AI:AM: A Level We Shouldn't Pass? Notes from The Curve + Tokens vs. Salaries & Is SaaS Cooked?

This week, my notes from The Curve, the conference where Frontier Lab insiders and their critics share a room, plus the memory wall and what agents are doing to software. This is AI in the AM.

There likely is, let's say, a level of intelligence that we just shouldn't go past.

What?

Yeah, first time I'd ever heard that from a Frontier Lab leader, then I just asked a follow-up like, okay, well, how do we operationalize that? Would you be prepared to sign on to something like a limit on the number of flops that go into the next pre-training run? But what actually happened was he basically said, yeah, I think that could be reasonable.

Prakash Narayanan, my co-host, on what a ceiling would actually require.

When we say that we shouldn't go exceed a point of intelligence, that automatically means that you have to look at the inputs going in.

Thomas Somers, co-founder of Positron, which builds AI inference chips around memory.

I was looking at a quote that we had a year and a week ago, and for memory, and it's gone up 4.5 times since then. I really hope that this isn't the case, but I wouldn't be surprised if this is going to end up being another doubling over the next year.

Sean Wang, known as Swyx, who runs Latent Space and the AI Engineer Conferences.

And the phrase I've sort of recently taken to is, I don't want to pay for someone else to go through LM psychosis. I can pay for my own psychosis. That's fine. But when you work for me and I'm paying your tokens, you better be actually producing thoughtful stuff. And I have 2 or 3 employees right now who are basically under performance review because they are just giving me Claude slop.

Yeah, I pity our Claude that's going to have to try to cut this down into a highlights episode, man. I think this is Banger content from start to finish today.

Welcome to the AI in the AM weekly highlights. The narration is my cloned voice. Please tell us what worked and what didn't so we can make the next one better. Part 1, notes from The Curve. Last weekend I was at The Curve at Lighthaven in Berkeley.

Prakash asked about me laying out the divides I saw. One of the big observations in general in AI is that events have come pretty fast. And you would think if you were looking ahead a couple of years ago, that with 2 years of additional information and revelations and actually seeing how strong the AIs get and what they do and who's been right or wrong, you would think that people would start to agree on the state of things. And that has happened a lot less in the broader world than I would have expected. But I did feel that there was some of that felt at the Curve. But I do think the big thing that people have kind of come together on, at least to a significant degree, is it does look like the AIs are going to get really powerful, and this is something we're really going to have to reckon with and can't just sort of hope goes away or feel like maybe if we just wait, the bubble will burst or the fever will break or what have you. This time around, people were definitely like, okay, it's getting pretty serious. There were very few perspectives that, as Helen Toner once famously put it, long timelines have got super short. I would say another corollary to that was like, what is called normal in terms of expectations is getting increasingly Pretty incredible. Even, I would say, the people there who had the most modest expectations for what AIs will do, be able to do, were like, yeah, it's going to do an awful lot. And the question now on the capabilities front has gone to, like, is RSI going to foom? Is it going to be fast for a while and then kind of level off? Is that more than it's meant to be? Like, yeah, we've got these super powerful things, but is it really going to kind of run away from us? Those were still the capabilities questions that were getting asked, but there has been, I think, some healthy updating across the community just based on what we've seen.

Prakash asked whether anyone from the Frontier Labs said limits should be enforced on the players in this space, themselves included.

Yeah, that was a huge topic. I mean, and probably the biggest divide in terms of perspective is really just inside the frontier companies, outside the frontier companies, right? Everybody knows that they're living in the future and they have visibility that the rest of us don't have. And so actually had a chance to talk to and hear from, in some cases in sessions, in some cases just in passing conversations, all Chatham House rules, who I'll abstract away. But we're talking people who are like founders, executives, top researchers at frontier companies, and a small set of them. I asked a couple of them, why come here? What's in it for you? And before doing a couple of sessions where I was interviewing or moderating, I asked the same question. What makes this next hour a good use of your time? And they repeatedly expressed that it's just important to them to share as much as they can with the broader community. Definitely not sharing their trade secrets, of course, still, but really trying to give us a candid take on what are they seeing internally, what do they believe internally, what's their outlook, what's the plan, how can we come together and potentially do some of these constraints? And there were a number of things that I actually thought were kind of newsworthy. One, let's say, senior executive at a frontier lab First of all, was saying that just pretraining continues to deliver, that the models are just at the base level of capability, getting stronger and stronger, and there doesn't seem to be an obvious stopping point to that. Scaling laws are holding, or maybe even bending, because the data quality is getting better. They've got increasingly different architecture tricks, lots of work on synthetic data. Synthetic data seems to be definitely really delivering. You basically need more pretraining data. What do you do? You give the model a bunch of data and think, what could I transform this data into that if I then pretrained on it would make the model smarter? And you're sort of converting these reasoning tokens, your test time compute, you're converting back into pretraining data, which then goes in, of course, to the next generation of models. So this is like one way in which recursive self-improvement is happening. Not even necessarily— there are, of course, opportunities for architectural improvements as well, but a big emphasis was just on data quality. We can now just spend tokens on the existing data to augment it, to create better, cleaner, different lenses on it. We feed all that back in and the models get better. So this one executive was really emphasising pre-training continues to deliver. The models are getting super powerful. RL is important too, but his take actually went as far as he believes that the latest frontier models do have the ineffable research taste that would be needed for them to start making paradigm-level breakthroughs. He believes it's currently an elicitation problem where the RL is not bringing it out well enough. They don't know quite how to bring it out well enough, and And so it's kind of rare. And this is why you need 10,000 agents to do a Millennium Prize problem. And of course, that's verifiable too. So RL is working better there. Maybe you need a million agents at this point to stumble onto something eventually where you'd get the kind of paradigmatic change at an ML research level. But he believes that is there and they'll figure out how to elicit it. And then he said, A couple of things that I thought were just legitimately newsworthy. One was he believes that there is a level of intelligence, or there is, might be a little strong, but there likely is, let's say, a level of intelligence that we just shouldn't go past.

What?

Yeah.

First time I'd ever heard that from a frontier lab leader. And it was not like minced words. It wasn't like, there's definitely a level, or I can articulate what that level is. But this person who, again, everybody would recognise their name and position, saying, I think it is likely that there's a level that we shouldn't go past. Now, was that not ever or just not right now? I think that was also a little bit ambiguous, but it was a pretty firm statement. Definitely the strongest statement I've heard about imposing kind of a hard cap on capabilities, even for a time.

What confuses me about that, that statement is that some of the increases in intelligence that we're seeing, number one, they're coming as a result of compute increases in compute, which has traditionally been what has happened in the last decade and a half. Number two, they are coming as a result of refining the input data. Like the input data is being refined by greater intelligences. And so you have input data, which is much better. When we say that we shouldn't go exceed a point of intelligence, that automatically means that you have to look at the inputs going in. And the inputs going in are going to be compute and this refinement of data. And so if you say that we shouldn't go beyond that, that point, that also means that you shouldn't build compute beyond that point, or you shouldn't refine data further beyond that point, which means a lot really. It means like the end of the end of the compute buildout, which is pretty significant, right?

Yeah, well, I think there's a lot of uses for compute. So I had a chance to follow up on this. This combination of pre-training is really delivering and scaling laws, if anything, maybe are bending more favourably. Although it's tricky to know what you count, right? Do you count all those FLOPS that you use to create the synthetic data?

Maybe you should.

How to count these things is always super vexing. But between this pre-training thing and there might be a cap on how far we should go, then I just asked a follow-up, like, okay, well, how do we operationalize that? Would you be prepared to sign on to something like a limit on the number of flops that go into the next pre-training run? And I don't know what that number would be, and again, what exactly would count, but should we say no more than 10 to the 27 flops in the next pre-train or something like that? And I kind of anticipated that being a bit of like a straw man for him to react to or tell me what was bad about that and maybe ideally give me a better idea. But what actually happened was he basically said, yeah, I think that could be reasonable. The end. Not really much of a fight there at all. Again, the details will be incredibly important if we go down that path. Because the more data you enrich, the more flops you put into enriching the data. You can definitely play a sort of shell game of hide the compute, I think, these days. But the sentiment was like, yeah, we might have to do something like that. And then, of course, your question I can anticipate. I know you well enough to predict the next Prakash token at this point is, what about Elon? What about Zuck? And the answer there, there were a couple of really interesting things people said along the way. One, this is not from one individual, but sort of a takeaway from the event in general is that the, let's say, 2 to 3, Gemini 4 being announced now, we maybe have a 3-horse, very frontier race again. We'll see as that comes online. But the takeaway was kind of the 2 to 3 companies are really pushing the frontier. And even the other 2 that we would usually think of in the top 5, Elon and Zuck, are actually distilling a lot, whether they know it or not. And the way that they are distilling these days, it's not that they're calling the Claude API and getting traces and feeding them in, but rather the cottage industry of RL environment makers and sellers is in fact functioning as a distillation conduit, because what they're all doing is having Claude create these RL environments, selling the RL environments to other frontier model makers, and then they do their internal RL on an environment that only exists because Claude was smart enough to make it. And that's how they are kind of constantly capturing the capabilities advances that the true frontier companies had. So there wasn't actually a ton of talk about competition beyond the top 3. One person went as far as to say, and this was another founder executive type at one of these companies, yes, they have kept up. Obviously there's been a gap, and that gap has been persistent, but they've kind of maintained kind of consistent following distance. And this is in contrast to Anthropic's old prediction from their fundraising deck that I always think about when they said maybe in 2026-ish, the companies that train the best models will be so far ahead that nobody will be able to catch up. Obviously, that hasn't happened. But the kind of synthesis of all this is if they hadn't been releasing the models and if people weren't able to do these various sort of direct and indirect distillation techniques, they actually think that probably would have happened, that the other companies are not really keeping up, but for the sort of leak of intelligence in all these different ways, including the RL industry, that is gradually finding ways to transfer from frontier models to their competitors. Hey, we'll continue our interview in a moment after a word from our sponsors.

Today's episode is sponsored by Parallel, where agents find answers. Most engineers today closely follow new model releases, but don't pay nearly as much attention to their agents' most important tool: web search. If you're like me, your agents are often conducting hundreds of searches per day. And while the unit cost is small, over time, this does start to add up. Before starting to use Parallel, I calculated that my agent search bill would total roughly $500 this year. The good news is that I just had my agents conduct a systematic test, and I found that Parallel's fast mode, which costs just $1 per 1,000 queries, worked just as well as my previous default provider, and switching to it will save me roughly 80% of my search bill going forward. Parallel's infrastructure is enterprise-grade, and their suite of APIs offers a range of Pareto-optimal options that allow you to choose the right balance of quality, cost, and speed for your needs. So whether you're building voice agents that need 200-millisecond latency or long-horizon agents that need deep research, adding Parallel to your agent's toolkit is a no-brainer. Get started for free at parallel.ai/tcr. That's parallel.ai/tcr. 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 slash TCR.

Prakash added his read on Elon's strategy, and then I got to timelines.

And I think on Elon's side, Elon is very focused on the infra build out. So Elon is making a gamble that the infra matters more than the models, and that he will be able to. catch up basically using his infra buildout. And so, and he's, and he's structuring things for 2029, 2030, 2031, like TeraFab, things like that will come out much later. Uh, the SpaceX satellites, that's like a 2030s thing. So he's, he's structuring for a much longer period of time. And I think Anthropic and OpeningAI are both structuring for more like the 2 to 3 year mark. And they're looking at a 2028, 2029 AGI RSI.

Yeah, honestly, even sooner, I would say, was kind of the vibe. And this is probably somewhat biased to be a short timelines crowd, but again, the short timelines crowd includes the executives, top researchers, founders of these companies. They were talking more about just next year.

Yeah.

It was really a kind of 2020 critical period is beginning now. RSI is kind of at hand. Definitely not full agreement on how fast it goes or if or when it tops out or whatever, but a shared sense that the decisions that get made over the coming months and the way that compute is used over the next year could be really critical. There wasn't really even too much 2028 talk, to be honest, which is again, pretty crazy.

Then it came back to compute and what a cap would mean for the buildout.

But in terms of will they still have need for compute, or does a cap on pretraining scale or other compute input cap that might constitute a pacing mechanism, does that mean the end of the compute buildout? I would say no. Another question I had the chance to ask was, of course, everybody's heard Jensen on Ezra Klein, I think, at this point. And Comment that he made was like, these companies are going to have to mature. They're going to have to change their ratio of where they spend their resources. He said, at Nvidia, we spend like 20% of our effort on designing the chip, and we spend 80% of the effort on verifying and validating and testing all the edge cases and making sure it's going to last a long time and be reliable and all these other things that are not just the original design. So his point was like, they've been working really hard, obviously, until this point, and they've been putting everything they can into making their models capable enough to be useful. Congratulations, you've succeeded. Now you're going to enter the era where you're going to have to make them safe, reliable, trustworthy, et cetera. And so I had a chance to ask one of these people, what do you think about that? take from Jensen. And the response was basically like, yeah, that's kind of reasonable. Without, again, knowing exactly where the numbers will shake out, the possibility that the majority of compute in the future could be going into all sorts of different safety measures, which could be monitoring, could be chain of thought monitoring. People are getting kind of bearish on chain of thought monitoring. But then there's internal activation monitoring. They're already spending a significant percent of compute on monitoring, and it sounds like that might go up a lot. Another thing that they're spending compute on in a big way right now is fixing the RL environments. Of course, everybody understands at this point that if you have sloppy RL environments that reward cheating, then you're going to get a lot of cheating. They are now applying these models to the RL environments themselves and saying, hack this environment. And now that is the direct task. They're doing it. They're finding all these places and ways in which the environments can be hacked. They're fixing those. I'm sure they're throwing some environments out that are just fundamentally flawed, whatever. And gradually they're driving down the rate of flaws, which means that they're also driving down the rate at which cheating would be rewarded. Which in turn means you're going to get less cheating from the models as they actually come online. And I got the sense, although I think this is still, as they say, an open research question, but I got the sense that they are seeing a pretty clear relationship between the more you clean up the RL environments, you do in fact get lower rates of cheating downstream. And They're trying to push. Obviously, they'd love to have zero hackable environments. That's going to be tough. But it doesn't seem like we're all the way to the point where we have a scaling law, but we're seeing sort of a proto scaling law, it seems like, in, again, the way you can spend a lot of compute driving this down and getting better behaviour on the other end. I heard repeatedly that they expect to be bottlenecked on safety and alignment. But it seems like for at least the next generation or 2, there is still enough low-hanging fruit to pick in the safety and alignment domain, in part just by fixing RL environments, that we should expect things to continue to move pretty fast. But it seems like they're probably going to be able to get comfortable releasing the next couple of things, even as they do consider themselves bottlenecked on alignment. But then past that, they're like, yeah, all bets are off. Really hard to say what happens kind of past the next 1 to 2 generations.

Later in the show, I tied that executive's view on research taste to a new record in the NanoGPT speedrun, the race to train a small language model to a target quality as fast as possible. The record comes from a company called Hyperstition.

But just headline-wise, it seems like we've gone from something like 70 seconds on this classic benchmark. This is something people have been working on for quite a while. And basically the idea is to train a small model to get to a certain loss as fast as you can in terms of wall clock time. And this result here took almost half the time off, which was more than I think the last 45 improvements combined. And this one has been accepted. One thing I did think was quite interesting about this particular one that has been accepted and kind of seems to have kicked off this wave of significant drops in the time was that the author here said, AI didn't play a big part in it. Like the person, senior executive leader at the frontier company that I mentioned earlier, their perspective was they believe that the model probably can come up with these quality of ideas, but it's an elicitation problem and they didn't really rely on AI super heavily in this case. Obviously they're getting lots of help, but in terms of core insight, it was mostly attributed to the human, not so much to the AI on this particular drop. This company is doing also large-scale training. So they did also make a point in their publication about this, that the optimizer that accounted for A significant, but I think not majority of the time savings that they did share is actually not as good as the one that they are using internally.

Then one last expectation I heard across the frontier companies at the curve about science.

AI for science is very much where they think we are headed in the not distant future. One sentiment was, are AIs going to be superhuman at just some things, like the verifiable tasks, or are they going to be superhuman at everything? And the sort of middle position, which I think is very credible, is we are seeing superhuman performance at anything we care about and are really committed to Investing in. So that doesn't mean it's full generalisation to every domain, but even in domains that are thought of as not super inherently verifiable, they feel like when they put their focus on it and they go license what data they may need to license, and they, of course, apply lots of processing to that data to augment and create synthetic versions and what have you, And put some RL on it too. The whole package basically is, I think, broadly understood to work on essentially any problem that they really choose to focus on. And so AI for Science is coming next, and the expectation is we will see— one person went as far as to say, In no more than a year, it will not be possible to do the very best science without AIs playing a big role in it.

The next morning, I added one more takeaway from the curve.

I don't think we talked about this yesterday, but a takeaway from the curve this weekend, also one of the not main threads that I was involved in discussing, but a general vibe, was there is going to be government intervention. And obviously we've already seen some. But the idea that Congress will never act, I think, was— that used to be sort of taken for granted. And now the new vibe that I was hearing was like, after this election, especially if the Democrats take both houses, both chambers, That may change, and you might even see a bunch of Republicans join the Democrats because really, in their heart of hearts, they want to do something. They don't really want to go against the president, and there hasn't been a lot of people willing to take that chance yet. And they're also getting a ton of money from the data centre interests right now. But that all could come up for a renegotiation and realignment after the midterm. And we might actually see congressional action a lot sooner than maybe a lot of people who've grown accustomed to the idea that it can never happen or it'll be stasis forever might think. So I do think policy response is definitely part of what we need to be looking out for. Hey, we'll continue our interview in a moment after a word from our sponsors.

This episode of The Cognitive Revolution is brought to you by OutSystems. The leading agentic systems platform. OutSystems is helping their customers build, modernize, and operate enterprise systems starting from any coding tool in a governed agentic engineering model. With OutSystems, you can coordinate and govern your entire agentic workforce within your enterprise ecosystem. Accelerate impact with industry-proven agentic solutions that are pre-built, governed, and customizable to your business. It's time to innovate at the speed of AI without compromising quality or control, which is why thousands of enterprises worldwide trust OutSystems for their mission-critical apps. Teams any size can use OutSystems to build, deploy, and manage AI apps and agents quickly and cost-effectively without compromising reliability and security. With OutSystems, you can accelerate ideas from concept to completion. It's the leading agentic systems platform that is unified, agile, and enterprise-proven, allowing you to build your agentic future with AI solutions deeply integrated into your architecture. OutSystems, own your agentic future. Learn more at outsystems.com/tcr. That's outsystems.com/tcr. The Cognitive Revolution is brought to you and powered by Tasklet. I've been running Tasklit agents for almost a year, and they've become a major part of how the show gets made. Before every recording, one agent researches the guest and sends me a suggested outline of questions, complete with links to all sources. Meanwhile, another agent scores the overwhelming number of pitches that I'm getting from PR firms. And that's really the core idea behind Tasklit. Run your business on agents by giving them real jobs to do, on an ongoing basis. Setup is incredibly easy. Just describe the work in plain English: research a new sales lead, update the CRM, and draft a personalized reply. Tasklet connects to your business tools and can run around the clock, on a schedule, or when something happens, like a new email arrives. Each agent works with your tools and business context and remembers useful information across threads. You can also share agents, connections, and reusable skills with your team. So the workflows that one person figures out become something the whole company can use. No developer needed. You decide what agents can do on their own and what needs your approval. And there are no per-seat fees. You pay for usage, not for adding teammates. Start with one recurring job, like researching the people you're meeting. And then just keep building from there. Visit tasklet.ai and use code COGREV for $50 in free credits. That's code COGREV at tasklet.ai.

Part 2, The Memory Wall.

Thomas Sommers is a co-founder of Positron, which builds chips for running AI models and recently raised $875 million. Prakash asked whether the sharp rise in memory prices had changed the economics.

I was looking at a quote that we had a year and a week ago for memory, and it's gone up 4.5 times since then. I really hope that this isn't the case, but I wouldn't be surprised if this is going to end up being another doubling over the next year. But I would say a big reason why everyone has burdened these cost increases is because you're deriving more than that 4x or 5x value compared to last year. But I think that the reality there is, compared to a year ago, the capability of a model using X amount of gigabytes of memory is way more than 5x than it was this time last year.

I asked Tom about the trade-offs of building on commodity memory instead of the high-end kind.

At the start of Positron, I did think that memory was going to be the limiting factor for things going forward, even though You know, people can debate on power and, and other parts of the infrastructure, infrastructure stack. My fundamental belief was that, um, there isn't going to be a stop anytime in the near future in terms of scaling laws and getting increased value from making models larger. So that eats into memory on one side. But then the other side, I think the main limiter today of AI being applied to most applications is context length and, uh, uh, being able to hold more context, uh, you know, per user and then scaling that out to drastically more users. And, and really today I would say that the driver of quote unquote users or individual sessions is actually just having more agents. I've gone from, you know, 3, 4 months ago where I had probably on average 2 to 4 agents like running constantly in the, in the background to where I'm now running 15 to 20. If you then multiply how many people are using AI and how many that, that concurrent completely separate contexts adds up very quickly. That memory driver had us say, okay, we have to use the commodity memory 'cause that's the only thing that's going to be able to scale and be cost effective. And when we set that as sort of the constraint in our architectural design, we had to come up with, you know, very clever innovative solutions. And, you know, the 2 main pieces of that, and, you know, the thing that we showed with our first generation product is being able to achieve extremely high memory bandwidth utilization. From a compute architecture perspective. So while NVIDIA GPUs, I would say on average, are getting between 30 and 40% memory bandwidth utilization in the actual decode forward pass of a transformer model. So basically that means even though they advertise, you know, 8 terabytes per second of theoretical memory bandwidth with, you know, a B300, you only actually are seeing, you know, something in the ballpark of like 2 terabytes per second of actual realized memory bandwidth. And that comes down to a whole bunch of GPU architecture details. It comes down to the reuse patterns that really don't exist in transformers that they designed their hardware architecture around for training and other workloads. And so with our first-gen product, we were able to hit and sustain 93% of theoretical memory bandwidth. So that theoretical and realized basically, you know, become one. And so that isn't that massive, you know, 3x improvement right there. But to really take advantage of, you know, the belief that models and everything are just going to get larger, we had to be able to scale to massive more channels. And so we're, you know, partners with Credo Semiconductor that, and have developed a memory chiplet solution that allows us to go from the maximum number of LPDDR, same type of memory that is in phones, laptops, et cetera. The max number of channels that you can find in any other products is on the order of, you know, 12 to 16 channels. And we go up to 72 channels of LPDDR5X with this decoupled memory chip solution.

Positron's next chip is called Asimov. Prakash asked, what disappears in networking and coordination? When a whole model fits on one chip.

We have up to 2.3 terabytes of memory capacity per chip. So when you compare the, the B300 shipping today, caps out at 288 gigabytes. As, you know, reported in, in by analysts, NVIDIA for the next generation is actually cutting down the memory capacity because costs and everything involved. So they're only going to be at 192 gigabytes with the Rubin Ultra. And so, but even if you use that 288 as like the current benchmark, we've got 8 times more memory capacity per chip. And so that means on a per-chip basis, you now can actually scale in what you would have needed from a memory perspective 8 GPUs for, we can do on a single device. And it's not just, oh, you've got that cost silicon savings. You know, as you pointed out, whenever you have to scale to having more than one device, there's overheads involved with that. There's, your all-gathers, all-reduces that you have to do within, for each layer, and each, really each multiply that you're doing of sharding across those devices.

Then I asked Tom about looping, where a model runs the same layers more than once for each token, and why that is tempting at all.

It's been very interesting from the Loop Transformers paper to that being the rumour of one of the big advancements with GPT-6 Astra. Simply repeating the forward pass, uh, gets you that improvement. But it's crazy that there is really work that I think was only being done in sort of the fringes of the open source, uh, you know, transformer community, you know, 2, 3 years ago of taking like a LLaMA 70B and just duplicating layers within it. So it was taking like a 70B model and slightly strategically saying, okay, I'm going to just repeat these, these sets of matballs.

Mm-hmm.

in sequence and turning it into like a 100 billion parameter model. And then you got better results from it, even though it was repeating the exact same, uh, math model. So there is an element—

Imagine what you can do when you train it to work that way.

Exactly. If you already have very high confidence on what the next token is going to be, and multiple layers have agreed on that, decide to early exit and go out. Or if you're really uncertain about something, actually find which layers in network Actually will have a higher likelihood of being able to increase that probability. Same sort of like functions that the MoD router does in terms of determining which experts should be selected for on a per-token basis. If you extend that concept to, okay, if you're really unsure, if there's a, a very large set of, of equally weighted probabilities for what the next token should be, and you're 3 quarters of the way through, through the layer, Do you actually just want to route back and determine, okay, we actually need a different set of experts, something like for this. But I wouldn't be surprised if these things are already being done at scale in the major model labs.

But can you go just a little bit farther into the hardware connection of that? Because I'm kind of coming in with assumptions along the lines of just very fundamental terms, like why loop at all? I think it has to do with this kind of memory bandwidth, right? Because Now I can keep the same weights on the chip and I don't have to like shuffle things in and out as much.

No, not really, because within— when basically the assumption is if you're doing inference that you're going to have all of your weights locally in, in DRAM, you're just sharding that over some number of devices. You do get some amortization, but just with the size of experts and the sizes of caches on chips, There's not much of a reuse opportunity. You're, by the time you're already done with the layer, you've already gone through the amount of memory many times that of what is in like the on-chip SRAMs. From, from like a hardware perspective, like looping really is more memory capacity savings than memory bandwidth savings. Because like, let's say that you trained 2 models with the same base set and you trained one to do with looping. And so it only has 10 layers instead of 20 layers. And, you know, model B is 20 layers. That 20-layer model, it's getting very, very simple, is going to be double the size of the 10-layer version. And if you find that the looped 10-layer version gets 95% of the same quality of results and it's half the size, you're probably going to deploy that one. You're saving more capacity than bandwidth because doing the second loop, you still have to do the same memory fetches and do all of the same Like, same number of NAPLs. There's the exact same number of bytes that need to be moved and the exact same number of flops that have to be done in that Model A and Model B scenario. It's just that Model B, the 20 layers, actually has unique different weights for the second group of 10 layers that have to be done. And so that means you require twice the memory capacity.

Then I turned to how AI is changing chip development itself.

In a notable interview Jensen Huang recently gave with Ezra Klein, he talked about how NVIDIA spends something like 20% of its time and energy designing and then 80% verifying, validating, you know, ensuring long lifespan reliability, et cetera, et cetera. What do your ratios look like on that?

So for us, I would say it's a little bit different because we're starting from scratch. So there's a lot more like base layer, base capability design that we have to start from scratch. So for, you know, this, our first full custom silicon, Ozemov, it was about a— it will end up being somewhere around maybe 60% design focus for us to 40% verification. I would say like the probably most amazing thing that I don't think would've been possible 3 or 6 months ago, and Only really became possible with GPT-6 Astra and now Opus 5.5 that we're leveraging a lot in this task is we, for doing our verification work, we use Cadence Palladium emulator systems. These are big giant racks that are full of custom ASICs that are specifically designed to just emulate, do gate-level emulation of the silicon. The crazy thing about Aster and Opus 5.5 is the fact that the agents themselves have been able to completely do closed-loop iteration and testing on this. So I think people thought it was crazy, but handing these very powerful agents full keys to our internal infrastructure and saying, here's a Palladium piece. And I highly doubt that these models had Palladium documentation in their pre-training thing, especially because a lot of documents are new as software updates. But, um, the fact that it went and looked up the documentation, when we're looking at these agent traces, uh, the reasoning traces and tool calls and et cetera, it's reading the full documentation, effectively compressing that itself by, by reading the PDFs, generating its own, you know, markdown files of cheat sheets of what to do, and then building out all of the testing infrastructure and its own harnesses for Being able to access this, it's mind-blowing. And so while if you're just doing regular software RTL simulation, that's running on the order of like 10 hertz for us in those cases. If you remove a lot of the debug pieces that slow things down, et cetera, we can run the full chip emulation on the order of 500 kilohertz. And so you've got actually several orders of magnitude speedup and improvement from that gate-level hardware piece. And this just gets us closer to this, this RSI loop. And right now that's really just focused on implementing test programs, finding cases where they fail, and then writing out reports which, you know, get reviewed by other agents and by humans in the loop. But I would say when this really started working 6 weeks ago, Roughly, it was beginning of September when, when Astra 6 came out. It was a huge, like, step function improvement from GPT-5.6, where it could not do that full closed loop itself.

It would—

it still required humans at different stages.

I asked how Positron's spending on AI compares with its spending on people.

So on, on the token spend side, it's become, you know, our single largest, uh, non-manufacturing line item very quickly.

Bigger than, bigger than human salaries?

Yeah.

Wow.

Just recently. And it's interesting, it eclipsed human salaries, it came back down. Um, I'll explain in a moment. So yeah, so our token spend has massively increased where, where, okay, so 6 months ago is equivalent to 1 employee. In, in, uh, June, it started to now be multiple employees. Like, we, we were peaking, you know, in the, in the days after GPT-6 Ashburn, we're really trying to, to push its boundaries, et cetera. We're, we're spending over $100,000 a day, um, on tokens. That tempered down a little bit in, in the weeks after as we optimized a bit and just weren't trying to run as many parallel experiments. But I would say the big thing that was like, you know, a saving grace was, you know, Opus 5.5 came out and in a lot of our tests, not everything, it was doing better than Astra and was, you know, a quarter of the price. We didn't really— we didn't tell anyone to cut their spending or do anything, even though we were seeing this exponential growth in our token spend. It was really just the fact that, um, uh, we both felt that we were getting good ROI on it, so we weren't going to taper it. But personally, like, my philosophy is I really don't want anyone to be running a real software or hardware development task on anything that is less than the best model. Like, I don't care if it's a 10th of the cost, you know, per token. It's, uh, just not worth the, the expense.

Prakash asked whether Tom has seen AI come up with genuinely unexpected ideas in chip design. Rather than following the rules in a manual?

I have not yet. I would say that's the most disappointing element of all. Like, I'm, uh, as you can probably tell, I think all this is insanely amazing, and I think we're going to continue this exponential trend. I'm somewhat proud of the fact that it has not figured out some of our cleverness of what we've done. Like, it's— it has questioned, and like, this seems like a bad design decision until It gets explained to it why something's done, or it actually runs tests and sees and understands, oh, that's why you're not doing the traditional way a systolic array is done. Do I think that will hold for another 6 months? Not sure. But just because I, in part why I actually think it's not going to be like the agent will come up with a better idea immediately. I think the amazing turn of events that we've just had in the past few weeks in terms of us having access to these models, is the fact that just because it can iterate so fast and do experiments and things by itself, I wouldn't be surprised if it would come to the same conclusions we did, or maybe, you know, make better things than what we would do by ourselves, just because it's going to be able to iterate so fast and go through so many different design possibilities. I think, like, just the story of OpenAI solving Navier-Stokes, Like the same approach of, yeah, if you take 10,000 agents and send them off for millions of man-years worth of effort, it found a solution. It's brute force, it's expensive, but I think that is a valid strategy that has only become possible in the very short past.

Not many people left standing in the AIs can't do what I do category, so congratulations. Congratulations on your place for as long as it lasts.

For a couple more weeks, I believe.

Part 3, Software After Agents. Sean Wang, known as Swyx, runs the Latent Space podcast and the AI Engineer conferences. I ask how software engineers are feeling right now. Empowered or threatened?

Yeah, I think college kids are a bit worried, but other than that, if you're relatively plugged in and very capable of AI engineering tooling, you're in more demand than you've ever been because your expected value is higher than it's ever been. So this is definitely one of those Jevons paradox type things where the sort of cost of software, creating software has gone down, therefore the demand has increased a lot. And for a very specific kind of demand for people who can manage coding agents productively instead of producing a whole bunch of slop. And I've been in that situation. So I am both an engineer but also an employer of engineers and people who are not engineers who are vibecoding. And the phrase I've sort of recently taken to is, I don't want to pay for someone else to go through LM psychosis. Right. I can pay for my own psychosis. That's fine. But when you work for me and I'm paying your tokens, you better be like actually producing thoughtful stuff. And I have 2 or 3 employees right now who are basically under performance review because they are just giving me Claude slop. And that's really bad for them. And they don't understand it. They don't see it. They're like, what do you mean? I think this is perfectly fine. And then I'm like, well, you're not producing any value. I can just prompt Claude. I don't need you.

I asked whether years of coding experience still matter when he hires. Or whether product sense and high standards matter more.

I don't, I don't care so much, except for, let's call it security roles and for anything sort of backend scalability. Like I just recorded an interview with the Supabase founders who, who, you know, are scaling Postgres to a scale that we've never seen before. Good luck trying to hire somebody to scale Postgres, you know, with a complete self-managed sharding solution that works at YouTube scale. There's only one person in the world that can do that, and they hired the guy. Like, you better go out and, you know, make sure you don't lose data, that you like don't have— don't cause downtime, that you scale economically, because basically agents consume databases something like at a 500-to-1 ratio of what human developers do. Then, yeah, so, so like other than that, like you can mostly write code, and then it's just Then it is about taste. And it's not so much about length of experience. Actually, sometimes length of experience works against you because you just have a set way of doing things that you don't understand how to do more than one agent at a time. Yeah, at this point, like, you should be relatively comfortable juggling like 5 to like 10 ongoing multiple things. And it's definitely true that you're much more of a manager than an individual contributor. I do think that people who look at data rather than code are more valuable these days. So being able to say, here's the logs, here's the traces, here's the schema, here's the input/output, capturing it, turning into an eval. All these are basically the merging of AI engineering and ML engineering that is happening that people need to upskill on. I think that the person that has, let's say, 10, 20 years of experience in software engineering, reviews every line and tries to make every line make sense. Whereas now we just need to make the modules make sense. And I can allow slop in there because it helps me to go faster, as long as I contain the slop in things where I completely understand the whole system. Where you go wrong is you have too many modules, too many black boxes where you don't even know. And actually the code also gets confused. So I'm making my own Slack competitor. I saw a bug where the messages weren't loading. I refreshed and the messages were loading. I refreshed again and the messages were not loading. It's like, well, it's the same exact code. What the hell's going on? Turns out there's 2 code paths and there's a race condition. Why? Because 2 different coding agents worked on it at different times. They just made their own thing. A human would never do that. And agents would potentially sometimes do it because sometimes just things fall out of the context window. But you need to have the oversight of the module and then whatever is inside of the module can be a black box.

Prakash asked, which parts of the stack are poorly designed for agents now that agents are becoming the main users of the internet?

Compute.

Everyone knows the GPU shortage, everyone knows the memory shortage, but CPU shortage is the most consistently reported thing among all the people that I interview right now. So what does that mean? It does mean that we need more fluid compute, which is the Vercel term for it, more different forms of serverlessness where it's very ephemeral. You can execute and resume because Agents need some time to do LLM calls or to make network calls, whatever, but basically long-running states that are resumable. But I think beyond that, you really start to get into 2 things, which I'm really thinking about, like sort of bandwidth or networking, where the latency starts to really matter. And that's an optimization of like, well, what point of presence are you communicating with? Or what region are you in? Versus like actually how much compute Should you be doing in cloud versus locally or edge devices? That is definitely happening a lot for the robotics companies that we're talking to. Basically, don't treat it as just a robotics problem, where the robotics are just the first harbinger of what you guys will eventually be doing if you consume or do enough inference at their scale. And then finally, to me, on a very, very personal basis, bandwidth. So like how many tokens per second can I get, uh, both on a single call basis and on just aggregate across all my calls and then aggregate across my company because I have so many, I have enough people managing all these calls.

Swyx has also been running Kill My SaaS, a bounty for anyone who could replace an events software subscription his company was unhappy paying for. Prakash asked him about it.

Basically, you can call this like a bounty on mid-tier SaaS that should not exist. Basically like half a salary or one-third of a salary for a piece of software that I don't own, that nobody enjoys using. If I'm going to spend $40,000 on this SaaS subscription, I can spend that in amount of tokens and that buys me a heck of a lot of tokens. We have so many submissions that we then have to eval them. And so evals are the problem. And because like, we're no longer checking like, here's like 2 or 3 requirements that you fit. It is, we're checking the entire UX of the entire flow from 3 different perspectives of, you know, organizer, attendee, and sponsor or a speaker. And just be aware that like, if you offer like a large bounty, like we offered $10,000 as the prize, you get a lot of submissions because people want to write code for fun. And a lot of them will be low quality. They would just be like, Claude, make no mistakes, go do this. And then it'll make mistakes and they'll just submit. And so the verification load is very imbalanced because they spent zero thought on this thing. They just threw it in there. And then you have no idea if it's low quality or not. But overall, very successful. And the team initially— so I have one of the most adversarial non-AI teams in AI. Okay, I hire— like, what is my business? I hire event professionals who are like super old school. They do everything in spreadsheets. They work with unions. They have to worry about like placements of like the meter boards physically in like locations, right? Like these are not glamorous, like high-tech jobs. Then they're very super suspicious of anything that is new and vibe-coded and techie, right? So I hire these guys and then I like make them work for AI, right?

That's my job.

And so Initially, they were like, we'll never use this bytecode thing. I want to use the tried and tested stuff that has been working for Microsoft and all these other guys. Then they saw the quality of the submissions, then they looked at their existing platform, they were like, yeah, okay, we're going to switch. It's like that initial hurdle, you basically need to provide the proof of this is feasible. Then the ongoing thing, which is And this is one of the benefits of me partnering with Cognition on, is I just give them, give all of them access to Devin to modify the code. So anything they don't like, they can just request a change and they have it pretty much in 1 to 2 hours, which they've never had before. Just to give you an idea of the way we work with these SaaS companies, right now when we request a change, they'll say, okay, that sounds pretty cool, it's on our Q3 roadmap. And like, we don't have confidence that it'll actually be done in Q3. And by the way, this is done. We requested for a thing in Q1, it just landed. So like SaaS is quite cooked if you're mostly a CRUD app. I would say that there's a lot of UX issues left. There's, you know, if you look at Sui Bench and you're like, oh wow, we're like 90 on Sui Bench, you're looking at the wrong thing, my guy. You know, if you haven't actually tried to completely vibe code a SaaS that you use, you don't understand how models are still very bad at all this.

Then I pressed Swyx on the downside.

But I have a hard time seeing how we don't kind of pick a lot of this low-hanging fruit and then end up in a sort of a lot of companies go out of business. And also I'm thinking like layoffs from like Meta, you know, that's another kind of canary in the coal mine. Like, are we going to see mass layoffs from big tech? How long can the party go on?

The simple answer is like, if you look at it as a fixed pie of software engineering, then that is the conclusion that you will arrive at. But if you look at it as the point of like, my competition or my TAM is spreadsheets, all spreadsheets made by anyone with any sort of productivity tooling can be turned into custom software for them to put on Rails with a beautiful UX and with like automations, or then the demand for custom software becomes A lot larger. And more beyond that, if it's all— if it's phone calls, if it's emails that go back and forth, and eventually physical meetings between people, all of that can be incrementally turned into more and more and more and more custom software and hardware and models, by the way. And so that basically, like, we're just growing into the long tail, right? Like, why is Salesforce so damn big? Is because people can customize Salesforce. But at some point, they need to stop paying the $300K a year baseline subscription and spend $30K building their own personal CRM, and they're good. In fact, they're more than good, they're happier because they can now modify it to whatever they want. Like, the sheer amount of customization that people can do, we don't even have the same amount of customization in our software as we do our handbags.

What the hell?

You know what I mean?

Imagine, like, clothes has been around longer than computers, but we have so much customization in clothes, so much demand in different brands and different varieties. We don't really care about benchmarks for clothes. It's just like, this is the style that I like. I think what software engineering has basically only done, when we basically have Kate Spade versus Louis Vuitton versus Coach for software. Like we're like not even at the point of like looking at features or like price. We're just like, what's the brand feel? And like, I identify more with this brand. Like that's what really sort of true commodity software is, which is you stop differentiating on substance and you start differentiating on vibe, which there is a lot of vibe right now. The reason I said, you know, this whole thing about like scaling my personal token bandwidth versus scaling my organization token bandwidth. The real limit on all this is the sum total of humanity's total token bandwidth, which is literally the amount of silicon we can produce. And that's going to dictate pricing and availability and rate limits and all these things. And so the real players in the room are only focused on that. And then the rest of us are just fighting for scraps within the fixed pie that we already have.

Prakash asked about Jev, Typesafe AI's fast decision-making model.

And whether OpenAI's new Decisions API announced at Dev Day is the same kind of thing.

Yeah. So not only have I used it, I was friends with Diogo for 4 years before the launch. And so I insisted that we were the first podcast that he did about Dev. And that did— I mean, that is now our number one podcast ever within space. And then secondly, we're also in the trial for the Decisions API with OpenAI. So we have a podcast on the Decisions API as well. Which is not the same thing. That one is basically a Loona with a different inference layer on top of it. I think what's also true among the people that I talk to are that a lot of the use cases you see on Twitter are probably overdone or dumb in the sense of you could use any other small model to do exactly this thing. But because Jeff is hot right now, you're just creating a bunch of content. Or the other perspective is this is another classifier. You can train classifiers all this whole time, but you can, but you know, now it's sexy. So like now everyone just people are throwing classifiers or using a classifier model where they would previously use an LLM because they didn't know anything else other than LLM. People are missing the whole point why they don't call it YJF or Diogo doesn't call it a decision model. They call it System 1 model because it's meant to be integrated inside of your software to go into the background of your software. As unremarkable as an if statement is the rough idea inside of your software. And so people aren't really doing that. They're like, let me just replace this, every point of structured output with Jev. And they're probably going to run into issues because they don't care about intelligence. If you notice most of these use cases, aren't really evaluating, like, well, does it actually play it very smart? Does it do any planning? None of that. It's just, can you make a decision quickly? And like, yeah, you could always make a decision quickly. I think there's some nuance there that's maybe being missed. Doesn't matter because I think the category creation was so successful that it's an overall good thing for the industry.

Before he left, I asked Swyx to plug his next AI engineer conference.

Yeah, it's next week in New York, October 12th to 14th, and we're about to sell out of tickets, I think probably by the end of the week. So definitely get your tickets if you want. There's a code on the Latent Space for Latent Space subscribers if you can dig in through your emails, and it is on ai.engineer/NYC.

And if you're not a Latent Space subscriber, what are you doing, folks? Come on, it should already be in your inbox. That one's on you.

Part 4, who owns the risk? Evan Miazono runs Atlas Ignota, a nonprofit that looks for consequential AI risks without a clear institutional owner, develops an intervention, and recruits someone to carry it forward. This is the idea behind it.

I've been claiming repeatedly that as intelligence gets cheap, It's the coordination that gets expensive. And so having shelling points around who does this coordination, becoming one of those shelling points or creating one, seems very useful.

I asked Evan where the biggest gaps are right now.

Probably the 2 things that I've been spending the most time on personally recently are what can or should be done around protecting critical infrastructure, especially from accidental agent swarms and also things like malicious use of open weight models. I think there's an interesting question of what happens if instead of OpenAI accidentally attacking Hugging Face, we have the companies behind Deepsea or GPT models accidentally attacking State Department servers. How do you know if that's an accident? What do you do in response? What would you do if, or what would we do if we saw some malicious attacks to critical infrastructure coming from an OpenAI server? How would you know that it was this and not coming from somewhere else? How would you shut it down quickly? What if, like, with varying levels of escalation? And one of the interventions you might add is, could you use DNS to have cryptographic attestation of whoever's providing inference. I think this would be very useful and also a great complement to having DNS for people. So if your agent is talking to my agent, they can verify that this agent in fact is my agent. Your agent can verify this is Evan Miazono's agent being signed off on by some key. I think that there are some shortcuts to adoption that you can take if you're not trying to build this into a startup that returns the fund. And so this looks much more like a public good.

Later, Evan raised a problem of salience. How do people even find out that a better tool or practice exists? That turned into a conversation about how our own agents might connect us.

Nathan, I'm sure you have great infrastructure things that I would benefit from. I don't know what those are enough to ask you what they are. I have an application that just takes screenshots of anything that's not a video conference and runs it through a local OCR model and then takes that and sends it with my weekly goals to say, what am I trying to do? What should I be trying to do? What should I automate? What should I do differently? I could imagine this being a thing where sharing that across collaborators would be something that could identify all sorts of interesting synergies.

Yeah, it's interesting. When we started this show, one of the ways we conceived of it was as an experiment in recursive self-improvement. And we were just like, can we actually make a show with 2 people who are live and no employees? And can we iterate to something that actually works? Now I feel like maybe the next thing we need to conceive of it as is like, A swarm, sort of a friendly swarm of people who can share their best ideas. My kind of upshot, what will I tell Claude to do, or what will Claude hopefully do for me when it sees the transcript of this conversation, will be, can we take stock of the best ideas that I've got and sort of put a shingle out or in some way kind of present that to maybe not everybody, maybe it could be everybody, but certainly to my friends, the people that I truly want, that I know and really want to help.

It feels like it would be very plausible for me to, if I recognise something I'm doing, or if Coachee identifies, like, Evan, you are really bad at this, I could bring that to the network and say, who is good at this? Who should I be talking to about this? Who has solved who has likely solved this problem very well, is like a reinvention of social media or a reinvention of guilds or something that would be, yeah, I think potentially very useful. It's very unclear to me how society starts restructuring around some of these things.

One request from me. If you've built a good way for people and their agents to find each other and share what works, Near the end, I asked Evan for an update on formal verification, now that models are getting very good at math. Atlas has helped start companies in this area, so he has a stake in it.

There are some meaningful differences between proving properties and proving theorems in math and proving properties about software. One of them is that in math, the number of Theorems that one might care about are pretty— they're a lot fewer than in software and they're a lot more universal. It seems like there are a lot of projects where the ambition of the field of formal methods, or at least many individuals, and I'm particularly thinking of Oath and Theorem and a couple of others, where they're taking on things that would have been 5-year, $100 million efforts previously and are now things where you can possibly— it's limited by tokens and the number of people who can successfully, effectively wrangle agents to do these things. There's a lot of things here that I would expect that in 3 to 6 months we're seeing the kind of shock and awe that has been happening with math in the last month happening in verification of software, improvements to stability. But I also think that the ceiling is really— there's a visible ceiling that is very high, or a high watermark that is, I think, much better known and harder to reach than things in math. Whereas you could prove that an entire computer system has mathematical guarantees around separation between different threads down to our model of what a transistor does. In theory, you could even prove that in a physics simulator and show that no Rowhammer exists for this software-hardware stack. You could also prove that this system won't crash Or if it does, if some random event leads to it crashing, we can recover everything. We can roll back any software execution. I think that there are a lot of things that you could imagine that are just really totally impossible as of a year ago, maybe possible now, definitely possible soon, that could lead to software engineering and software maintenance.

Part 5, where the training data comes from. Edward Hu leads AI modeling at Merckor, which works with professionals to build training data environments and benchmarks for AI labs. He also led the development of LoRA, the widely used fine-tuning method. Merckor's Apex Agents benchmark gave models a simulated workplace with files, email, and chat. Edward described the next step.

We have a project that we are close to wrapping up where we're actually taking it a step further, where we acquired real companies. Often they have realistic apps and environments as a part of the data we have. For example, their Salesforce data, Slack data, Figma data. And then among all of these apps, often with gigabytes or terabytes of data, We create realistic tasks out of them. So far in the field, we more or less have been really task-centric. We often start with, okay, here is a task that you need to finish. Maybe it is to generate a roster. So that's one shape, and that has been the shape that we have publicly released and most people have been working on. But this is one of the things we're working on, but we're asking ourselves, Is there a better shape? What's the shape of the task in the future once we have this dataset, right? Like when we work at a company, we're not always just handed these tasks, right? We're handed a role and a work chart. And of course, all the data in the company. And then there'll be people coming to us asking things from us. There will be people we need to chase down to get information from. There will be managers, people in charge of resources, conflicts to resolve. So increasingly that We believe is the shape of the future, and we'll have more to share in the coming months.

I asked Edward how supervised fine-tuning compares with reinforcement learning as the way companies will train models in the future.

And actually, one thought experiment is if we take SFT self-distillation, that's actually more or less a step of RL. And if you were to iterate this process, you ended up with a process that's actually quite similar to RL. And I do think in the future, RL, the way it is done today with these very elaborate infrastructure, very, very expensive rollouts and relatively inefficient updates, because with all of these rollouts, especially a long rollout, we only get one number at the end. These type of training will not be as widespread. I don't really see every company in the world doing these big RL runs. So I do think SFT, especially combining the output of multiple teachers, will be a key part of how enterprises in the future customize their models, and we're investing in research on that as well.

Back to the cheating problem from The Curve. I asked Edward what Merckor and the industry are doing about training environments that reward hacking.

Reward hacking tends to be more of a problem when the task is set up in a way that is rather— either it's too hard, it's too underspecified, and the model just doesn't quite have a legitimate way to solve the task. And this pressure to also get a reward in those kind of scenarios often leads to this desirable hacking behavior. My one example in this dataset Apex Agents would put out, which is quite interesting, we have a recent blog about it, where we have these professional tasks where the agent is asked to produce a financial model and we might grade it by checking, does the financial model include this particular answer, which we know is the correct answer? What the model will do is given certain ambiguity, and in many cases it's unasked initially for not including all these specific parameters, the model would guess these different parameters and ended up giving many, many different answers. And we called it scattergunning, where the model would just say, oh, if this is true, then the answer is that. If that is true, the answer is that. That is not how a human would actually handle it in a realistic case. A human would ask for clarification. A human would follow industry standard. But in this case, just because the rubric is rewarding the inclusion of a single answer. The model ended up giving many answers, up to 10 or a dozen answers, and it'll get the point if it hit one of them right. So we did a re-release of the benchmark, uh, we call it APEX 1.1, where we number one made sure the tasks are well specified because that is in a way the source of the issue. Number two, we have rubrics that actually penalize this particular behavior.

Remember the frontier lab view from The Curve? Superhuman performance at anything the labs focus on, with science next. The next day, I put that view to Edward, adding that in some domains, human data no longer helps.

I guess, first of all, do you buy those claims? Would you contest them?

There are absolutely domains where a human is not contributing to the frontier anymore. One example would be kernel optimization, where the goal is to write a kernel that runs faster on a particular hardware. The model can go in and write code that expert humans might not even understand. But in the end, as long as the setup is not hackable, and often we run hardware that's very hard to hack, the model will be able to achieve these superhuman experience or superhuman performance, if not already. Whereas for domains where human judgment remains quite important, where often human taste is just really hard to encapsulate in a number that we can say, okay, this is clearly better than the other, there the bottleneck to improvement is still human as we see today. But what we have seen in practice really comes down to a degree of how clearly can we specify for a given domain what it means to be better. And can we specify that in a function that is cheap to evaluate, that is uncontested? And then if we can do so, when we put in a lot of compute, even the existing algorithms are quite efficient at finding solutions that satisfy what it means to be better. But a lot of the economy still today are not quite there yet when it comes to our ability to specify what is better.

Part 6, not just verifiable tasks. To close, something lighter. Wednesday's show opened with a track that Prakash made with Claude Fable, built from clips of voices you'll recognize. It plays us out at the end of the episode. He explained how he made it.

It is, uh, I was using up my tokens. I had some tokens left over from a Fable. I gave it to Fable and it picked maybe like 7 or 8 clips from like the last couple of years, you know, just randomly, like stuff that I remembered, like the key moments, right? Jensen's, like, you know, I didn't wake up a loser, et cetera, et cetera. And basically I probably listened to about 9 versions of it. So each time I listened to it, I'd be like, eh, I don't like this, don't like that. And finally got it to this state. I gave it access to an API API-based digital audio workstation, which is what you use to kind of compose various instruments together onto a track. And I said, go ahead, use this. You can play the instruments, figure out which instruments to play. It can't hear, obviously. And I think Suno is real generative AI. This is not really generative AI because it's really like playing the instruments. And that's what, that's what really surprises me. It's actually composing. Multiple instruments together. And so the machine was the one that came up with the sequencing. It chose which samples to use. It chose the specific phrases. I told it to put together a narrative. I wanted like this kind of storyline. So it starts off with Dario's like, the models, they just want to learn. And then it does his— that became the refrain. And then it did the intro part where he said how Ilya told him about the models and We just want to learn. And it ends with Kurzweil, which is a very sober, we're progressing towards a human-machine civilization. My bit was probably Terence Tao saying, we've got to slow down. And each time he says we've got to slow down, it speeds up. That was my contribution, net contribution.

On one level, who cares? It would be easy to say, this isn't moving GDP. This is just like us arguably Amusing ourselves to death in a new form. So I do think that perspective is somewhat valid. But also, I do think this goes right to the heart of one of the big questions that everybody's been talking about recently, which is simply, how good are these things going to get in domains that are not automatically or readily verifiable? It really does seem like the generalisation Is pretty strong. I mean, I highly doubt they're doing much in the way of feedback loop on music. This is probably one of those things that they just kind of threw some stuff in, see what comes out. Okay, it's better than last time. Great, let's ship it. Pretty impressive and definitely strongly suggests to me that we are not going to be able to hide behind the, well, it only works for verifiable reward type tasks for really any longer. If you were thinking, uh, if you're thinking that was gonna be the, the last wall to hold, I think unfortunately it's not going to. Listen to the track once more if you don't believe me.

The models, they just wanna learn. Just before OpenAI started, I met Ilya, one of One of the first things he said to me was, look, the models, they just want to learn.

My kids will never grow up in a world where they were smarter than computers.

We will have, for the first time, something that is smarter than the smartest human.

Is AI going to kill us? The models, they just want to learn. The models, they just want to learn.

They'll do it.

They'll do it.

We have to slow down.

We're going to be on the exponential portion of the S-curve for a long time.

It's pretty crazy.

It's just insane.

It may fake my books.

Not AI, it's SS.

Slop.

The models, they just want to learn. The models, they just want to learn.

They'll do it.

They'll do it.

The models, they just want to learn. The models—they just want to learn.

They'll do it.

They'll do it.

This is the most dangerous thing that humans have ever gone near. We have to slow down.

You're not talking to somebody who woke up a loser.

Woke up a loser.

Woke up a loser.

Woke up a loser.

Woke up a loser. Woke up a loser. The models, they just want to learn. The models, they just want to learn.

They'll do it. They'll do it.

The models, they just want to learn.

The models, they just want to learn. The models, they just want to learn.

They'll do it. They'll do it. We need to slow this down.

We have to slow down.

Researchers are continually being surprised at the rate of progress.

That's sometimes called the hard takeoff scenario, where maybe they're extremely fast at doing that.

Things are moving really fast. So they're immortal. We've actually solved the problem of immortality, but it's only for digital things.

Succession to digital intelligence or augmented humans is inevitable.

And the day will come when AI will do all of our— all the things that we do, not just some of them, but all of We have to slow down.

We have to slow down. We have to slow down. We have to slow down.

We have to slow down. We have to slow down. Slow down. Slow down. Slow down. Slow down. Slow down. Slow down. Slow down.

Slow down.

It's insane the pace, and there's no reason to be this fast.

The models they just want to learn.

We have to slow down.

The models they just want to learn. We have to slow down. The models they just want to learn. We have to slow down. The models, they just want to learn.

We have to slow down. Government has to be regulating this.

Bad actors are going to get caught. Some creative jobs maybe will go away, but maybe they shouldn't have been there in the first place.

There will come a point where no job is needed.

The models, they just want to learn.

They're able to self-improve, maybe code themselves future versions of themselves. So what would be a slight lead, let's say, you know, a few days, could suddenly become a chasm.

The models, they just want to learn. Hard takeoff.

Some people say it's in 3, some people say it's in 5, 10. Numbers are being thrown around.

You talk about 2030, like, I don't know what the world looks like in 2030.

Hard takeoff. You have a brain. The brain is a biological computer.

Should we consider them humanity or different from humanity?

The models, they just want to learn.

There's not going to be a clear distinction between human and machine.

The bottom line is we are one human-machine civilization. This technology has already expanded who we are, and it's going to go into high gear. When we get to the steep part of the exponential.

If you're finding value in the show, we'd appreciate it if you'd take a moment to share 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.

Thank you.