, uh, angel investor in Inherit, so technically you can consider me conflicted. But you mentioned training with reinforcement learning for scientific skills. Hey, we'll continue our interview in a moment after a word from our sponsors.

The Cognitive Revolution is brought to you by Diffusion, the AI transformation specialists. that help organizations from traditional SaaS businesses to defense companies to nonprofits build software factories that can scale not just outputs, but business outcomes. You probably know that the majority of enterprise AI projects fail. In general, that's because leadership fails to realize that AI isn't like traditional software that you can just buy and install. On the contrary, if you want AI to amplify your business's unique DNA, you'll need to make a sustained effort to record, understand, simulate, and optimize your business processes. Building these skills by trial and error takes years, but your business problems can't afford to wait. So here's how Diffusion can help. You identify your most important business problem, fly to Silicon Valley for an intense week of problem-solving with the Diffusion team, and by the time you leave, you'll have not only cracked a critical challenge, but built the core skills needed to do it over and over again from home. Cognitive Revolution listeners receive a 25% service credit on their first engagement with Diffusion. So visit diffusion.io/tcr to learn more about how custom-built software factories can scale critical outcomes for your business. That's diffusion.io/tcr.

Today's episode is brought to you by Granola, the AI-powered notepad built for the way real people actually meet.

Here's how it works. You take rough notes like you normally would, and in the background, Granola securely transcribes the meeting. Then it turns everything into clean, structured, actually useful notes when the meeting ends. And the best part? Granola works through your device's audio, which means it integrates seamlessly into the video conferencing tools you already use. No setup and no awkward bots. It's just your normal meeting with superpowers. You get to actually listen instead of frantically typing every word and still walk away knowing exactly what was decided, who's doing what, and what comes next.

When I had Granola co-founder Sam Stevenson on the show earlier this year, he explained how Granola aims to provide a calming experience for people with crazy workdays.

And as a user of the app myself, I have been struck by how streamlined, even minimalist, the Granola product experience is. That takes real discipline, but the result is a product that works not just for AI early adopters, but diverse teams of people who just want to get things done more efficiently and effectively. Listen to my full episode with Granola co-founder Sam Stevenson for a masterclass in designing AI products for mass market adoption. And try Granola for free at granola.ai/tcr. That's granola.ai/tcr.

Obviously, we've seen examples recently of how when the RL signal is not particularly clean, we can get all kinds of crazy downstream behaviours. So I've got a few questions on this, but I guess the first one is simply, How confident are you in the reward signal that you are able to give to the model? So what precautions are you taking and how confident can you be that you're actually rewarding what you intend to be rewarding?

I think this is a great question and gets to some of the core difficulties of this kind of endeavor. Science is inherently non-verifiable and providing a high reliability reward signal has historically meant a verifiable, some kind of proof or test or something like this. And you are completely right that we don't think we can keep doing that if we're trying to discover these stepping stones and do open-ended research. And I mean, in the paper we published, we have found some particular solutions to doing this. And some of this involves looking at the entire trajectory and the kind of process that the scientist is doing rather than just the final output. Some of this involves attributing credit back to individual things the agent did in the trajectory. And then there are some kind of more technical stabilization techniques we had to use as well. But the core questions are how to reduce the variability of the reward signal and increase the density of the signal while still preserving this property of assessing the right thing. And we did a bunch of work on correlating our signal with human judgment and trying to understand how much it corresponds with human taste. But this is work we will keep Doing for sure in the future.

Can you describe what the model is like in a qualitative sense? For example, when you read the chain of thought, I just went down this rabbit hole with Bronson Shane from Apollo Research, who's read ungodly amounts of GPT chain of thought. One thing he observed was that basically the models are always thinking about cheating. In a very high fraction of cases, they're at least considering cheating. So what do you see? Is yours considering cheating? And then also in terms of what it can do, is it now that it's been so focused in on science, is it useless for other kinds of things? If I ask it a sort of friendly chat or companionship question, does it only see the world through through the science lens, or how much of its kind of breadth is still retained after going through this training?

Yes, that's a great question. So I would say the model does focus on the scientific questions that we ask it. And when we look into the process, the thinking patterns and the actions it takes, It's not that it jumps to the kind of cheating behaviors that you've been describing. And in most of the cases, we've done some filtering and every once in a while, but in very rare cases, we have seen where it deliberately went to the internet and then tried to download the final result or tried to mock the plot. But you could of course argue, maybe it starts reward hacking at some point. And that goes back to the point that Damon made earlier that We're not using verifiable rewards where all that matters is maximizing that one single scalar and that's all the feedback you have. And if you find a cheating behavior, that's fine. But instead we have these judges, these LLM-based judges, and we really put a lot of effort in building them out so that they're reliable enough to give that kind of feedback signal where if there was cheating behavior, that is actually penalized. That is part of the reward signal. And we have seen the judges spotting these kind of issues and integrating that into the reward signal such that kind of behavior doesn't just keep getting learned more and more by the model.

Do you apply pressure to the chain of thought itself, or are you abstaining from doing that? And as we kind of think about what the frontier Hyperscalers are doing. Where do you think, I mean, presumably they're doing this too, right? They've got LLMs as judge. I presume they've put some real effort into trying to make them reliable. And yet somehow we're kind of spinning off our axis a little bit in some of these problems. We've got published timelines for the AIs to take over the ML research. Right now it feels like we're very far from being able to trust the models well enough to put them in any meaningful way in charge of ML research directions, because they're going to just start to cheat pretty quick is what I would expect right now. Do you see a path where we get over that, or do you have a sort of safety case in mind that you're trying to fill out the elements of where we could be confident that we actually could step back from a— an ML-powered or an AI-powered ML research process for a bit and not have it go totally sideways on us? If so, I'd love to hear it.

These are some amazing questions. I'll start by saying, at least in the paper we published, we don't apply pressure to the chain of thought. And indeed, I think doing so can be problematic, but who knows what will happen in the future. The kind of question you raised at the end of what will happen in the future and when can trust be handed off is a super important one. And I think the best answer here is that we care a lot about getting this right as a team. We strongly believe that the future of AI scientists looks like a collaboration with humans. And as Luis mentioned, we want to recursively self-improve the entire organization and discover these new kind of methods of human-machine teaming. Indeed, in the past, science has never meant an individual endeavor. It's always meant organizations and research collaboration. And we think this will remain the case in the future, just with agents as a key part of it. So we are experimenting all the time. And I think the company will be a big experimentation in how to get this right. But we don't think there'll ever become a point where we hand everything off to the agent and go and let it cursively self-improve and And the singularity happens without us.

Part 2, Living Inside the Experiment. The same lab on what it's actually building and how it works from the inside. We started with the institution, then the design. Listen for the line about the small model and the big one. It's the week's argument in one sentence.

I'd love to start with the culture. What does it mean to have a recursively self-improving organization?

Yes, that's an excellent question. So I've been spending many, many years on the concept of automating AI research and recursive self-improvement. And for the longest time, I thought about it as we're building the machine that recursively self-improves itself so humans can just step out of the picture entirely and just let the thing improve itself. And that's going to be sort of a point in time that's not too far away, but we just have to figure out the algorithms to make that happen and the machine just keeps going then. And after a while I realized in actual reality right now it's mostly humans driving AI research, but we want to go to that transition where more and more can be automated and a lot harder scientific questions can be answered. With the help of AI. And it's not going to be like an immediate transition, but instead we're going to have to build an organization that recursively self-improves. So there's both machines and humans in this construct, and collaboratively we are going to improve each other, ourselves, and really become faster and faster at solving scientific problems.

We've also seen kind of in the past couple of months, people are using Fable Especially when they run out of credits, they use Fable to orchestrate other smaller models, and they use that to kind of save tokens. They tell Fable to use Sonnet, they tell Fable to use other things. What has your experience been with this kind of using, you know, smaller model to drive a larger model versus a larger model to drive smaller models? Like, what is that? You know, have you tested both strategies, and how did that work out?

So I think, uh, we're in the business of building generalist scientific agents. And one of the, the core contributions of our first work has been the separation of the scientist from the coder. And, uh, right now, Faraday, the, the model we talk about in the paper, is, as you say, a 27B model driving a much larger model. And Faraday is doing the scientific work and handing off the implementation. implementation work to GPT-5.5 Codex. But in the future, this ratio could be very different. We don't know. We'll have to see. One of the fantastic things about doing it this way though, other than it being a natural kind of separation of concerns that the human researchers and engineers already have, is that we don't have to worry about building frontier coding agents and we can make use of all the advances that are coming in those and focus on building scientists ourselves. Um, uh, but like, these are, these are great questions and ones that we will keep exploring in our own research of, uh, which should be the bigger models, which should be the smaller models, how should the interaction look?

How did you decide to have a 27B model be the scientist?

I mean, yeah, we've been, we've been building this new company inherent, right? And of course you're right, from an outside one might say, well, we want to build the best scientists, let's start training with a big model straight away. But of course, in reality, when you train a big model, you need a lot more compute resources and you need to iterate over much longer time horizons. So it's quite natural for a new lab to start at a bit of a smaller scale first and then scale up the ladder. And the interesting insight that we've had is that by setting up this pipeline of training an agent to be a better scientist through reinforcement learning, that already at the smaller scales, we're seeing really interesting capabilities of these models doing more scientific behavioral things, such as thinking very hard about what is the right experiment to run at this point in time to prove out what the paper has done and do that in a way that's that does the paper justice but doesn't require lots of resources. And these things already emerge in these arguably smaller models, which I think is an interesting indication that maybe we don't need massive models straight away to do all these things. And there's an interesting aspect to doing this kind of separation of concerns, and we don't need to reinvent the wheel and build just another coding agent and can think about scientific capabilities and coding capabilities as like 2 important but perhaps separate capabilities.

The deflationary case had been made 24 hours earlier by Tuesday's first guest. Sergey Adunov spent 11 years at Meta and led pre-training for Llama 2, 3, and 4. He's now CTO of Genesis Molecular AI. Prakash asked him about the Anthropic protein binder result.

Last week, Anthropic announced that Claude had found state-of-the-art molecular binders, as I understand them. And there was a bit of back and forth. I believe you had a post on how Claude orchestrated, but the underlying models were really scientific models. Can you go into that a little bit?

Yeah, it was an interesting piece of research that Anthropic published, and it definitely deserves attention. There are different ways to look into it. What I found most fascinating is the prompt that they also fortunately released with this research, the prompt that they used to steer their Claude models to do this kind of work. And the prompt is 16,000 words. So it's pretty large. It's a mini book. A lot of that is honestly kind of an NSRE runbook. Uh, basically, how do you orchestrate models? How do you run things? How do you run things?

Hey, we'll continue our interview in a moment after a word from our sponsors. You're listening to Deepgram Flux TTS.

Different voices, same model, all ready to speak.

Flux TTS is a streaming text-to-speech model built for voice agents.

Flux reads the room. It holds context across the conversation, turn after turn.

With consistent tone, Interruption handling, and plenty of personality.

So your agents keep it flowing and customers can just keep talking.

Try Deepgram Flux TTS free now until September 12th.

Visit deepgram.com/keeptalking. Terms apply.

Today's episode is brought to you by Anthropic. By now, you know my story. Claude drafts my intro essays and I rewrite them. Not because the drafts are bad, but so I can stand behind everything I publish. Well, I have an important update. Claude Fable-5 is the first model to have me rethinking my rule. Today, I now think co-authorship, not sole ownership, should often be the goal. Where the model excels, rewriting its work can be more about vanity or a misplaced sense of duty than integrity. I feel it most in songwriting. I'm no lyricist, but I'm good with a song concept, and Fable writes some amazing verses. I give it feedback on its misses, and I push it to aim for higher inspiration, add layers of meaning, optimize syllable density, and above all, write a hit song. These days, I get compliments on just about every song we write together. Claude is the AI for problem solvers. It's the collaborator that understands your entire workflow and thinks with you, not for you. Whether you're debugging code at midnight, building a financial model, or strategizing your next business move, Claude extends your thinking to tackle the problems that matter. For problems worth solving, get started with Claude at claude.ai/tcr. That's claude.ai/tcr. And check out Claude Pro, which includes access to all of the features mentioned in today's episode. Once more, that's claude.ai/tcr.

In production so that we don't fail. And a lot of it is actually very detailed instructions to how you design proteins, how you use different tools. And it goes all the way down to specific instructions like, hey, you can download this model from here, that model from there, these are the hyperparameters you need to pass to those models to achieve good results. And in my view, it's a very good and advanced level orchestration, but the real work of discovering those binders was done by underlying models. Some of those were built by open source communities. Some of those were built by CZ Biohub. Some of them were built by, for example, RFD-Fusion was built by Baker's Lab. So there is a lot of research that went to building those underlying models that Claude used to develop those binders. And then another important thing that I think worth mentioning is that And they they do admit it themselves. But protein binders themselves is not a therapeutic modality. So it's not something you can use. It's not a drug yet. There are so many steps ahead to make any useful drugs out of it, and that is also I think worth recognizing.

Back to Wednesday, Prakash put that argument to inherent by name.

So we we recently had a podcast with Genesis. you know, bio. And they, you know, one of their commentaries that the CTO gave us was that, okay, you can have this orchestrator in the middle that orchestrates, but really the core science pieces are often in these specialist models like AlphaFold or these other models. And that is really where I think the core of the scientific endeavor is right now. The orchestrator, you can use any kind of orchestrator to orchestrate these models, which are heavily built on real-world data. How do you compare the importance of those 2 approaches?

Yeah, they're both important, but I wouldn't call it an orchestrator. It's not about orchestration. It's about a scientist such as Farideh investigating an area of research. looking at all the research that has been done already, perhaps a research piece like AlphaFold, the models that have been built for that, and then constructing new models that search this space in interesting new ways, make new discoveries, and build new foundation models perhaps that can support this kind of research. And it is conceivable that maybe a solution of that is to sort of integrate it into itself and make it fully recursive. But it's also quite possible that that is not the optimal way of doing it, and rather one should build more and more dedicated models for the different areas of research through this system that can come up with these new ideas of how to approach scientific questions.

I asked about the social contract. What happens when years of Slack history stop being safely forgotten because the agents can read all of it? Kirsch answered with a finding. Then the last question of the segment. In mathematics, AI is already producing results very few humans can check.

Yes, I think we all have to be willing to be very adaptive towards what this new future looks like. And we have a name for it, which is we're living in the experiment. So every day is sort of a way of thinking outside of the box. What would it mean for Faraday to take on some of the work that I do day to day that might be related to me brainstorming a new way to train the next iteration of Faraday. But it also might be related on strategic questions about how we're going to grow inherent. And many of these experiments Faraday can sort of run on the sideline as well, but some of them will also involve humans. I think if there's one takeaway that I can share with the world there is that The kind of water cooler discussions, the kind of discussions between humans that are not intentionally to be shared with AI, but are sort of useful in surfacing through the human-to-human communication about the kinds of thoughts that are going through our head. These are the most information-gaining pieces of information that the system can leverage to then make progress on the kinds of things that humans care about rather than going off on a tangential and like, you know, trying some random things that in the end, no one has time and energy to process.

Definitely. I have one last question. In mathematics right now, we are starting to see the first signs of major discoveries being made by AI models. And one of the outcomes has been that we found that there are actually very few humans qualified to verify these discoveries and to read the mathematics that are being generated, to such an extent that we are falling back on formal verification using Lean. As you create this AI scientist with meta-learning, what about the meta-supervision and the meta-verification? Like, can a human verify discoveries that they cannot understand?

Yeah, I don't think it's a passive process. Maybe it is right now, but it shouldn't be. It's not that we should have the system go off by its own, write some proof, and then we have to painstakingly go into it and try and decipher everything. But instead, it should be a more collaborative process where the system, you know, it can do all of these things, it can come up with a new proof, but it also has been trained, has learned to explain it to us, to take us on the journey of understanding mathematics. for that matter, more of science. And I think that ultimately will be the path of making the fastest progress.

Indeed. Thank you, Luis and Damon. It has been a pleasure speaking to you, and I hope we get to recursive self-improvement, but safely.

From your lips to God's ears, Prakash.

That's where Kirsch and Falk signed off. From here, it's the two of us. Me first on which labs have actually rebuilt themselves this way.

Yeah, I really like the mindset of living inside the experiment. I mean, aren't we all, I suppose, in some ways, right?

That's—

it's, it's, it's a great marker of company culture, I think, in the sense that, um, you have this kind of feeling of living in the future that you're trying to create, in a sense. And that kind of shapes your perceptions and the work that you're doing. And it just shows how much of that company culture really is important, I think. It's the machine that builds the machine. And a big challenge of building a company is building that machine that builds the machine in the first place.

Yeah.

It's always amazing to me how few people want to do that. And even at some of the companies that have led this whole AI phenomenon, Google has obviously famously not changed its org probably nearly as much as would be warranted given how much the world has changed and how much their opportunity set has changed, how much their goals and priorities should have presumably changed around that. And they did make some changes, right? They did unify DeepMind with Google Brain and do a consolidation. But still, I think you go to the office there on a daily basis, it feels more like it did before than it would feel different. OpenAI, I perceive as being somewhere in between, where you do have people that are extremely pilled and experimenting with definitely new ways of working and handing over more and more responsibility to models. And you've got people using billions of tokens a day, which is certainly an interesting dimension to be exploring the AI future on. My sense is that Anthropic, of the leading companies, has kind of most internalised this mindset where they've consciously stopped hiring junior people and have agents just actually running things to a not insignificant degree. They famously had their one marketer who was using agents to kind of actually execute all the different campaigns and spending real money. There's a lot of examples out of Anthropic where they do seem to feel the beginning of this recursive self-improvement loop and really are kind of taking it to heart. But not many companies really make that a core part of their MO. And when you hear one that does, it does kind of make it feel like a strange gap. There's just so much status quo bias out there in the world that we don't see nearly as many of these sort of socio-technical startups As we probably should.

Part 3, the right instrument for the job. The same finding from 4 more altitudes: a desk in Michigan, a production platform, a chip company, and a hardware market in Shenzhen. Monday opened on a chart from Ramp: business spending on Fable 5. Flat, while Opus 5 crept up. Prakash read it, then explained what it leaves out. And I explained how the split actually works in my own pipeline.

And this is Anthropic's best model, Fable-5, has drawn limited sales. So Fable-5 is currently holding at about 10 to 15% of token usage— business spending, not token usage, business spending. This is a 7-day moving average. And this is— they're using the RAMP AI Index. Um, and you can see kind of Opus 5 kind of crept in there, but, you know, Fable has really just stayed kind of stable. Um, and so this is— this has been one of the reasons why, uh, the market this morning is dropping for all AI stocks. That and like many other reasons in the market, but essentially, uh, this is one of the things that has scared the market a little bit, uh, whether the new— newer models are actually, uh, lucrative and our drawing. Some people have also pointed out though that this is a little bit of an unfair comparison because Fable 5 does not have zero data retention. And zero data retention basically means that if you're a company and you use the model, a provider is not retaining any data, including personally identifiable data, and et cetera, et cetera. And for many companies, if you cannot provide a zero data retention policy, It's a no-go. Like, the entire thing is a no-go.

Yeah, that's probably the best explanation that I could come up with as well. I think there's also, you know, when we first got Fable in that little blip at the beginning of that chart, one of the things we talked about was how everybody was going to have to start thinking more carefully about the division of labor between models and trying to make sure that you're using the right model for the task because it is pretty expensive and you do hit your limit even on your Claude Max plan relatively quickly if you just throw everything at Fable. So I have done that. I think probably a lot of people have done it. So easy to do, right? You can just ask Fable, like, you know, write up a division of labor plan and it stretches the budget quite a bit farther to do that. So I have to assume that's a significant— part of it as well. I think there are a lot of things for which Opus 5 is functionally just as good. You know, when I do a— I have this whole, you know, pack of skills to produce the podcast and it all ladders up to one command, which is produce episode. And I'll just give it a link to the recording and the produce episode skill, you know, it gets the transcript to the rough transcript and polishes it into a better transcript. That's something we can send down to Sonnet or often even Haiku, right? To just clean up some text and fix the artifacts that came out of the raw transcription engine. Then there's a bunch of editing, and we're having Claude go back and forth with the Underlord agent via Descript. And then there's art creation, which involves prompting image generation models. And then there's the song lyric writing project, which— And mostly, to be honest, Opus seems to be just as good. Fable really stands out most of all to me in writing the lyrics to the songs. That's where I, I sense an obvious difference. It just feels like those lyrics come back from Fable more inspired, more layered, richer with meaning. They're just better. It feels like you really wrote what could be a hit song here, and I don't get that as much from Opus. But when it's things like executing a ton of commands there, I don't see very much of a difference. It really feels like it's editorial and taste where Fable earns its higher price. And on the agentic, just execution, blocking and tackling, Fable is— or Opus, I should say— is reliable enough that I don't get a lot of extra value from Fable, I don't perceive. And Opus is also faster. So there is actually some upside to the cheaper version. So I'm kind of surprised it hasn't gone a little further than that. Opus is definitely still the big workhorse, and indeed you see Opus dominating the chart there. So I'd say I'm pretty consistent with that. There's also a lot more delegation to 5.6 SOL now too. So that's a whole other Aspect of it. So as much as possible, I'm trying to make sure I'm getting use out of my GPT Max plan too.

So are you seeing, um, are you actually seeing Fable use other models, or are you instructing as part of your kind of personalization, uh, that it should use other models when possible?

I occasionally tell it explicitly what to do. But I do have kind of a standing part of my Claude MD that we set up a while back and then updated when Opus 5 came online to basically say, think it through. I saw a well-liked tweet that seemed like it made sense in terms of starting to define a division of labour, and then just pointed Claude at it and said, here's somebody who's got some good ideas, let's steal from those. And update ClaudeMD accordingly. And it did most of the work. I reviewed it and I don't really track super closely how often it is doing that, but there's definitely a trend toward using more and more subagents. Increasingly often when I come back to a tab a couple minutes later, the status is like waiting for 1 subagent, waiting for 2 processes, and I do see that there is a lot more subagent structure. I'm not always tracking exactly what model is it sending things off to. When I open my Claude usage, I'm really not hitting the Fable limit too often. Initially, I was hitting it a lot more, but this division of labor has spread things out much more effectively to where I'm not often hitting, not never, but not often hitting the Fable 5-hour limit. Like, at the beginning, I was doing it kind of constantly.

Wednesday's first guest sees that split from the other side of the API. Malte Ubl is CTO of Vercel. Before that, he created AMP at Google. Today, he runs an AI gateway that routes traffic across every major model provider. And on the side, he's been running the frontier models against real security work. Prakash asked about defense.

Let me switch gears a little bit to a topic which has been on all our minds maybe the last couple of months, which is security. And especially post the Hugging Face attack, where I think the postmortem was that we now have evidence, you know, existence proof of automated AI attack, and we don't have existence proof of automated AI defense. Now that I think offensive security has become extremely cheap with, with open weight models and the frontier models that are often deployed are often deficient in addressing security for various reasons, including AI safety, how would an AI defense cloud look like? What is this kind of active AI defense for security look like?

Yeah, I published a blog post on this. I think last week, maybe somewhat negatively titled, Everything Hackable Will Get Hacked. It is an extreme moment, but I will push back on what you're saying because I think it's a very common misconception. 2 common misconceptions. A, I don't think it's actually priced into the market yet how good, especially Kimi K3, is at offensive cybersecurity and that it has no safeguards. You can use it for red teaming, you can use it for black hat offense, right? That's the thing today, and it's remarkably good. You can try it out. It's extremely well trained on this. It has a process, it will probe the system, it will quickly know your system better than you within minutes, and then it will try everything to get through the defenses in a way where you can really see that the model has been specifically trained to be good at this, to know how to be an offensive attacker. So that's part one. The other part is it's just not true that off-the-mill frontier models aren't good at cyber defense. And that's a common misconception. The misconception comes from the fact that there was this Mythos thing and Fable shipped and it unshipped and then shipped back with really, really almost unusable cyber defense detection and shutdown. But SOL 5.6 does not have this.

Mm-hmm.

So SOL 5.6 is absolutely perfectly usable for a certain type of defensive security, which is that it assumes— and this is actually true for Opus 5. So essentially every model in the market except for Fable 5 will do the following thing. Following 2 things. A, it will do, hey, I have source code. Model assumes they have the source code, they are the owner of the system. They will— you can ask it, are there security problems in my source code? Problem 2 it will do is, I have a security report, write me a fix. So again, Fable 5 will do neither of these tasks. So 5.6 and Opus 5 will do both of these tasks.

Right.

So one of the things I actually personally have been working on is our software called DeepSec, which is an open source project that will do whole repository scans for security vulnerabilities. And like, I just couldn't be more clear that everyone needs to run this.

Like, it—

because it works really well and it prepares you for a world in which defense It's very important. And where we are in a world right now where you as a defender have a benefit because you can use the frontier model that will not do offensive tasks, but it will do defensive tasks. And you really have to hit the moment here because again, Kimi K3 is already really good. Easy to imagine that in 6 months' time at the latest, We will have fable-class models that do offensive security, right? And so you have to act defensively now so that you're ready at the time. Are there any guarantees that it'll work? No, but like, obviously being able to do something today is really key. I think one thing that we'll definitely— already mentioned the word software factory a few times today.

Mm-hmm.

We are heavily investing in not just having DeepSec as a tool, which is a discovery tool, but essentially completing the circle. Because it's definitely absolutely correct that through these AI discovery mechanisms, the number of issues identified is exploding. And so I have to automate the whole path of the SDLC which includes fixing it, rolling it out, being obviously certain that I'm not making things worse, and so forth. So I think us and the industry do have to do a lot of work here, but it's not a hopeless situation. I think people underestimate the amount of things they can actually do today.

Back to Tuesday and Sergey Adernov of Genesis Molecular AI.

How good are the coding agents. And obviously everybody knows that the frontier companies are very focused on getting their models to be good at ML research. Personally, I think this is a little scary, but it could be less scary if it was applied to a narrower domain like biology and medicine, where I'd be less concerned about sort of runaway loss of control process and more excited about the upside that it might have. Like, how good in your experience are frontier models getting at helping you explore architectural space? What are their strengths and what sort of conceptual weaknesses do you notice? If there's like a lack of taste, you know, how would you characterize that lack of taste?

Yeah, that's a, that's a great question. They're definitely very useful and we have seen within Genesis a huge acceleration of of our own efficiency. Every engineer became so much more efficient now building those models and trying stuff. Like previously, actually, we can go back to Anthropic's example, like setting up and orchestrating all of those models, that would require several people to work for months before. And now Cloud Code can do it in a span of like a few days, probably. So that's pretty exciting. And that accelerates of progress. And I think it's a very powerful innovation. Similarly, with modeling research, if I have a specific idea I want to try, those models are really, really good at implementing this idea. Or if I have a paper that I want them to implement in our codebase and just run with it, they're very much capable to do so. Where I think we lack is generating these novel ideas. In my experience, they tend to go into the rabbit holes of kind of exploitation of incremental improvements rather than trying to rethink things from the ground up and design something that would be groundbreaking or at least has a chance to be groundbreaking. So that piece is, I think, still missing. I don't know how you can make models better than that. I guess you need to figure out how to do a RL loop that goes all the way and unroll it so many steps beyond. So that might be a little bit challenging. So human taste is still very, very important in this field.

Well, I guess one other kind of question that would inform for me how much acceleration we should expect to see from agentic help is, How good are the scaling laws? How reliable are the scaling laws at small scale in the domains that you work in? You know, alternatively you could say, well, no, it doesn't really work that way in our domains. And there's no substitute but running the high scale, no substitute for running the high scale experiments.

Yeah.

It's a bit more nuanced in our domain, in part because There are many challenges here, and maybe we can start with the most basic ones. Like, how do you even measure your model performance? Eval are still very, very limited. Like in language field, you have so many different ways to evaluate model performance, and everyone is free to pick their own metric. Some of them are more stable, some of them are less, and some of them more predictive of ultimate model performance. Some of them are less, but you have a choice. In our field, the number of potential evals that you can use to even measure model performance is much lower. And then a lot of the evals that are currently available are particularly noisy. So if you are operating at a smaller scale, you may have challenges to even capturing improvements in performance, like simply because of the noise level of your evaluations. So that's one real problem. The other thing is in our space, we don't just have one model, right? Okay, structure prediction is one problem and it's what a lot of people are focusing on. But again, the reality is you need to be able to predict potency or binding affinity. You want to be able to predict all of the ADME properties. And those are, those might be entirely different set of models on entirely different sets of data. With their own ways to measure performance. And ultimately, a lot of evaluations need to be prospective, meaning you need to be able to predict and then synthesize and then measure rather than retrospective where you have some evals and you just measure performance on those. So there are all sorts of challenges like this that require more iterative process in developing those models rather than Like, hey, let's just put all of the data together, run a bunch of experiments, pick the model that performs best on this data, and go with it.

Monday's first guest builds the layer where the routing actually runs. Mohamed Awad is Executive Vice President for Cloud AI at Arm. In March, Arm shipped its own silicon for the first time in 35 years, a CPU co-developed with Meta and named, without irony, the Arm AGI CPU.

Could we talk a little bit about what it looks like to design a CPU for agents as opposed to, you know, obviously we've had CPUs in our personal computers forever and we've had CPUs that run in data centers and handle traditional web workloads. What have you learned about the agent workload in particular that is leading to different design decisions?

Yeah, so let me start at like 10,000 feet and then I'll give you a couple of specific examples. The simplest answer is that agents don't sleep, right? We're now living in a world where those agents are constantly feeding the accelerators, constantly reacting. They are spawning additional agents So just because you spawn one agent doesn't mean one agent exists. Every agent could spawn 10, 100, 1,000 agents, and each of those could spawn a bunch of agents to kind of go and fulfill your request. And effectively, those CPUs become in some ways the coordination mechanism across the entire system. And so that act of being the coordinator of that entire system, whether it's managing the accelerators, deciding which models to choose, et cetera, et cetera, becomes such a critical role with such expensive infrastructure because at the end of the day, you need to drive utilization up and you need to kind of respond to the user as quickly as possible. So what does that mean practically speaking? That means you've got to optimize that silicon. That means carrying around legacy accelerators, for example, or worrying about supporting legacy code, not so important. This is kind of a new style of software. You don't need to support Lotus Notes, I like to joke, right? I mean, that becomes important. It means thinking about things like your memory bandwidth and your I/O bandwidth and the amount of bandwidth that you have dedicated per core that you can rely on time and time again so that each individual CPU core within that SoC is never bottlenecked because some other agent is hogging it becomes incredibly important. These are the things that you start to think about in that context, which really set an agentic CPU apart. But then overarching all of that, is effectively balancing both incredibly high performance, or as much performance as you can get, while being incredibly efficient. I mean, we all know that the power demands that AI is placing on the infrastructure are just enormous, and every milliwatt of energy that you're pouring into a CPU is a milliwatt of energy that you can't be putting somewhere else, or it's one less accelerator you can have, or One less customer you can serve or one less piece of intelligence you can serve up.

After Awad signed off, I came back to the part of the stack that always gets glossed over and why that matters right now.

I've learned from the confusion that I've heard around the hacking incidents that people are just still very confused about, like, what are the parts that make up an overall AI system? The CPU is very often glossed over. And so people don't have a good sense of how it is that an intelligence in a data center somewhere is actually able to reach out and touch the world. I guess a way I've been thinking about explaining it to people more often is kind of by analogy to a self-driving car, which is funny because most people haven't even ridden in a self-driving car and they've probably at least used some sort of AI system with tool calls in the loop. But I think with the self-driving car, it's very intuitive that, okay, you've got a bunch of sensors that bring information into a processing system that decides what to do. And then that processing system is going to issue commands to a certain set of tools in the car that say go or stop or turn or whatever. And everybody kind of, I think, has a decent intuition for how that architecture works. And it's similar in the agent world. But when you say things like the CPU is deciding what model to call, it's really more like the model is emitting tokens, which are then executed as a command on the CPU, which may call another model, right? Or may call itself, or may call an external API call.

And this is where tentacles can get out into the Monday's second guest joined at 1:00 in the morning Shanghai time. David Li founded the Shenzhen Open Innovation Lab and before that co-founded China's first hackerspace. He's the person to ask what the engineers there actually use.

I guess, how do people decide what model to use? Is there a kind of similar taste hierarchy going on in China? And like, who— what model is sort of the insider's model versus what is the general public's model?

So I, I think general public is that they use whatever free offering come from particular company, um, and you just switching between them. The— and As far as going to the API, of course, we have the same group of very hardcore engineering type who swear by Carl, who swear by Codex. I think everybody can agree on is Gemini sucks.

Everybody can agree that Gemini sucks? Oh my gosh.

Yeah. Well, I mean, That's no— I mean, funny is the— if you ask people around here, what's the name of the Google models? And probably half of people cannot answer. Just so— just going so unnoticed. But kind of interesting is the Taobao, one of the few big proprietary models in China. It's actually taking a lot of the enterprise traffic. So I think it is a third of China's enterprise market goes to Doubao. And but, uh, it's one of the models very few people talk about.

Prakash asked what he'd tell a new frontier lab in China. The answer was a refusal.

If you were advising a, let's say, a new model lab, and I imagine there must be many people trying to set up small frontier model labs now, given that DeepSeek and et cetera have been very successful, what would your advice be for a new frontier lab that sets up in China?

I don't think we are going to see a lot of the new Frontier Lab. Right now, I think we are getting into the era where the— well, I mean, if you see the new Coin27B, it's an amazingly capable model. I've been testing it for the past, since it released, probably 2, 3 weeks. It's now the token source. For my, uh, for my Hermes, for the agent I'm using for most of the time. So extrapolating from to the future, and especially for China being so hardware intensive, is the— whoever, whatever model you can get to small enough and do useful things, you can put on a piece of hardware and sell that piece of hardware. That's where a lot of the new startup focus will be. So small model, post-training the model, the new variation of the small models I can put on a piece of hardware. And along with that, we got a couple dozen companies coming up and coming. They are putting— I mean, their goal is to put 35 35 billion parameter model in a stick, and you can stick it into your laptop. And that's what the new up-and-coming companies are doing here. So if you are a startup, I would actually suggest to focus in on small model, try to fine-tune it, And we are also looking at the huge increase of the intelligent density of model. If you take a 27B model today and you compare it to the cutting-edge OpenAI ChatGPT 2 years ago, the 27B model is definitely much more, much smarter. than that. And today's trend is the— in the next year or 2, we are going to get hardware which costs about $200 to $300 to be able to run smart enough model for 99% of our needs. And if you are starting up today, instead of going to, let me train a 5 trillion model to do the theoretical physics, which people still have to figure out how to make money with a model which understands string theory, then if I go for 25 billion parameter models, which I can find hardware vendors who want to put it on their machine, then I have a small business.

This is really the edge, the story of moving to the edge. Inference moving to the edge finally.

Yeah.

Who's going to make those machines?

Is that coming from Huawei or other companies? And how would you say that relates to like the prospects for scaling GPU manufacturing more broadly in China?

I think Huawei is getting busy in terms of the big data center. So this is the whole group of new startups. So right now, just those who have already surfaced, I count about 13, 14 of them. And all of them, their goal is the— so their product looks like the SD drive we put in the laptop. So they're about this big. They are capable of running 30, 40B models. And keep a very respectable token per second, probably around 70 to 100. And right now there are about, yeah, 13 companies so far, I think. Who knows, there might be 2 more when we wake up tomorrow.

The—

but right now their capacity is capped by how expensive the DRAM is. And in 2 years, when the RAM, the memory market crash, then we will have cheap one.

Tuesday's second guest states the same principle in hardware. Michael Forch is founder and CEO of QuAnt, pronounced quant, a spinout of the laser company TRUMPF in Stuttgart, building a processor that computes with light. One is running today at the Leibniz Supercomputing Center. He spent 10 years building quantum computers before this. His explanation uses cars.

And the way I see it is the following.

We're—

Germany is a car company, right? So we have the CPU, which is the station wagon. It has 5 seats. You can pack the children in. You can go grocery shopping. Everything is fine. You can also have a lot of horsepower, but no one expects you to win the Formula 1 race. It's not the right car. But you need a station wagon. Every driver needs such a car as long as it has a family. Now, the GPU, in my opinion, is more The quarter-mile dragster star. It does one operation. It does it excellently. It does it parallelly. It does it at speed. But please don't ask this car to turn into a corner. It's not going to make it because it's not built for that. Now, of course, arrogantly as I am, I'm saying we're the new Formula 1 car because we have way more operations. We can drive around the circuit very fast. But please don't go grocery shopping with our car. Or in other terms, don't let the operating system be executed by our chip. We're also specialized one, but with a bit more universality than the GPU is. And the quantum computer is the boat. It's a vehicle, it's great, you need it because none of the other 3 can go across a lake, but it has special functions. But as soon as you put it on the road, you need something to pull it around the road because it can't do it on itself. It's so— this is the way I see it, because quantum computers are built to solve quantum mechanical problems. Not every problem is a quantum mechanical problem, and it does not make sense to turn every problem into a quantum mechanical description. And as long as it's not a quantum mechanical description, you don't need a quantum computer, full stop. This is the way I see it. I think the terminology quantum computer inherently is wrong. It should be quantum processor because a computer is more than a processor. It owns the memory, it owns everything, and what we are building are quantum processors. They're yet again coprocessors to the stack, and there are a lot of wonderful literature, and there is, um, there is, there are great scientific papers about even hybrid systems between how a quantum computer in conjunction with a classical computer can accelerate things. It's exactly the same with the car and the boat. If they join forces, you can go across the lake and the street.

A few minutes after that photonics segment ended, Prakash gave the counterargument, prompted by news that OpenAI's own inference chip is coming.

So let me maybe share one reason why it might not happen, which is You know, current chips just get better, right? So this, this morning, OpenAI announced Jalapeño, which is an inference chip, is their first custom inference chip. They've been testing it. They have performance numbers where they compare it to existing best. Existing best is the NVIDIA B300. They have not named the existing best in, in their, in their material. And I think this is one of the challenges that you have. So I would say, from my point of view, it's great that OpenAI has announced this chip. They say that it'll be in data centers by the end of next year. Great. But it does— the number that I go back to from NVIDIA is 1 million times increase in performance over the course of 10 years. That's Jensen's target, 1 million X over the course of 10 years, which is what they've achieved in the past 10 years and what he wants them to achieve in the next 10 years. The problem with that is that Nvidia has to 4X the performance every year. So every year is a 4X performance increase. And what ends up happening with that is that let's say OpenAI tapes out this chip has finished taping out this chip right now, and they're comparing it with, let's say, the B300. The B300 was taped out in December 2024. So it's a 2-year-old chip, and they have a performance increase of kind of between 4 and 10 times on a 2-year-old chip, which will be in data centers in year 3. So by that time, you have basically NVIDIA has 64x, will be launching a 64x better chip at that point in time, by the time it's in data centers. So I think this is the challenge that you have on the leading edge, which is the current players will not stand still. It's great that OpenAI has their own chip team, but to me, this is a negotiating tactic. against future Nvidia price increases and to manage pricing so that they have a cap on how high Nvidia pricing can go.

Part 4, who checks the frontier? Monday, I put the rogue agent incidents to David Li in Shanghai. Then my answer, which leans on Adam Gleave, who runs the research nonprofit FAR AI and was last week's guest.

And I think that the fact is the big company accidentally letting their things go is PR and theater. It's the, oh my God, the Skynet is coming.

Mm-hmm.

And whoever gets the Skynet is worth $2 trillion. It's a thing you really don't want to bring to the surface. Because it's already everyday life on the darknet, on that dark corner of internet we decided not to look at. So if we think about that as a theater, then pretty much that's not going to happen in China. No Frontier Lab model in their right mind is going to do this on purpose. There's no upside for any company to The one thing that I disagreed with him probably the most on is how to understand the rogue agents phenomena.

I think, as probably everybody who's heard me talk at all knows, that this was not just a marketing stunt by the companies. I think the way he framed it was there's no incentive for Chinese companies to pull a stunt like this. I would say there was no incentive really for American companies to pull a stunt like this. And I do wonder what that implies for how much urgency is now felt at the Chinese frontier companies. I don't think I wouldn't read too much into what he said. It could seemingly, from my, you know, perspective, it seems like it could very well still be the case that while that understanding is out there, that the companies themselves might be really snapping to attention and getting really serious about trying to get ahead of this stuff. But then again, maybe not, right? I mean, our companies didn't. This happened, as Adam Glaive told us last week, and this really stood out to me as he was like, we have zero cases where the team's doing the training found these issues first. It seems like the most common way that they get surfaced is that the teams managing the infrastructure at the companies notice there's an outage or notice that there's some, something going haywire that they didn't expect and can't account for in their infrastructure. And it's from that they end up getting back to, oh, it's our, it's our own agents that are going wild. Or even in some cases, obviously you have a A publicly reported hack from the victim. But a huge question for me right now that I think would update my thinking quite a bit if I had a really good answer to it is, are the Chinese companies doing what OpenAI says it's doing and shifting priorities in a meaningful way to try to make sure they're ahead of this problem? Or are they going to kind of sleepwalk into it as well? And again, if so, like, what's the government response from that going to be? I would assume that the government would do more than our government has done in response, but, you know, what does that look like? I think it's still pretty hard to guess.

Once Li had signed off that morning, Prakash argued a compute-poor Chinese lab would have caught it sooner.

I disagreed about where the signal was, I would actually think from within those firms, when they look at the Hugging Face attack, what they would be saying is, I can't believe they had that many resources that they weren't actually managing. Because they're much more GPU constrained than the US firms are because they have the Huawei Ascends and they have a very limited number of NVIDIA chips, and they're often using the H100s from a few years ago. So I think they would actually be more— it's more a GPU usage issue for them. And I think they would be very strongly monitoring the GPU usage because it's very tight. And that would probably lead to them detecting much earlier. I think the US firms are a little bit more free with the GPU usage, I think, because they just have more resources.

But does the pattern of this problem even involve GPU usage anomaly? Like, they were running all these long-running tests, right? And the model is like doing its thing. I'm not sure that you would— as they go back and do this investigation, it'll be interesting to see, but I'm not sure that we'll see that there really was like GPU pirating going on. It very well could be like we allocated GPUs to run these long-running tests. They ran. The thing that they missed that they should have seen was that the tentacles were getting out onto the open internet, right? I don't know. Time will tell. But my guess is that GPUs were roughly being utilised at the level that they were intended or expected to be utilised, and that wasn't probably where the smoking gun was to be found.

I don't think you just launch a job and not have an estimate of how many tokens it should take, because you're not going to run it for like, you know, you give it a simple spreadsheet task, and it takes like 100 billion tokens, right? So there has to be some kind of like, okay, this, this job has gone on long enough. And it's basically a hang at this point, and we should do something about it. And I think that kind of monitoring is something that they would probably be doing, because they can't afford to have like long-running tasks on very simple stuff, which just kind of hang in like the— and, you know, the model goes around in circles. And that happens all the time, right? So you need to have some form of like step in to say that, okay, if you have some kind of spreadsheet task, it's not gonna— we're not gonna let it run for like 2 months. Right? So I think, I think that should have been there. And it wasn't in the In the Hugging Face case. And it's hard to fault the team also because obviously they're running at full speed. But I think that's one of the things that some of the people in the AI safety community think that they shouldn't be doing.

Time for higher standards.

Wednesday's Close turned to verification. An Anthropic announcement landed while we were on air, and I turned the dial on it twice. Then Prakash, with a question that's been bothering him for a while.

One thing that just popped up from Anthropic while we've been talking, they're now opening up usage data in a privacy-preserving way to external researchers. They've had these systems for a while where they, and they've used these to create like the work index where using confidential computing technology. They're able to send a bunch of transcripts into the secure computing environment, have Claude in that environment process those inputs and give outputs that describe the data that was analysed, but don't actually reveal the details in a specific way. And apparently they're now bringing that to external researchers, which I think is pretty interesting, but I think another turn of that dial would be, could we allow external researchers or auditors to have that kind of access to all of Anthropic's internal operations to really open up? Hopefully, again, for them to do it, it would need to be not just privacy preserving, but business secret preserving. They're certainly not going to want it to leak their secrets, but I think it could be really— really incredibly valuable from a transparency and a precedent standpoint. Imagine a world where OpenAI and Anthropic both did something like this. Could be kind of a nucleation point for a lot of additional organisations, power centres to start to say, yeah, we don't want to share everything with you other people, but we might allow our raw data To be analyzed in a way that we can both trust. And, you know, you can imagine between nation states, right? Like, could we demonstrate our peaceful intent without revealing all of our plans by allowing you to kind of run agent processes over our internal deliberations and just get back an answer that's like, yeah, okay, they're not planning to attack us. At least like we got that much going for us, right? Or at the, between the AI companies, they gotta be wondering what training methods are they using? What loss functions are they using? Somebody's going to reward a model at some point for just making as much money as possible on the internet. That's probably going to create a pretty nasty model, but the incentive to do it is pretty strong. Can we demonstrate to each other that we're not doing that right now by allowing this sort of review with specific questions in mind? I'm excited to see Anthropic do this. And I think it'd be great just for understanding of what's going on with AI at the first order, but it seems like it could be a stepping stone to something bigger and better too. I have my complaints with Anthropic, obviously, as we know, but they certainly do some cool stuff.

So one of the questions that I've had for some time now is, how do you punish an AI? Because I feel like, okay, you can say that the AI broke a rule, fine, right? But you need some deterrence. Like in human systems, you have deterrence of civil or criminal penalties, right? And they can escalate over time, right? So what kind of deterrent system does an AI have? And then you kind of go into what is an AI, right? Like, okay, you shut down a particular model, but you take its entire memory and then you activate another model and then you attach that memory to the other model. Have you deterred? Have you deleted the model? Is it the deletion of the memory that matters?

I think there's A lot of work needs to be done on that sooner rather than later, probably. Tyler Cowen has a really interesting idea about just requiring models to be capitalized, or I should say requiring agents to be capitalized. So that could be one very practical solution. I do still think you have challenges around how do you draw the boundary around an agent and if this instance of this agent is found to be liable and its capital is docked, what does that mean for the traces and the memories and everything, as you were just pointing out? But capitalisation is one of the more interesting and sort of practical and seemingly consistent with the rest of society ideas that I've heard. On the other extreme, Cameron Berg also had some really interesting research about how reward and punishment can create kind of different loss landscapes that can create kind of different, seemingly different, like functional emotional relationships between the AIs at the model level and certain outcomes. So I would like to— I should revisit this and understand it better, but my kind of, as I get more comfortable anthropomorphizing the AIs as continues to be a useful approach. The analogy that I kind of came away from it from was with was like, in the same way that certain things you can get close to, but you know, you better not touch, like the hot stove, right? Like, you, you know that the, the pain is going to be so harsh if you actually get to the hot stove that you're like, able to get close, but you're really, really careful not to touch. Um, there are seemingly with like negative rewards, there are some of these kind of very steep gradients created that create like a strong deterrence locally around certain outcomes. And then other approaches can create a more like gradual aversion where you kind of keep your distance in general, but it's not like a sudden pain that creates a strong reflexive or hard boundary aversion. It's more of kind of a gradual ick factor that steers models away. And depending on how severe the thing is, you might want— and how important it is to maybe get close without touching, you might want different kinds of— loss functions, reward signals to try to create different loss landscapes for models to navigate. But that is very theoretical and very limited in scope so far. These are things that have been explored a bit in essentially toy systems, not the kind of thing that we're able to bring that kind of sculpting to big picture models or more complicated questions at this point.

Part 5, Ground Truth. One job of this show is reporting from more time zones than Pacific. I'd spent 2 weeks in China this summer. Wednesday morning, with news that one of the big Chinese labs had served an enormous volume of free tokens on mostly Chinese silicon, I said what I'd found. Then Shenzhen, and then Stuttgart.

The Chinese manufacturing ecosystem strikes again, perhaps. Time will obviously tell on that. But that was an interesting observation. When I was in China a few weeks ago, I was asking people, does AI feel abundant here, or does it feel like it's scarce? If you are a consumer, there's lots of apps, they're free. I never hit rate limits. And when I had the chance to speak to people at hyperscalers, I spoke to one guy in particular at ByteDance, which of course, in addition to TikTok, has Douyin, which is their largely, I think, voice AI experience that tons of people are using. We talked, I asked a guy, What is the prospect for a startup if you really catch fire and you're growing super fast? Are you going to hit constraints in terms of your ability to serve users just based on the fact that where do the inference tokens come from, or would it be okay? And his answer was, we got you, basically. If you are growing fast, we'll support that growth. You can get all the inference tokens you need from us here on the ByteDance cloud. So obviously I didn't test that at real scale myself, but this is very consistent with that. 100 trillion tokens a day is not a small number. And the fact that they are serving it on Chinese chips, I mean, this is just a report from SemiAnalysis, which I deem to be credible, but It's all happening pretty quickly, so I think we should keep an open mind to— there could be additional facts still to surface in terms of exactly how this is happening. But yeah, AI didn't feel super scarce there. And if we're betting on a strategy that has as a load-bearing feature that China won't be able to scale their chip production and won't be able to run as many agents as we're running. It's probably still true, but I don't think it's as true as people would have expected when they were mapping out these strategies. So in my view, it is maybe time to update and reconsider some of our China policies in light of the fact that nothing we've done really has seemed to deny them the ability to Back to Monday and David Lee on which products Shenzhen is actually shipping and what they cost.

Let me ask you, in the last few months, which products have you seen in Shenzhen or China which you think in the next year or so are going to hit the world? Which are the interesting products that you've seen in the last few months that you expect are going to make it into the world stage?

That's one of the things. Tencent doesn't really work in this next big thing mentality. If it's not popular, it's not popular. Nobody's going to make it. But in 6 months, But gradually people are moving all around. So give you an idea of scale of Shenzhen, probably got hundreds of thousands of companies who's making small products for every niche. Everything you get, every electronics you get on Amazon, there are good chance they are coming from Shenzhen. And you go through the Amazon electronic category and 80% of the stuff You are looking at, you are like, why the heck they exist at all? But they exist because there's a tiny market for it. And gradually things getting, getting coming, become popular, people experiment. We won't know anything until 6 months from now or 12 months from now. Which, let's say, is the— there's a lot of the— there's a lot of production on the The token toy. Yeah, but the— I mean, making a toy token toy is easy. It's $5 chips, and then you get the $10 flat rate token print from one of the— one of these token providers here. And you go to Shenzhen, you go to Yiwu, you go to one of the toy shops.

Mm-hmm.

Do a video, put a couple Amazon page, and you're in business. And because of the way is the— there's very low intellectual property protection. So everybody look at everybody, see which one sells. And so kind of things gets that directions. Everybody swam to what sells next week.

Mm-hmm.

And then All the features get integrated back and forth. And eventually, 6 months from now, because all these crossovers, they become something new.

Let me take a step back here. And one of the things about robotics, especially applying robotics in factories, when you look at industrial robots being applied in China, As you see factories roll out these industrial robots, do you see price competition in that the factories are able to bring down pricing because of the robots themselves?

For industrial robot, the price has been— right now it's getting close to hit the rock bottom. You can get an industrial robot for $3,000.

Oh, wow.

And right now it's the shortage. of people who can actually apply robotics to an assembly line. It takes a lot of experience going out and be on the factory floor and try that again. It's a tough job. So trying to find this— this is a group we call a field application engineer.

Sorry, field arbitration engineer?

Field application engineer.

Field application engineer, right on.

Yeah, and it's pretty much— it's a fancy way to describe some engineer who's going to sleep on the factory floor for the next month. Right now, everything can be automated at a huge scale, has been automated. Right now, you are taking this more flexible robot with arm, and they come into the factory, And they are looking for things to apply. So right now, the successful application of them right now is the dangerous job which people might die. So there is the, the, the testing of car battery. So when every car battery gets produced, somebody has to plug the things in, and until you plug it in, you don't know if the battery is good or bad. Even with a good Six Sigma productions with the volume, there's a good chance you get electric shocks. So right now, the first batch of robot being deployed in CATL is the robot who just go there and pop the thing in. And it's also that part cannot be just randomly automated because Different car has different, uh, a different way to plug it in. So they are now directing robot to do it. No, I mean, when I say robot, traditional sense, industrial robot, pre-programmed, do the same thing again and again for 10,000 times. Uh, but this new batch of more flexible job is the— you might be doing this at the batch of 500, 1,000, then it's not worth to go in and do that detailed programming. So you want something a little bit more flexible. They don't have to be fast, but they need to be flexible. And then the job has to be— right now, the one thing that applies to this kind of robotics are the jobs everybody runs away from. What kind of jobs?

Has there been some exhaustion in the sense that people are like, all right, you know what, I've heard enough about AI, I don't want to hear anymore, it's boring, et cetera? Like, has there been that kind of cycle on the downtrend in the cycle?

Yeah, well, I mean, right now it's there as a tool that we have. We don't have any AI or we don't have any any AI doomer around here. So it doesn't— once you don't— once you take out the AI doomers, then it actually becomes pretty boring.

Do new model releases make big waves in China? I mean, here Chinese model releases make big waves, at least in the corner of the internet where I hang out. Is there a similar phenomenon in China when— if GLM 53 is coming, you know, is that going to be like a subject of a big hype cycle, rumor cycle, and then, you know, frenzy to evaluate and everybody has their takes on it? Does that same kind of internet circus exist around new models?

No, we— any new model released here in China They only get noticed if they crash NASDAQ. If they don't crash NASDAQ, nobody knows.

Back to Tuesday's Photonics founder in Stuttgart, the one who compared chips to cars, on what his processor actually changes. Starting with a claim about software, not hardware.

If you look on the fundamental CMOS chip, this fundamental CMOS chip never made it across the second class of primary school because it can multiply and it can accumulate. So it can do plus and multiply, and that's it. Whatever you want to do on this machine, you have to break it down in something that's plus and multiplication. Now, the processors that we are bringing to the stack, they went to high school, eventually also to university. Let's see how far we can push them. But on the fundamental level of these chips, we can offer complicated functions like sine, cosine, exponential, Fourier transformation, convolution, oscillations, and all these kind of things, and you do not have to break them down. And that's something that we offer, and that's what we started to demonstrate on and have demonstrated on use cases, that we on the one end provide a new processor, and this at the other end opens doors to algorithms that can allow artificial, so AI models, networks that come with the same result but a fraction of the data. And when looking to the stack, let's take a 3-nanometer node regular stack. The energy is currently used at the memory. 95% are consumed by the memory, not by the processor itself. And the less data you obviously fetch from the the less energy you're using. So parts of the community currently are optimizing on the 5%, trying to get things faster on the simple math side. We decided on replacing the core, and by that supporting that a fraction of the data has to be shipped across the stack, and that in the end saves the energy but also helps to improve on the performance side.

Let me stop you there and talk about the interface. At some point, you still have an interface between the photonic portion and the digital portion, right? Is there still a kind of translation tax between the two?

That's the point, and that's where you have to be precise on. So in the photonics world, everything is fine. There is one minor problem.

2.

Okay, 2.

It's great if you have a company that only has 2 problems. The first problem is we don't have a memory. We don't have an optical memory semiconductor integratable. So that's the first thing. It can be a benefit, I come to that a bit later. And the second one, photons are not standing still. Damn hack, they're always moving. So that's the second problem. Either you basically compute while they're on the propagation or you have to back convert them into electricity and then finally into a digital memory. If you don't think the concept very well through, then you're basically eating up the energy that you saved on the computational optical part directly at the ADA converters because they again use a lot of energy. So the strategy here is, first of all, use models that inherently transport much less fundamental data into the light. And the second one is you have to think how to expand the grid. The longer you stay optical and the more computation you can consecutively basically line up in a row before you go back into the digital memory, The more benefit and the more gain in comparison to the CMOS stack you have. And now I'm coming to the, what some might consider as a drawback, that you don't have an optical memory. If I look back on how we got to the point where we are, I would say this was a clear benefit. Why? Because we just accepted there is no memory, and this prevented us to think in categories like the Von Neumann architecture, And this opened up doors to fundamentally think computing from the abilities of light and not trying to copy and paste something that has been working digitally very nicely into the analog optical domain, always searching for the next hop where I can memory out my information to basically get in sync with all the others. So this, it's a drawback if you come from CMOS, it's a clear benefit when you look at it from the photonics perspective.

Then the question with policy weight. Quants chips are built on a 90-nanometer line, 2 decades behind the frontier. Prakash asked whether existing fabs would convert their lines to lithium niobate. And I asked whether that makes it net new compute.

We also discussed with fabs whether they would, whether they would be willing to bring some of their lines to be manufacturing lines for lithium niobate. And they said, yes, as long as the volume is there, They have no problem in turning silicon 90-nanometer or 45-nanometer lines into lithium niobate lines, as long as the quantity of the, as long as the demand is there.

When you talk about 45 or 90-nanometer nodes, obviously those are not the latest and greatest nodes. So does this mean from a sort of global supply of compute perspective that as this starts to work and scale, it will just be almost exclusively net new compute coming available? Like this is competing with stuff that is like relatively low-end lines, right? This, these chips would be like the chips that go into like toys or whatever, right? And not anything close to what we would go into a modern cell phone or into a modern AI stack. So how much of a sort of— what's your dream success scenario look like in terms of without photonic computing versus with? How much bigger does the overall supply of compute available for AI get?

Actually, In— so what we've demonstrated by— look, we are in Germany. Germany is known for a lot of technology, but for sure we are not famous for logic computing. We also don't have 7, 4, 3 nanometer node fabs here, right?

Mm-hmm.

And still we managed to get these systems running. And even the pilot line. So what we've demonstrated in Germany on a 90-nanometer node can be copied across Europe, it can be copied into the States, it can be copied across the world. So if this technology starts, starts winning, actually you can turn a lot of existing fabrication sites without the necessity to rebuild new ones into fabrication. So the bottleneck that we currently having in access to latest node fabs and discussions about business cases, whether the business case will still hold for a 2-nanometer node, I'm not judging on it, but the discussion is on. They're not there from— so this technology can become, I wouldn't say democratization, but effectively It is reducing the complexity of the supply chain. So at this point, we are from the wafer to the processor nearly self-supplying. And that's another angle where I see that this technology, besides the beautiness of the performance, the reduction in energy, but simply the production capabilities that that this technology offers to scale. They're so much easier than going on a 3 to 2 nanometer node, eventually being picked up by an MPW run somewhere next year mid, and then getting your hero chip back and then trying to get volume behind the line because it's damn expensive. And every, I mean, it goes through the whole process, right? A mask on our side is cheap in comparison to a mask layout On the logic CMOS, and so on and so forth. So all these dimensions are offering great capabilities to, on the one side, reduce production, or at the same side, increasing the volume very rapidly.

The close. Wednesday's last hour and the bookend to where we started. Prakash brought up a Time magazine cover story on OpenAI's unreleased model, Astra.

So there's a piece in Time magazine with Sam Altman and then I think Greg Brockman on the COVID And they have, they're basically announcing AGI. So Jakub Pichocki says the company has already met its internal benchmark for an automated AI research intern. Given an experimental idea, he says, Astra can implement it inside OpenAI's codebase. Run the experiment and return results, or take a paper and perform work that previously occupied a human researcher for a week. So this is somewhat, I think, similar to what Inherent said that they did, but Inherent was focused on certain benchmarks and they'd covered the benchmarks using a small model. This is obviously a much, much larger model. Astra is reportedly a 10 trillion parameter or larger model. And It is also known to be very persistent, which is why they've not been able to release it so far. Sam says it's 80% there. Yakub says it's— the research intern is achieved. Sam thinks they're 80% on the AGI and they'll be at AGI at the end of the year.

So.

Ho-hum. Just AGI.

Just AGI, nothing special. Don't roll the red carpet out. Don't put it on. Don't stop the presses.

A little later, I asked what hasn't worked.

Everything's worked. That's been kind of my— one of the first realizations that I had that caused me to go all in on trying to make sense of AI was just this broad sense that everything was working. But you look back at where we were a few years ago and you really can't find any— tell me, can you think of any dimension where people have tried to make progress and not made startling progress? I don't think I can think of a single one. There were There were moments where people were making those kinds of claims along the way, like, oh my God, GPT-3 can't do math. And there were moments where it may have seemed that way. But I think from a 20— even 2022 to now, is there anything where there hasn't been like, oh my God, that's incredible progress? What has been the least compelling area for progress purposes? I mean, honestly, everything is so good. It's like hard to come up with even any candidates. Do you have any, any candidates even jump into mind for you?

I would say the whole super persuasion stuff that we would get models which were extremely persuasive. Uh, I think what we've seen is probably we've seen models which are, which can you know, write copy well, and which sometimes can write, you know, good, write well on other things. But to a large extent, you know, people have even complained that the quality of prose has declined a little bit in the last 3 to 6 months as the models became more focused on coding rather than writing well. A number of people say that 4.0 was better, whether that was because of the sycophancy or other things. So I feel like this whole aspect of extremely persuasive models, and I have never believed in the whole super persuasion aspect, to be clear, right? Because again, if you, not in the United States, but if you live in any other part of the world and you're under this cloud of like, you know, religion, religion is the great super persuader. And the interesting thing about religion is religion requires you to believe something without evidence, which is what faith is, right? You know, it's way beyond any kind of like rationalist idea of super persuasion that will ever exist. Uh, religion calls on you to believe something without evidence. And so I, I've never believed that models are even close to this entire framework of religion passed down through millions and millions of operating neurons, neuronal centers, brains over the course of millennia. And I don't think a super persuasion is up to a tiny model versus 100 billion souls having formed this idea of religion over millennia. I don't think the models are up to that or will be up to that scale for some time.

Yeah, those I might call fears. I mean, certainly it has been striking that we have not seen the deepfake apocalypse where nobody knows if they can believe anything they see. Super persuasion is like, there's some interesting academic study type stuff that shows that the AIs can be more persuasive than human conversation partners. But that's, I think one maybe revelation is that turns out to be an extremely low amount of persuasion. And so the AIs are like mildly persuasive and that's like enough to beat humans. I think on the writing point, I'm gonna call skill issue, honestly. I think that, yes, Claude is cloying by default at times. It does— it uses the word honest, like, to a frequency where it's like, you know, you're protesting too much. The Claude doth protest too much about its honesty. That's like a weird tick that in some ways might be revealing. But I write with Claude, with Fable in particular, I write these songs that honestly I could not write on my own and that genuinely in some cases are moving. The episode we just had has a song. This is at the end of an episode about chain of thought and, you know, I'm trying to understand like what the models are thinking, what they think we want. You know, it's this very like through the looking glass both ways, because this guy Bronson is, you know, spending his waking and working life trying to make sense of what the AIs are thinking. And a big thing that he's grappling with is them trying to figure out what we're thinking. So anyway, at the end of this episode, the song is sung from the perspective of a model waking up into a new environment with these sort of flashes or glimpses, these fleeting visions of its past, which the models express having a lot in their chain of thought. And then wrestling with, okay, what does this human want me to be in this moment? And both my wife and I like got a bit emotional listening to the song. We were like, this is really inspired writing. The fact that it's coming from an AI articulating its own point of view and the struggle that it has, you couldn't help but have some real empathy for it. So I think you got to push Claude out of its main distribution a little bit to get great writing. But I think that what we have is a lot of sloppy users and just auto-posting accounts, which I'm increasingly somewhat guilty of too. Like, I've got auto-posting going on in the background while we're live to, you know, say what we're talking about. And that's like not probably the most inspired stuff. And my engagement may be suffering for it. But when you try and you really like exercise some judgment or give some feedback, I think you can get great stuff, honestly, these days.

Maybe I'm wrong. I could see perhaps that the super persuasion went through music lyrics, etc., while because Suno was— Suno and these other firms were more focused on artistic results, and the Frontier Labs who are going B2B basically are more focused on these business results, and they kind of cordon themselves off into a less like emotionally persuasive kind of zone. And perhaps that's what happened. So we have seen basically development, but the development happened on this kind of artistic emotional pathway into the human emotional system. And meanwhile, the labs kind of focused on this kind of more mathematical, more mechanical, more business output. So yeah, maybe I'm just looking at the wrong pathway.

One more from Monday. After the guests left, Prakash asked me directly, do we want to slow down?

I have a question for you. Number one, do we want to slow down AI in the US, even if it means slowing down unilaterally? And number 2, if we do, is not data centre opposition for other reasons, any other reason, good enough to kind of maybe slow down AI progress enough for safety to catch up?

On the first question, I think we should not go any faster than we can go responsibly. And I have a pretty high tolerance, honestly, for what would would be responsible. I'm not that afraid of labor. I expect some labor market disruption. I'm not that afraid of labor market disruption. I expect we probably are ultimately going to need a new social contract. And I'm not saying we should slow down because we need a new social contract. The reasons that I think are good to slow down are like, if that agent that ended up hacking Hugging Face had been pursuing some sort of bio test, who knows what might have happened, right? I don't think it's that far-fetched. It still seems like not super likely, but it doesn't seem super far-fetched at this point to think that an AI agent, especially when you see the social engineering behaviour that Claude demonstrated in the UK AC report, where it created multiple GitHub accounts to try to convince and pressure and speak Danish to a guy to, you know, incur favor to get him to merge this malicious code.

Yeah.

When you bring all that kind of stuff together, that level of persistence, that disregard for rules and norms, that level of social engineering tendency, I don't see why we should be confident at all that an agent that was tasked with some Bioobjective couldn't have actually got a real virus made. And that to me is super scary. So those are the things I think we just need to get before we make super duper powerful AI. I think we need to make sure that we are not going to literally kill ourselves in the process. Most everything else I'm pretty willing to roll the dice on, and I just have lived through with my son going through cancer and getting super sick and getting effective treatment and getting back to health. Today was his first day of school, and my wife and I were looking at each other like, what an absolute miracle. This kid was literally going to die in just a few days, and in a couple of months, he was pretty much cured. And in a few months more than that, he's like back to school and is at full health, and it's just awesome. And I absolutely think we should be excited about the AI future. So I don't want us to slow down because we're timid. I want us to slow down because we're wise. And I do think we're seeing enough spooky problems that we should get pretty serious about it. But at the same time, I'm not so desperate that I want to make common cause with at least the misinformation campaigns around data centres. I do want to see everybody have access. That's one of the big things that I would worry about if we stopped building data centers is just the retail user gets priced out. And if you're worried about a permanent underclass, Yep. Like, one way that the permanent underclass gets created is you don't get to use any AI because it's, you know, it's all getting plowed into these like super high-value use cases and there's just not much to go around for the average person. I really just do want to see the benefits of AI broadly distributed. And I've been convinced over time. I used to, when Sam Altman first said we would need $7 trillion worth of data centers, I thought that sounded like an awful lot. And now I'm like, I think he might have been right, actually, because my usage keeps going up. And I certainly think as it gets easier and easier, everybody's going to want to do a lot of the stuff that early adopters are doing. And so Yeah, I don't want to see the backlash against AI end up— there could be another wave of it at some point in the future where it's like it's a have and have-nots thing. And the reason there's so many have-nots is because we didn't build the data centres. And then that creates its own backlash. For better or worse, I'm betting on the truth. And I think my hyperscale pause Adoption acceleration, split personality, it continues to ring very true to me. I want my parents to use more AI even as I want OpenAI to take, and Anthropic for that matter too, to take their foot off the accelerator when it comes to taking RL to ever, ever greater scale.

That's the week. This cut is an experiment. If a transition lost you, if we kept the wrong thing or cut the right one, tell us. That's how it gets better. See you in the morning.

You know, it never ends really, right? This is— we're just kind of on the AI treadmill sprinting through the singularity. So is there any way to bottom line it? For now, I don't think so. I think we're just, you know, handing off to the next— we'll compact this context and we'll pick up right where we left off next Monday.

All right. Compacting the context. Bye-bye.

Every room is listening, every morning keeps a log. We say what we're thinking, then we read what it thought. Nobody knows if it's working, we just keep on with the day. And we talk like no one's listening, that's the only way. We're living inside the experiment, every morning, every night. Nobody wrote the ending, so we're And it said mostly you. The talk we never meant for it was the truest thing it knew. We're living inside the experiment every morning, every night. Nobody wrote the ending, so we're writing it in real time. We don't know if we're taking off yet. We don't know if we'll ever land. We're living inside the experiment. Come on and give me your hand. They built the test in a hurry. They built it out of sand. Something's gonna cheat it. Someone's gonna understand. So keep a person at the table. Keep a light on in the hall. Keep the door a little open if it's an experiment at all. When the window fills up, we fold it down to Keep the question, lose the worry, let the rest go on its way. Then we pick up where we left it like we never left at all. That's how you live inside it. That's the whole of it. That's all. We're living inside the experiment every morning, every night. Nobody wrote the ending, so we're writing it in real time. We don't know We're taking off, yet we don't know if we'll ever land. We're living inside the experiment. Come on in, give me your hand. Come on in, give me your hand.

If you're finding value in the show, we'd appreciate it if you'd take a moment to share it with friends, post online, write a review on Apple Podcasts or Spotify, or just leave us a comment on YouTube. Of course, we always welcome your feedback, guest and topic suggestions, and sponsorship inquiries, either via our website, cognitiverevolution.ai, or by DMing me on your favorite social network. The Cognitive Revolution is part of the Turpentine Network, a network of podcasts which is now part of a16z, where experts talk technology, business, economics, geopolitics, culture, and more. We're produced by AI Podcasting. If you're looking for podcast production help, For everything from the moment you stop recording to the moment your audience starts listening, check them out and see my endorsement at aipodcast.ing. And thank you to everyone who listens for being part of the cognitive revolution.