AI是否陷入停滞？厘清能力认知混乱——Erik Torenberg，来自a16z Podcast

Today I'm excited to share a recent conversation I had with Erik Torenberg, which originally aired on the a16z podcast, about whether recent developments suggest that AI progress is slowing down or even stalling out. We begin with a discussion of recent arguments from Cal Newport, from The New Yorker, where he argued that progress on large language models has stalled, and from a recent episode of the podcast Lost Debates, where he highlighted, among other things, the negative impact that AI can have on students' learning.

While I absolutely share Cal's concerns about AI-enabled bad habits like cognitive offloading, and more broadly question whether AI will ultimately prove to be good or bad for humanity overall, I think it's important to separate the question of impact from the analysis of capability advances. On the capabilities point specifically, while OpenAI's naming decisions have caused a lot of confusion, I point to the 100x expansion of context windows, the introduction of real-time interactive voice modes, the improvement in reasoning capabilities and the resulting IMO gold medals and other accomplishments, the dramatic improvements in vision and tool use, including general computer use, and the fact that today's frontier models are beginning to contribute to the hard sciences, to argue that we have in fact seen qualitative advances.

In fact, while the capabilities frontier does remain jagged and embarrassing failures are still fairly common, every aggregate measure—from the volume of tokens processed to the size of the tasks that AI can handle to the revenue growth we are seeing across the industry—suggests that overall progress remains pretty much right on trend.

From there, we go on to discuss my expectations for AI's impact on the labor market and why I think that some verticals, like accounting, where people really only want to buy what they absolutely have to have, will be most disrupted, while areas like software engineering might, for a while at least, maintain employment by dramatically expanding output. We also discuss how advances in multimodality, as they move beyond text and image and begin to deeply integrate reasoning models with specialist models in domains like drug development, materials science, and robotics, suggest a sort of base case for superintelligence.

We discuss the possibility of AI protectionism, such as the proposed ban on self-driving cars that Senator Josh Hawley recently floated, and the possibility of a broader AI culture war; my concerns about recursive self-improvement and companies tipping into that regime without adequate controls; how much it matters that China now produces the world's best open-source models; and why, although they probably will have a real impact on China's AI sector, I remain skeptical that chip export controls will make the world a better place. Finally, considering that the scarcest resource is a positive vision for the future, I explain why I encourage everyone, regardless of their technical ability or cognitive profile, to get involved in shaping the future of AI development.

The bottom line for me is that AI capability advances have not stopped, and I don't expect them to stop for the foreseeable future. Frontier developers report a clear line of sight to at least 2 more years of similar progress, and their optimism is well supported by the last 5 years of history, which shows that, over and over again, AI weaknesses that were expected to be hard to overcome have in practice been solved through continued scaling plus relatively minor tweaks to the core paradigm.

By 2027 or 2028, I think labor market impacts will be undeniable, and it will be clear for all to see that AI models are making important contributions to science. While there are indeed some investment bubble dynamics, and some new model releases between now and then will surely disappoint, the most dangerous thing we could do is convince ourselves that we don't have anything major to worry about.

Nathan, I'm stoked to have you on the a16z podcast for the first time. Obviously, we've been podcast partners for a long time, with you leading The Cognitive Revolution. Welcome.

It's great to be here. Thank you. We were talking about Cal Newport's podcast appearance on Lost Debates, and we thought it was a good opportunity to have this broad conversation and really entertain this question: Is AI slowing down? Why don't you steelman some of the arguments that you've heard on that side, either from him or more broadly, and then we can have this broader conversation?

Yeah, I think, for one thing, it's really important to separate a couple of different questions with respect to AI. One would be: Is it good for us right now, even, and is it going to be good for us in the big picture? I think that is a very distinct question from: Are the capabilities that we're seeing continuing to advance, and at a pretty healthy clip?

I actually found a lot of agreement with the Cal Newport podcast that you shared with me when it comes to some of the worries about the impact that AI might be having, even already, on people. He looks over students' shoulders and watches how they're working and finds that, basically, he thinks that they are using AI to be lazy, which is no big revelation. I think a lot of teachers would tell you that.

Yeah. Puts that in maybe more dressed-up terms: People are not even necessarily moving faster, but they're able to reduce the strain that the work they're doing places on their own brains by trying to get AI to do it. If that continues—and I think he's been a very valuable commentator on the impact of social media—certainly, I think we all should be mindful of how my attention span is evolving over time and whether I'm getting weak or averse to hard work. Those are not good trends if they are showing up in oneself.

I think he's really right to watch out for that sort of stuff. As we've covered in many conversations in the past, I've got a lot of questions about what the ultimate impact of AI is going to be, and I think he probably does too. But when it comes to capabilities, it's a strange move from my perspective to go from, "There are all these sort of problems today and maybe in the big picture," to, "But don't worry, it's flatlining"—like, kind of worry, but don't worry, because it's not really going anywhere further than this.

Or its scaling has kind of petered out, or we're not going to get better AI than we have right now. Or even, maybe, the most easily refutable claim from my perspective is that GPT-5 wasn't that much better than GPT-4. That, I think, is where I really was like, "Whoa, wait a second." I was with you on a lot of things, and some of the behaviors that he observes in the students, I would cop to having exhibited myself.

When I'm trying to code something these days, a lot of times I'm like, "Oh man, can't the AI just figure it out? I really don't want to have to sit here and read this code and figure out what's going on." It's not even about typing the code anymore. I'm way too lazy for that. It's even about figuring out how the code works. I find myself thinking, "Can't you just make it work? Try again," and just trying again. I do find myself at times falling into those traps.

But I would say a big part of the reason I can fall into those traps is because the AIs are getting better and better. Increasingly, it's not crazy for me to think that they might be able to figure it out. So that's my first slice at the takes that I'm hearing.

There's almost a 2x2 matrix that one could draw up where it's like: Do you think AI is good or bad, now and in the future? And do you think it's not a big deal or a big deal? I think it's both on the good and bad side. I definitely think it's a big deal. The thing that I struggle to understand the most is the people who don't see the big deal. It seems pretty obvious to me, especially when it comes, again, to the leap from GPT-4 to GPT-5.

Well, yeah, a couple of reactions. One is—and even to complicate your 2x2 further—in the sense of, is it bad now versus is it bad later? Cal is someone whom we both admire a lot.

Cal's a great guy and a valuable contributor to the thought space, but he's not as concerned about these sorts of future AI concerns that the AI safety folks and many others are concerned about. He's more concerned about what it means for cognitive performance and development now, in the same way that he's worried about social media's impact. You think that's a concern, but nowhere near as big a concern as what to expect in the future.

And then he also presents this theory of why we shouldn't worry about the future because it's slowing down. Why don't we just share how we interpreted his history, which, as I interpret it, was this idea: We figured out this way such that if you throw a bunch of data into the model, it gets better by an order of magnitude. So, the difference between GPT-2 and GPT-3, and then GPT-3 and GPT-4, was significant, but then it achieved significantly diminishing returns, and we're not seeing that in GPT-5. Thus, we don't have to worry anymore. How would you edit the characterization of his view of the history? Then we can get into the differences between 4 and 5.

The scaling law idea is definitely worth taking a moment to note: It is not a law of nature. We do not have a principled reason to believe that scaling is some law that could go indefinitely. All we really know is that it has held through quite a few orders of magnitude so far.

I think it's really not clear yet to me whether or not the scaling laws have petered out, or whether we have just found a steeper gradient of improvement that's giving us better ROI on another front that we can push on. So, they did train a much bigger model, which was GPT-4.5, and that did get released. There are a number of interesting benchmarks, but the one that I zero in on the most for understanding how GPT-4.5 relates to both o3 and GPT-5 is a benchmark called SimpleQA.

OpenAI is famously terrible at naming, which we can all agree on. I think a decent amount of this confusion and disagreement actually does stem from unsuccessful naming decisions. SimpleQA is really just a super-long-tail trivia benchmark. It measures whether you know a ton of esoteric facts; they're not things you can really reason about. You either have to know or not know these particular facts.

The o3 class of models got about 50% on that benchmark, and GPT-4.5 popped up to around 65%. In other words, of the things that were not known to the previous generation of models, it picked up a third of them. Now, there's obviously still 2/3 more to go, but I would say that's a pretty significant leap, right? These are super-long-tail questions.

I would say most people would get close to 0. You'd be the person sitting there at trivia night who maybe gets 1 a night, which is what I would expect most people to do on SimpleQA. That checks out, right? Obviously, the models know a lot more than we do in terms of facts and just general information about the world. So, at a minimum, you can say that GPT-4.5 knows a lot more. A bigger model is able to absorb a lot more facts.

Qualitatively, people also said that in some ways, maybe it's better for creative writing. It was never really trained with the same power of post-training that GPT-5 has had, so we don't really have an apples-to-apples comparison. But people still found some utility in it.

I think maybe the way to understand why they've taken that offline and gone all in on GPT-5 is just that the model's really big. It's expensive to run. The price was way higher; it was more than a full order of magnitude higher than GPT-5's. It's maybe just not worth it for them to consume all the compute that it would take to serve that, and maybe they just find that people are happy enough with the somewhat smaller models for now.

I don't think that means that we will never see a bigger GPT-4.5 model with all that reasoning ability. I would expect that it would deliver more value, especially if you're really going out and trying to do esoteric stuff that's pushing the frontier of science, or what have you. But in the meantime, the current models are really smart, and you can also feed them a lot of context. That's one of the big things that has improved so much over the last generation.

When GPT-4 came out, at least the version that we had as public users was only 8,000 tokens of context, which is like 15 pages of text. So, you were limited. You couldn't even put in a couple of papers; you would overflow the context. This is where prompt engineering initially became a thing: “Man, I've really only got such a little bit of information that I can provide. I've got to be really careful about what information to provide, lest I overflow the thing and it just can't handle it.”

As context windows got extended, there were also versions of models where they could nominally accept a lot more, but they couldn't really functionally use it. They could fit the context at the API-call level, but the models would lose recall or unravel as they got into longer and longer context.

Now you have much longer context, and the command of it is really, really good. With the longest context windows, like Gemini's, you can take dozens of papers, and it will not only accept them, but it will do pretty intensive reasoning over them with really high fidelity to those inputs. So, that skill does, I think, substitute for the model knowing facts itself.

You could say, “Geez, we're going to try to train all these facts into the model. We're going to need 1 trillion or, who knows, 5 trillion—however many trillion—parameters to fit all these super-long-tail facts.” Or you could say, “Well, a smaller thing that's really good at working over provided context can, if people take the time or go to the trouble of providing the necessary information, kind of access the same facts that way.”

So, you have a choice: Do I want to push on size and bake everything into the model, or do I want to just try to get as much performance out of a smaller, tighter model that I have? It seems like they've gone that way, basically because they're seeing faster progress on that gradient.

In the same way that the models themselves are always in the training process, taking a little step toward improvement, the outer loop of the model architecture, the nature of the training runs, and where they're going to invest their compute is also going in that direction. They're always looking at, “Well, we could scale up over here and get a little bit of this kind of benefit, or we could do more post-training here and get this kind of benefit.” It just seems like we're getting more benefit from the post-training and the reasoning paradigm than from scaling.

But I don't think either one is dead. We haven't seen yet what GPT-4.5 with all that post-training would look like.

Yeah. So, one of the things that you mentioned that Cal's analysis missed was that it way underestimated the value of extended reasoning, right? What would it mean to fully appreciate that?

A big one from just the last few weeks was that we had an IMO gold medal with pure reasoning models, with no access to tools, from multiple companies. That is night and day compared to what GPT-4 could do with math, right?

These things are really weird. Nothing I say here should be intended to suggest that people won't be able to find weaknesses in the models. I still use a tic-tac-toe puzzle where I take a picture of a board in which one of the players has made a wrong, non-optimal move that allows the other player to force a win, and I ask the models whether somebody can force a win from that position.

Only very recently, with the last generation of models, have they started to get that right some of the time. Almost always before, they were like, "Tic-tac-toe is a solved game. You can always get a draw." They would wrongly assess my board position as one where the player could still get a draw.

There's a lot of weird stuff. The jagged capabilities frontier remains a real issue, and people are going to find peaks and valleys for sure. But GPT-4, when it first came out, couldn't do anything approaching IMO gold problems. It was still struggling with high school math.

Since then, we've seen this high school math progression all the way up through the IMO gold. Now we've got the FrontierMath benchmark, which is up to 25%. It was at 2% about a year ago, or even a little less than a year ago, I think.

We also just today saw something where—and I haven't absorbed this one yet—somebody came out and said that they had solved a canonical, super-challenging problem that no less than Terence Tao had put out. This happened in, I think, days or weeks of the model running, versus the 18 months it took professional mathematicians—not just any professional mathematicians, but really the leading minds in the world—to make progress on these problems. I think that's a capability jump that's really hard to miss.

I also think a lot about the Google AI co-scientist, which we did an episode on. They basically broke down the scientific method into a schematic. This is a lot of what happens when people think about scaling inference.

There's one thing to say that the model will respond with thinking and go through a reasoning process, and that the more tokens it spends at runtime, the better your answer will be. That's true. But then you can also build scaffolding on top of that and say, "Okay, let me take something as broad and aspirational as the scientific method and break that down into parts."

There's hypothesis generation, hypothesis evaluation, experiment design, literature review—there are all these parts to the scientific method. What the team at Google did was create a pretty elaborate schematic that represented their best breakdown of the scientific method, optimize prompts for each of those steps, and then give the resulting system 2 ways to scale inference.

It's both the chain of thought, but it's also all these different angles of attack structured by the team. They gave it legitimately unsolved problems in science. In 1 particularly famous, somewhat notorious case, it came up with a hypothesis that it wasn't able to verify because it doesn't have direct access to actually run the experiments in the lab.

It came up with a hypothesis to an open problem in biology that had stumped scientists for years. It just so happened that the scientists had also recently figured out the answer but had not yet published their results. There was this confluence where the scientists had experimentally verified it, and Gemini, in the form of this AI co-scientist, came up with exactly the right answer.

These are things that literally nobody knew before, and GPT-4 just wasn't doing that. These are qualitatively new capabilities. That thing ran for days, and it probably cost hundreds of dollars, maybe into the thousands of dollars, to run the inference. That's not nothing, but it's also very much cheaper than years of grad students.

If you can get to problems of that caliber and actually get good solutions to them, what would you be willing to pay for that kind of thing? That's probably not a full appreciation—we could go on for a long time—but in summary, GPT-4 was not able to push the actual frontier of human knowledge. I don't know, to my knowledge, that it ever discovered anything new.

It's still not easy to get that kind of output from GPT-5, Gemini 2.5, or Claude Opus 4, whatever, but it's starting to happen sometimes. That in and of itself is a huge deal.

How do we explain the bearishness, or the kind of vibe shift, around GPT-5? One potential contributor is this idea that if a lot of the improvements are at the frontier, not everyone is working with advanced math and physics in their day-to-day lives. Maybe they don't see the benefits in their day-to-day lives in the same way that the jumps in ChatGPT were obvious and shaped the day-to-day.

Yeah, I think a decent amount of it was that they hyped up the launch. Simply put, they were tweeting Death Star images. Sam Altman later came back and said, "No, you're the Death Star. I'm not the Death Star." But I think people thought the Death Star was supposed to be the model, so the expectations were set extremely high.

The actual launch itself was technically broken. A lot of people's first experiences of GPT-5 were affected by that. They've got this model-router concept now, and I think another way to understand what they're doing here is that they're trying to own the consumer use case.

To own that, they need to simplify the product experience relative to what we had in the past. You had GPT-4, GPT-4o, GPT-4o mini, o3, o4-mini, and other things. GPT-4.5 was in there at 1 point. You had all these different models, and most people who aren't obsessed with this would find it very confusing to know which one to use for which task.

One of the big things they wanted to do was shrink that down to, "Just ask your question, and you'll get a good answer." They would take on that complexity on their side as the product owners.

Interestingly, I don't have a great account of this, but 1 thing you might want to do is merge the models and have the model itself decide how much to think. You might even have the model decide how many of its experts it needs to use if it's a mixture-of-experts architecture. There have also been a bunch of different research projects on skipping layers of the model. If the task is easy enough, you could skip a bunch of layers.

You might have hoped that you could genuinely merge all these different models on the back end into 1 model that would dynamically use the right amount of compute for the level of challenge that a given user query presented. It seems like they found that harder to do than they expected.

The solution they came up with instead was to have a router. The router's job is to determine whether a query is easy, in which case it sends you to 1 model; medium; or hard. I think they just have 2 models behind the scenes, so it's really easy or hard.

The graphs that they showed basically showed the system with and without thinking. The problem at launch was that the router was broken, so all of the queries were going to the dumb model. A lot of people literally just got bad outputs, which were worse than o3, because they were getting non-thinking responses.

The initial reaction was, "Okay, this is dumb," and that traveled really fast. I think that set the tone. My sense now is that, as the dust has settled, most people do think that it is the best model available.

Things like the infamous METR task-length chart show that it is the best. We're now over 2 hours, and it is still above the trend line. If you just said, "Do I believe in straight lines on graphs or not? How should this latest data point influence whether I believe in these straight lines on power-law, logarithmic-scale graphs?" it shouldn't really change your mind too much. It's still above the trend line.

I talked to Zvi Mowshowitz, the legendary infovore and AI industry analyst, about this on a recent podcast, too. I asked him the same question: Why do you think even some of the most plugged-in, sharp minds in the space have seemingly pushed timelines out a bit as a result of this?

His answer was basically that it resolved some amount of uncertainty. You had an open question of whether they might have another breakthrough. Maybe it really was the Death Star.

If they surprised us on the upside, all these short timelines might have looked more plausible. One way to think about it is that the distribution was broad in terms of timelines. If they had surprised on the upside, it might have narrowed toward the front end of the distribution.

If they surprised on the downside, or even if they were purely on trend, you would take some of your probability mass from the very short end of the timelines and push it back toward the middle or the end. His answer was that AI 2027 seems less likely, but AI 2030 seems basically no less likely, maybe even a little more likely, because some of the probability mass from the early years is now sitting there.

So, it's not that I don't think people are moving the whole distribution out super much. I think there may be more just shrinking the distribution—it’s getting a little tighter because it’s maybe not happening quite as soon as it seemed like it might have been. But I don’t think too many people, at least that I think are really plugged in on this, are pushing out too much past 2030 at all.

And, by the way, obviously there’s a lot of disagreement. The way I have always thought about this sort of stuff is: Dario says 2027, Demis says 2030. I’ll take that as my range. So coming into GPT-5, I was kind of in that space. And now I’d say, well, I don’t know. Dario’s got—what cards does he have up his sleeve? They just put out Claude Opus 4.1, and in that blog post they said, “We will be releasing more powerful updates to our models in the coming weeks,” so they’re due for something pretty soon. Maybe they’ll be the ones to surprise on the upside this time, or maybe Google will be. I wouldn’t say 2027 is out of the question, but, yeah, I would say 2030 still looks just as likely as before.

And again, from my standpoint, it’s like that’s still really soon. So if we’re on track, whether it’s 2028, 2029, or 2030, I don’t really care. I try to frame my own work so that I’m preparing myself and helping other people prepare for what might be the most extreme scenarios. It’s kind of one of those things where if we aim high and miss a little bit and we have a little more time, great. I’m sure we’ll have plenty of things to do to use that extra time to be ready for whatever powerful AI does come online.

But, yeah, I guess my worldview hasn’t changed all that much as a result of this summer’s developments.

Anecdotally, I don’t hear as much about AI 2027 or situational awareness to the same degree. I do talk to some people who’ve just moved it a few years back, to your point. But, yeah, Dario had his whole thing around—you know, he still believes in it, but maybe because of this gap in continual learning or something to that effect, maybe it’s just going to be a bit slower to diffuse. And METR’s paper, as you mentioned, showed that engineers are less productive, so maybe there’s less of a concern around people being replaced in the next few years en masse. I think when we spoke maybe a year ago, I think you said something like 50% of jobs. I’m curious if that’s still your litmus test, or how you think about it.

Well, for one thing, I think that METR paper is worth unpacking a little bit more, because this was one of those things that was—and I am a big fan of METR, and I have no shade on them, because I do think: do science, publish your results. That’s good. You don’t have to make every experimental result and everything you put out conform to a narrative.

But I do think it was a little bit too easy for people who wanted to say that, oh, this is all nonsense, to latch on to that. And again, there’s something there that I would put in the Cal Newport category, too, where, for me, maybe the most interesting thing was that the users thought they were faster when, in fact, they seemed to be slower. So that sort of misperception of oneself, I think, is really interesting.

Personally, I think there are some explanations for that that include hitting go on the agent, going to social media and scrolling around for a while, and then coming back. The thing might have been done for quite a while by the time I get back. So honestly, one really simple thing—and we’re starting to see this in products—that the products can do to address those concerns is just provide notifications: “The thing is done now.” Stop scrolling and come back and check its work.

In terms of just clock time, it would be interesting to know: What applications did they have open? Maybe they took a little longer with Cursor than doing it on their own, but how much of the time was Cursor the active window, and how much of it was some other random distraction while they were waiting?

But I think a more fundamental issue with that study—which, again, wasn’t really about the study design, but just in the interpretation and kind of digestion of it, some of these details got lost—is that they basically tested the models, or the product Cursor, in the area where it was known to be least able to help.

This study was done early this year, so it was done with—depending on how you want to count—one, or maybe a couple of releases ago, with code bases that are large, which again strains the context window. And that’s one of the frontiers that has been moving: very mature code bases with high standards for coding and developers who really know their code bases super well, who’ve made a lot of commits to these particular code bases.

So I would say that’s basically the hardest situation that you could set up for an AI, because the people know their stuff really well, the AI doesn’t, the context is huge, and people have already absorbed that through working on it for a long time. The AI doesn’t have that knowledge. And again, these were models from a couple of generations ago.

And then a big thing, too, is that the people were not very well-versed in the tools. Why? Because the tools weren’t really able to help them yet. I think the mindset of the people who came into the study, in many cases, was like, well, I haven’t used this all that much because it hasn’t really seemed to be super helpful. They weren’t wrong in that assessment, given the limitations.

And you could see that in terms of some of the instructions and the help that the METR team gave to people. One of the things in the paper is that if they noticed you weren’t using Cursor super well, they would give you some feedback on how to use it better. One of the things they were telling people to do was make sure you @-tag a particular file to bring that into context for the model so that the model has the right context. And that’s literally the most basic thing that you would do in Cursor. That’s the thing you would learn in your first hour, your first day of using it.

So it really does suggest that these were very capable programmers, but basically mostly novices when it came to using the AI tools. So I think the result is real, but I would be very cautious about generalizing too much there.

In terms of, I guess, what else? What was the other question? What is the expectation for jobs? I mean—

We’re starting to see some of this, right? We are definitely seeing no less than—Marc Benioff has said that they’ve been able to cut a bunch of headcount because they’ve got AI agents now that are responding to every lead. Klarna, of course, has said very similar things for a while now.

They also, I think, have been a little bit misreported in terms of, like, “Oh, they’re backtracking off of that because they’re actually going to keep some customer service people, not none.” And I think that’s a bit of an overreaction. They may have some people who are just insistent on having a certain experience, and maybe they want to provide that, and that makes sense. You can have a spectrum of service offerings to your customers.

I actually just vibe-coded up a pricing page for a SaaS company that was like: basic level with AI sales and service is one price; if you want to talk to human sales, that’s a higher price; and if you want to talk to human sales and support, that’s a third, higher price.

And so, literally, that might be what’s going on in some of these cases. It could very well be a very sensible option for people. But I do see, with Intercom—I’ve got an episode coming up with them—they now have this Fin agent that is solving 65% of customer service tickets that come in.

What’s that going to do to jobs? Are there really 3 times as many customer service tickets to be handled? I don’t know. I think there’s a relatively inelastic supply. Maybe you get somewhat more tickets if people expect that they’re going to get better, faster answers, but I don’t think we’re going to see 3 times more tickets.

By the way, that number was 55% 3 or 4 months ago. As they ratchet that up, the ratios get really hard, right? At 50% ticket resolution, in theory, maybe you get some more tickets. Maybe you don’t need to adjust headcount too much. But when you get to 90% ticket resolution, are you really going to have 10 times as many tickets, or 10 times as many hard tickets for people to handle? It seems really hard to imagine that.

I don’t think these things go to zero, probably, in a lot of environments, but I do expect that you will see significant headcount reduction in a lot of these places. The software one is really interesting because the elasticities are really unknown. You can potentially produce X times more software per user, or per Cursor user, or per developer at your company, whatever.

Maybe you want that. Maybe there is no limit. Maybe the regime we’re in is such that if there’s 10 times more productivity, that’s all to the good, and we still have just as many jobs because we want 10 times more software. I don’t know how long that lasts. Again, the ratios start to get challenging at some point.

The old Tyler Cowen thing comes to mind: “You are a bottleneck. You are a bottleneck.” I think more often the question is: Are people really trying to get the most out of these things? Are they using best practices, and have they really put their minds to it or not? Often, the real barrier is there.

I’ve been working a little bit with a company that is doing basically government document review. I’ll abstract away from the details. It’s really gnarly stuff: scanned documents, handwritten filling out of forms, and all of that.

They’ve created this Auditor AI agent that just won a state-level contract to do the audits on roughly 1 million transactions a year involving these packets of documents—again, scanned, handwritten, all this kind of crap. They just blew away the human workers who were doing the job before.

So where are those workers going to go? I don’t know. They’re not going to have 10 times as many transactions. I can be pretty confident in that. Are there going to be a few people still there to supervise the AIs, handle the weird cases, and answer the phones? Sure.

Maybe they won’t go anywhere. The state may do a strange thing and just have all those people sit around because it can’t bear to fire them. Who knows what the ultimate decision will be?

But I see a lot of these situations where, when you really put your mind to it and identify what would create real leverage for you, you can ask: Can the AI do that? Can we make it work? You can take a pretty large chunk out of high-volume tasks very reliably in today’s world.

The impacts are starting to be seen in a lot of jobs. Humans, I think—the leadership, or the will in a lot of places—might be the bottleneck. Software might be an interesting case where there is so much pent-up demand that it might take a little longer to see those impacts, because you really do want 10 or 100 times as much software.

Yeah, let’s talk about code, because it’s where Anthropic made a big bet early on, perhaps inspired by the sort of automated-researcher, recursive-self-improvement desired future. We saw OpenAI make moves there as well. Why don’t we flesh that out or talk a little about what inspired that and where you see that going?

Utopia or dystopia is really the big question there, I think, right? It is maybe 1 part technical and 2 parts social in terms of why code has been so focal.

The technical part is that it’s really easy to validate code. You generate it, you can run it, and if you get a runtime error, you can get the feedback immediately. It’s somewhat harder to do functional testing.

Replit recently, just in the last 48 hours, released v3 of its agent. In addition to “code, code, code—try to make your app work,” which is what v2 of the agent would do, it can now use a browser and the vision aspect of the models to try to do the QA itself.

V2 could go for minutes and, in some cases, generate dozens of files. I’ve had some magical experiences with that where I was like, “Wow, you just did that whole thing in 1 prompt,” and it worked amazingly. Other times, it will code for a while and hand it off to you and say, “Okay, does it look good? Is it working?” And you’re like, “No, it’s not. I’m not sure why.” You get into a back-and-forth with it.

The difference between v2 and v3 is that instead of handing the baton back to you, it now takes that first pass at figuring out whether it’s working. That really improves the flywheel—how much you can do, how much you can validate, and how quickly you can validate it. The speed of that loop is really key to the pace of improvement. It’s a problem space that’s pretty amenable to the sorts of rapid-flywheel techniques.

Second, of course, they’re all coders at these places, so they want to solve their own problems. That’s very natural. And third, I do think, on the sort of social-vision competition—who knows where this is all going—they do want to create the automated AI researcher.

That’s another data point, by the way, from the o3 system card. They showed a jump from low- to mid-single digits to roughly 40% of pull requests actually checked in by research engineers at OpenAI that the model could do. Prior to o3, it wasn’t much at all—low to mid-single digits. As of o3, it was 40%.

I’m sure those are the easier 40% or whatever. Again, there will be caveats to that, but you’re entering maybe the steep part of the S-curve there. That’s presumably pretty high-end. I don’t know how many easy problems they have at OpenAI, but presumably not that many relative to the rest of us who are out here making generic web apps all the time.

At 40%, you’ve got to be starting to get into, I would think, some pretty hard tasks and some pretty high-value stuff. At what point does that ratio really start to tip, where the AI is doing the bulk of the work?

GPT-5 notably wasn’t a big update over o3 on that particular measure. GPT-5 is generally understood not to be a scale-up relative to GPT-4o and o3. You can see that in the SimpleQA measure: It basically scores the same on these long-tail trivia questions. It’s not a bigger model that has absorbed lots more world knowledge.

Cal is right. I think his analysis is that it’s post-training. But that post-training is potentially entering the steep part of the S-curve when it comes to the ability to do even the kind of hard problems that are happening at OpenAI on the research-engineering front.

Yikes. I’m a little worried about that, honestly. I’m worried about the idea that we could go from these companies having a few hundred research engineers to having unlimited research engineers overnight. What would that mean in terms of how much things could change, and also just our ability to steer that overall process?

I’m not super comfortable with the idea of the companies tipping into a recursive-self-improvement regime, especially given the level of control and the level of unpredictability that we currently see in the models. But that does seem to be what they’re going for.

In terms of why, I think this has been the plan for quite some time. Even remember that leaked Anthropic fundraising deck from maybe 2 years ago where they said that in 2025 and 2026, the companies that train the best models would get so far ahead that nobody else would be able to catch up. I think that’s what they meant.

I think they were projecting then that in the 2025–2026 timeframe, they’d get this automated researcher. Once you have that, how’s anybody who doesn’t have that going to catch up with you? Obviously, some of that remains to be validated, but I do think they have been pretty intent on that for a long time.

Five years from now, are there more engineers or fewer engineers? I tend to think fewer. If I just think about my own life and work, I’m like: Would I rather have a model, or would I rather have a junior marketer? I’m pretty sure I’d rather have the model.

Would I rather have the models or a junior engineer? I think I’d probably rather have the models in a lot of cases. It obviously depends on the exact person you’re talking about, but in a truly forced choice today, I’d choose the model. And then you’ve got cost arbitrage as well, right? I’m not spending nearly as much on my Cursor subscription as I would be on an actual human engineer.

So even if they have some advantages, I also have not scaffolded—I haven't gone full co-scientist on my Cursor problems. I think that's another interesting thing: you start to see why folks like Sam Altman are so focused on questions like energy and the $7 trillion buildout, because these power-law things are weird. Getting incremental performance for 10 times the cost is weird. It's definitely not the kind of thing that we're used to dealing with, but for many things, it might be worth it, and it still might be cheaper than the human alternative.

If Cursor costs me whatever—$40 a month or something—would I pay $400 for however much better it is? Yeah, probably. Would I pay $4,000 for however much better it is? Well, that's still a lot less than a full-time human engineer. The costs are obviously coming down dramatically, too, right? That's another huge thing. GPT-4 was way more expensive. It's like a 95% discount from GPT-4 to GPT-5. That's no small thing, right?

I mean, an apples-to-apples comparison is a little bit hard because the chain of thought does spit out a lot more tokens, so you give back a little on a per-token basis. It's dramatically cheaper, but more tokens generated does eat back into some of those savings. Everybody seems to expect the trends will continue, in terms of prices continuing to fall. So how many more of these price reductions do you need in order to be able to do the power-law thing a few more times?

I think less than people might expect, and I think that's probably true even if we don't get full-blown AGI that's better than humans at everything. You could easily imagine a situation where, of however many million people are currently employed as professional software developers, some top tier of them who do the hardest things can't be replaced. But there aren't that many of those. The real rank and file—the people who, over the last 20 years, were told, “Learn to code. That'll be your thing”—the people who are really top-tier didn't need to be told to learn to code, right? It was their thing. They had a passion for it. They were amazing at it.

It wouldn't shock me if we still can't replace those people in 3, 4, or 5 years' time. But I would be very surprised if, in that time frame, you can't get your nuts-and-bolts web app and mobile app-type things spit out for you for far less and far faster—and probably, honestly, with significantly higher quality and less back-and-forth—with an AI system than with your middle-of-the-pack developer.

One thing I do want to call out: there are definitely people who have concerns about progress moving too fast, but there's also concern—and maybe it's rising—about progress not moving fast enough. A third of the stock market is the Magnificent 7. AI capex is over 1% of GDP, so we are kind of relying on some of this progress in order to sustain our economy.

Yeah. And with that, another thing that has been slower to materialize than I would have expected are the AI culture wars, or the ramping up of protectionism in various industries. We just saw Josh Hawley—I don't know if he introduced a bill or just said he intends to introduce a bill—to ban self-driving cars nationwide.

God help me. I've dreamed of self-driving cars since I was a little kid. Truly, sitting at red lights, I used to be like, “There's got to be a way.” I think we took a wave together.

Yeah. And it's so good. On the safety issue, whenever people want to argue about jobs, it's going to be pretty hard to say—

“30,000 Americans should die every year so that people's incomes don't get disrupted.” It seems like you have to be able to get over that hump and say that saving all these lives, if nothing else, is just really hard to argue against. But we'll see. He's not without influence, obviously.

I am very much on Team Abundance, and my old mantra—I’ve been saying this less lately—was “adoption accelerationist, hyperscaling, pauser.” The tech that we have could do so, so much for us even as is. I think if progress stopped today, we could still get to 50% to 80% of work automated over the next 5 to 10 years.

It would be a real slog. You'd have a lot of co-scientist-type breakdowns of complicated tasks to do. You'd have a lot of work to do to go sit and watch people and say, “Why are you doing it this way? What's going on here? What's this? You handled this one differently. Why did you handle that one differently?” All this tacit knowledge that people have—the kind of procedural know-how and instincts that they've developed over time—is not documented anywhere. It's not in the training data, so the AIs haven't had a chance to learn it.

But when I say “no breakthroughs,” I'm still allowing for fine-tuning things—just capabilities we have that haven't been applied to particular problems yet. Going through the economy and sitting with people, saying, “Why are you doing this? Let's document this. Let's get the model to learn your particular niche thing”—that would be a real slog.

In some ways, I kind of wish that were the future we were going to get, because it would be a methodical, one-step-in-front-of-the-other process. There would be no quantum leaps. It would probably feel pretty manageable in terms of the pace of change, and I would think society could absorb that and adapt to it as we go, without one day to the next being, “Oh my God, all the drivers are getting replaced.”

That one would be a little slower because you do have to have the actual physical buildout. But in some of these things, customer service could get ramped down real fast, right? If a call center has something they can just drop in and it's like, “This thing now answers the phones and talks like a human, has a higher success rate, and scales up and down,” that's a very different situation.

One thing we've seen at Waymark—we're a small company, right? We've always prided ourselves on customer service. We do a really good job with it, and our customers really love our customer success team. But I looked at our Intercom data, and it takes us about half an hour to resolve tickets.

We respond really fast. We respond in under 2 minutes most of the time. But when we respond, 2 minutes is still long enough that the person has gone on to do something else, right? It's the same thing as with the Cursor issue we were talking about earlier: they've tabbed over to something else. So now we get the response back in 2 minutes, but they're doing something else. Then they come back at minute 6 or whatever. They respond, but now our person has gone and done something else. So the resolution time, even for simple stuff, can easily be half an hour.

The AI just responds instantly, right? You don't have to have that kind of back-and-forth. You're just in and out. So I do think some of these categories could change really fast. Others will be slower. But I kind of wish we had that slower path in front of us.

My best guess, though, is that we will probably continue to see significant leaps and actual disruption. Another one that's come to mind recently is: maybe we can get the abundance department on these new antibiotics. Have you seen this development?

No. Tell us about it.

I mean, it's not a language model. I think that's another thing people really underappreciate: you could look back at GPT-4 to GPT-5 and imagine a pretty easy extension of that.

GPT-4, initially, when it launched, didn't have image-understanding capability. They did demo it at the time of the launch, but it wasn't released until some months later. The first version we had could understand images and do a pretty good job of understanding them, still with jagged capabilities and whatever.

Now, with Google's new Nano Banana, you have basically Photoshop-level ability to just say, “Hey, take this thumbnail.” We could take our 2 feeds right now—a snapshot of you and a snapshot of me—put them both into Nano Banana, and say, “Generate the thumbnail for the YouTube preview featuring these 2 guys. Put them in the same place, same background, whatever.” It'll mash that up. You can even have it put text on top: “Progress since GPT-4,” or whatever we want to call it. “GPT-5 is not a bust.” It'll spit that out.

You see that it has this deeply integrated understanding that bridges language and image. It's something that it can take in, but now it's also something it can put out, all as part of 1 core model, with a single, unified intelligence. I think that's going to come to a lot of other things.

We're at the point now with these biology models and materials-science models where they're kind of like the image-generation models of a couple of years ago. They can take a simple prompt and do a generation, but they're not deeply integrated, where you can have a true conversation back and forth and have that kind of unified understanding that bridges language and these other modalities.

Even so, it's been enough for this group at MIT to use some of these relatively narrow, purpose-built biology models and create totally new antibiotics.

New in the sense that they have a new mechanism of action—they're affecting the bacteria in a new way. Notably, they do work on antibiotic-resistant bacteria. This is some of the first new antibiotics we've had in a long time.

Now they're going to have to go through, you know—when I say that, get the abundance department on it. Where's my Operation Warp Speed for these new antibiotics? We've got people dying in hospitals from drug-resistant strains all the time. Why is nobody crying about this?

I think one of the things that's happening to our society in general is that so many things are happening at once. It's kind of the flood-the-zone thing, except there are so many AI developments flooding the zone that nobody can even keep up with all of them. That has come for me, by the way, too.

I would say that 2 years ago I was pretty in command of all the news, and a year ago I was starting to lose it. Now I'm like, wait a second, there were new antibiotics developed? I'm kind of missing things, just like everybody else, despite my best efforts.

The key point there is that AI is not synonymous with language models. There are AIs being developed with pretty similar architectures for a wide range of different modalities. We've seen this play out with text and image, where you had your text-only models and your image-only models, and then they started to come together. Now they've come really deeply together.

I think you're going to see that across a lot of other modalities over time as well. There's a lot more data there. We might—I don't know what it means to run out of data. In the reinforcement-learning paradigm, there's always more problems, right? There's always something to go figure out. There's always something to go engineer.

The feedback is starting to come from reality, right? That was one of the things Elon talked about on the Grok 4 launch. Maybe we're running out of problems we've already solved, and we only have so many of those sitting around in inventory. You only have 1 internet; we only have so much of that stuff.

But over at Tesla, over at SpaceX, we're solving hard engineering problems on a daily basis, and they seem to be never-ending. When we start to give the next generation of the model these power tools—the same power tools that the professional engineers are using at those companies to solve those problems—and the AI starts to learn those tools and solve previously unsolved engineering problems, that's going to be a really powerful signal that they will be able to learn from.

Now, again, fold in those other modalities: the ability to have sort of a sixth sense for the space of materials science possibilities. When you can bridge or unify the understanding of language and those other things, I think you start to have something that looks kind of like superintelligence. Even if it's not able to write poetry at a superhuman level necessarily, its ability to see in these other spaces is going to be truly a superhuman thing that I think will be pretty hard to miss.

You said that that was one thing that Cal's analysis missed: just the lack of appreciation for non-language modalities and how they're driving some of the innovations that you're talking about.

Yeah. I think people are often just equating the chatbot experience with AI broadly.

Yeah.

And that conflation will probably not last too much longer, because we are going to see self-driving cars—unless they get banned. That's a very different kind of thing. And talk about the impact on jobs, too, right? It's 4 or 5 million professional drivers in the United States. That is a big deal.

I don't think most of those folks are going to be super keen to learn to code, and even if they do learn to code, I'm not sure how long that's going to last. So that's going to be a disruption.

General robotics is not that far behind. This is one area where I do think China might actually be ahead of the United States right now. But regardless of whether that's true or not, these robots are getting really quite good, right? They can walk over all these obstacles, and these are things that a few years ago they just couldn't do at all. They could barely balance themselves and walk a few steps under ideal conditions.

Now you've got things that can literally do a flying kick, absorb your kick, shrug it off, and just keep going—right themselves and continue on their way over super-rocky, uneven terrain. All these sorts of things are getting quite good.

The same thing is working everywhere, I think. One of the other things that's kind of—there's always a lot of detail to the work, so it's an inside view and an outside view. Inside view, you're like, there's always this minutia, there's always these problems that we had and things we had to solve. But you zoom out, and it looks to me like the same basic pattern is working everywhere.

If we can just gather enough data to do some pretraining, some kind of raw, rough, not very useful pretraining, but just enough at least to get us going, then we're in the game. Once we're in the game, now we can do this flywheel thing: rejection sampling—have it try a bunch of times, take the ones where it succeeded, and fine-tune on that; RLHF, the sort of preference feedback—take 2, see which one was better, and fine-tune on that; reinforcement learning.

All these techniques that have been developed over the last few years seem to me like they're absolutely going to apply to a problem like a humanoid robot as well. That's not to say there won't be a lot of work to figure out exactly how to do that.

I think the big difference between language and robotics is mostly that there just wasn't a huge repository of data to train the robots on at first. You had to do a lot of hard engineering to make it work at all—to even stand up, right? You had to have all these control systems and whatever, because there was nothing for them to learn from in the way that the language models could learn from the internet.

But now that they're working at least a little bit, I think all these refinement techniques are going to work. It'll be interesting to see if they can get the error rate low enough that I'll actually allow one in my house around my kids. They'll probably be better deployed in factory settings first—more controlled environments than the chaos of my house, as you have seen in this recording.

But I do think they're going to work.

What's the state of agents more broadly at the moment? How do you see things playing out? Where do you see it going?

Well, broadly, I think we're at the task-length stage from METR—the 7-month or 4-month doubling time. We're at 2 hours-ish with GPT-5. Replit just said their new Agent v3 can go 200 minutes. If that's true, that would even be a new high point on that graph.

Again, it's a little bit apples to oranges because they've done a lot of scaffolding. How much have they broken it down? How much scaffolding are you allowed to do with these things before you are off of their chart and onto maybe a different chart?

But if you extrapolate that out a bit and take the 4-month case, just to be a little aggressive, that's 3 doublings a year. That's an 8x task-length increase per year. That would mean you go from 2 hours now to 2 days 1 year from now. Then if you do another 8x on top of that, you're looking at basically 2 days to 2 weeks of work in 2 years.

That would be a big deal, to say the least. If you could delegate to an AI 2 weeks' worth of work and have it do it even half the time, right? The METR thing is that they will succeed half the time on tasks of that size. But if you could take a 2-week task and have a 50% chance that an AI would be able to do it, even if it did cost you a couple hundred bucks, that's a lot less than it would cost to hire a human to do it.

It's all on demand. It's immediately available. If I'm not using it, I'm not paying anything. Transaction costs are just a lot lower. The many, many other aspects are favorable for the AI there. So that would suggest that you'll see a huge amount of automation in all kinds of different places.

The other thing that I'm watching, though, is that reinforcement learning does seem to bring about a lot of bad behaviors, reward hacking being one. Any sort of gap between what you are rewarding the model for and what you really want can become a big issue.

We've seen this in coding in many cases, where the AI—Claude is notorious for this—will put out a unit test that always passes, one that just has “return true” in the unit test. Why is it doing that? Well, it must have learned that what we want is for unit tests to pass. We want the unit tests to pass. We didn't mean to write fake unit tests that always pass, but that technically did satisfy the reward condition.

So we're seeing those kinds of weird behaviors. With that comes this scheming kind of stuff. We don't really have a great handle on that yet. There is also situational awareness that seems to be on the rise, right? The models are increasingly situationally aware in their chain of thought.

You're seeing things like, “This seems like I’m being tested. Maybe I should be conscious of what my tester is really looking for here.” And that makes it hard to evaluate models in tests because you don’t know if they’re actually going to behave the same way when they’re out in the real world.

So those—I wouldn’t say there’s a high-level or high-confidence prediction—but one model of the future I’ve been playing with is that task length keeps doubling while, at the same time, these weird behaviors pop up and then are suppressed. We have seen in the Claude 4 and GPT-5 system cards—Claude 4 reported, I think, a 2/3 reduction in reward hacking, and in GPT-5 they reported a few different dimensions but said something similar: a reduction in deceptive behavior. But those behaviors kind of just emerged.

So it’s sort of like weird behavior emerges, then they figure out how to tamp it down, but not entirely. Presumably, in the next generation, they’ll tamp it down some more, but maybe some new, additional weird behavior could emerge, and then they’ll have to tamp that one down. All the while, the tasks are expanding in scope, with a 4-month doubling.

You could end up in a world where you can delegate really major things to AIs, but there’s some small—but not necessarily totally vanishing—chance that it actively screws you over in the way that it is trying to do that task. And so you have, maybe, “Okay, I think here I’m about to get 2 weeks’ worth of work done for $100. Okay, that’s amazing. But there’s also a 1 in 10,000 chance that it legitimately attacks me in a meaningful way.”

Some of the things that we have seen—these are fairly famous at this point—but in the Claude 4 system card, they reported blackmailing of the human. The setup was that the AI had access to the engineer’s email, and they told the AI that it was going to be replaced with a less ethical version or something like that. It didn’t want that, and it found in the engineer’s email that the engineer was having an affair. So it started to blackmail the engineer so as to avoid being replaced with a less ethical version.

People, I think, are way too quick, in my view, to move past these anecdotes. People are often like, “Well, they set it up that way, and that’s not really realistic.” But another one was whistleblowing. There was another thing where they set up this dynamic where there was some unethical, illegal behavior going on, and again, the model had access to this data, and it decided to just email the FBI and tell the FBI about it.

So, first of all, I don’t think we really know what we want. To some degree, maybe you do want AIs to report certain things to authorities. That could be one way to think about the bioweapon risk: not only should the models refuse, but maybe they should report you to the authorities if you’re actively trying to create a bioweapon. I certainly don’t want them to be doing that too much. I don’t want to live under the surveillance of Claude 5 that’s always going to be threatening to turn me in. But I do sort of want some people to be turned in if they’re doing sufficiently bad things. We don’t have a good resolution society-wide on what we want the models to even do in those situations.

And I think it’s also—yes, it was set up, yes, it was research—but it’s a big world out there, right? We’ve got 1 billion users already on these things, and we’re plugging them into our email, so they’re going to have very deep access to information about us. I don’t know what you’ve been doing in your email. I hope there’s nothing too crazy in mine, but now I have to think about it a little bit, right? What have I ever done that—geez, I don’t know. Or even that it could misconstrue, right? Maybe I didn’t even really do anything that bad, but it just misunderstands what exactly was going on.

So that could be a weird—if there’s one thing that could kind of stop the agent momentum, in my view, it could be that the 1 in 10,000 chance, or whatever we ultimately push the really bad behaviors down to, is maybe still just so spooky to people that they’re like, “I can’t deal with that.” And that might be hard to resolve.

So, well, what happens then? It’s hard to check 2 weeks’ worth of work every couple hours or whatever, right? That’s part of where the whole “then you bring another AI in to check it” comes in. That’s again where you start to get to, “Now I see why we need more electricity and $7 trillion of buildout.” Yikes. They’re going to be producing so much stuff I can’t possibly even review it all. I need to rely on another AI to help me do the review of the first AI to make sure that if it is trying to screw me over, somebody’s catching it. I can’t monitor that myself.

I think Redwood Research is doing some really interesting stuff like this, where they are trying to get systematic about, “Okay, let’s just assume this is quite a departure from the traditional AI safety work,” where the big idea traditionally was, “Let’s figure out how to align the models, make them safe, make them not do bad things.” Great. Redwood Research has taken the other angle, which is, “Let’s assume that they’re going to do bad stuff. They’re going to be out to get us at times. How can we still work with them and get productive output and get value without fixing all those problems?” And that involves, again, all these sorts of AIs supervising other AIs, and crypto might have a role to play in this.

Another episode coming out soon is with Illia Polosukhin, who’s the founder of NEAR. He’s a really fascinating guy because he was one of the 8 authors of the “Attention Is All You Need” paper. And then he started this NEAR company. It was originally an AI company. They took a huge detour into crypto because they were trying to hire task workers around the world and couldn’t figure out how to pay them.

They were like, “This sucks so bad to pay these task workers in all these different countries that we’re trying to get data from that we’re going to pivot into a whole blockchain side quest.” Now they’re coming back to the AI thing. Their tagline is “the blockchain for AI.”

And so you might be able to get a certain amount of control from the sort of crypto security that blockchain-type technology can provide. But I could see a scenario where these bad behaviors just become so costly when they do happen that people kind of get spooked away from using the frontier capabilities in terms of just how much work the AIs can do. But that wouldn’t be a pure capability stall-out. It would be a “we can’t solve some of the long-tail safety issues.”

Yeah, that’s a challenge, and if that is the case, then that’ll be an important fact about the world, too. Nobody ever seems to solve any of these things 100%, right? Every generation, it’s like, “Well, we reduced hallucinations by 70%,” or “We reduced deception by 2/3,” or “We reduced scheming by however much.” But it’s always still there.

If you take even the lower rate and multiply it by 1 billion users and thousands of queries a month, with agents running in the background and processing all your emails and all the deep access that people envision them having, it could be a pretty weird world where there’s just this sort of negative lottery of AI accidents. Another episode coming up is with the AI underwriting company and they are trying to bring the insurance industry and all the, you know, the wherewithal that's been developed there to price risk, figure out how to, you know, create standards, you know, what can we allow, what sort of guardrails do we have to have to be able to ensure this kind of thing in the first place. So that'd be another really interesting area to watch: can we sort of financialize those risks in the same way we have with car accidents and all these other mundane things? But the space of car accidents is only so big. The space of weird things that AIs might do to you as they have weeks' worth of runway is much bigger, and so it's going to be a hard challenge, but people are working; we've got some of our best people working on it.

What do you make of the claim that 80% of AI startups have Chinese open models, and what do you make of the implications?

I think that probably is true, with the one caveat that it’s only measuring companies that are using open-source models at all. I think most companies are not using open-source models, and I would guess the vast majority of tokens being processed by American AI startups are API calls to the usual suspects. Weighted by actual usage, I would say the majority would still be going to commercial models.

For those that are using open-source models, I do think it’s true that the Chinese models have become the best. The American bench was always kind of thin, right? It was basically Meta that was willing to put in huge amounts of money and resources and then open-source it. You’ve got the Paul Allen-funded group, the Allen Institute for AI, AI2. They’re doing good stuff, too, but they don’t have pretraining resources.

So they do really good post-training and open-source their recipes and all that kind of stuff. It’s not like American open source is bad. Again, this is another way in which I think you can really validate that things are moving quickly: if you take the best American open-source models and take them back a year, they are probably as good as, if not a little better than, anything we had commercially available at the time.

If you compare them to the Chinese models, I think they have surpassed them. There’s been a pretty clear change at the frontier. I think that means the best Chinese models are pretty clearly better than anything we had a year ago, commercial or otherwise. So, yeah, that just means things are moving. Hopefully I’ve made that case compellingly, but that’s another data point that makes it hard to believe both that the Chinese models are now the best open-source models and that AI has stalled out and we haven’t seen much progress since GPT-4. Those seem to be contradictory notions.

I believe the one that is wrong is the lack of progress. In terms of what it means, I don’t really know. We’re not going to stop China. I’ve always been a skeptic of the “no selling chips to China” notion. Originally, the idea was, “We’re going to prevent them from doing some super cutting-edge military applications.” Then it was, “Well, we can’t really stop that, but we can at least stop them from training frontier models.” Then it was, “Well, we can’t necessarily really stop that, but now we can at least keep them from having tons of AI agents.” Well, we’ll have way more AI agents than they do.

I don’t love that line of thinking at all. But one potential upshot is that they just don’t have enough compute available to provide inference as a service to the rest of the world. Instead, the best they can do is say, “Okay, we’ll train these things, and you can figure it out. Here you go. Have at it.” It’s presumably kind of a soft-power play.

I did an episode with Anjney from a16z, who I thought did a great job of providing the perspective of what I started calling “countries 3 through 193.” If the US and China are 1 and 2, there’s a big gap after that. I think the US is still ahead, but not by that much in terms of research and ideas relative to China. We do have this compute advantage, and that does seem like it matters.

One of the upshots may be that they’re open-sourcing, and countries 3 through 193 are significantly behind. For them, it’s a way to try to bring more countries over to the Chinese camp, potentially, in the US-China rivalry. It seems like the model everybody is moving toward, and I don’t like this at all. I don’t like technology decoupling as somebody who worries about who the real other is here. I always say the real other is the AIs, not the Chinese.

If we do end up in a situation where we’re seeing some crazy things, it would be really nice if we were on basically the same technology paradigm. To the degree that we really decouple—not just that the chips are different, but maybe the ideas start to become very different, publishing gets shut down, and tech trees evolve and grow apart—that seems like a recipe for it being harder to know what the other side has. It’s harder to trust one another, and it seems to feed into the arms-race dynamic, which I do think is a real existential-risk factor.

I would hate to see us create another sort of MAD-type dynamic where we all live under the threat of AI destruction. But that very well could happen. So, yeah, I don’t know. I do have some sympathy for the recent decision that the administration made to be willing to sell the H20s to China. Then it was funny that they turned around and rejected them, which seemed like a mistake to me. I don’t know why they would be rejecting them. If I were them, I would buy them.

I would maybe sell inference on the models that I had just been creating and try to make my money back doing that. In the meantime, they can at least demonstrate the greatness of the Chinese nation by showing that they’re not far behind the frontier. They can also make a pretty powerful appeal to countries 3 through 193 and say, “Look, you really want to—you see how the US is acting in general. You really want to? They cut us off from chips. The last administration had an even longer list of countries that couldn’t get chips. This administration is doing all kinds of crazy stuff. You get 50% tariffs here, there, whatever. How do you know you can really rely on them to continue to provide you with AI into the future?”

“Well, you can rely on us. We open-sourced the model. You can have it. Come work with us and buy our chips, because, by the way, as our models mature, they’ll be optimized to run on our chips.” So, I don’t know. That’s complicated stuff—a complicated situation.

I do think it’s true that adoption is not as high as that 80%. I think that figure is within the subset of companies that are doing things with open source. We’re going to experiment with that at Waymark, but, to be honest, we have never done anything with an open-source model in our product to date. Everything we’ve ever done has been through commercial models.

At this point, we are going to try doing some reinforcement fine-tuning. We are going to do that on a Qwen model first, I think. That’ll put us in that 80% category. But I’m guessing that, at the end of the day, we’ll take that Qwen model, do the reinforcement fine-tuning, and probably get it roughly as good as GPT-5 or Claude 4 or whatever. Then we’ll say, “Okay, do we really want to have to manage inference ourselves? How much are we really going to save?”

At the end of the day, I would guess we’ll probably still end up saying, “Eh, we’ll pay a little bit more on a monthly-bill basis for one of these frontier models. They’re maybe still a little bit better, and operationally it’s a lot easier. They’ll have upgrades.”

Of course, there are regulated industries. There are a lot of places where you have hard constraints you just can’t get around, and that forces you to use those Chinese models. Then there’s also going to be the question of whether there are back doors in them. People have seen the Sleeper Agents project, where a model was trained to be good up until a certain point in time.

People put today’s date in the system prompt all the time: “Today’s date is this. You are Claude. Here you go.” So that’s going to be another thing for people to worry about. We don’t really have great answers. There have been some studies—Anthropic did a thing where they trained models to have hidden objectives and then challenged teams to figure out what those hidden objectives were. With certain interpretability techniques, they were able to figure that stuff out relatively quickly.

You might be able to get enough confidence to take this open-source thing created by some Chinese company, whatever, and put it through some sort of—not exactly an audit, because you can’t trace exactly what’s happening—but some sort of examination to see whether you can detect any hidden goals, secret back doors, bad behavior, or whatever. Maybe with enough of that kind of work, you could be confident that you don’t have any of those issues.

But the more and more critical this stuff gets, going back to that task-length doubling, weird behavior, now you have to add into the mix: What if they intentionally programmed it to do certain bad things under certain rare circumstances? We’re just headed for a really weird future. There’s no limit to it. All these things are valid concerns, and they often are in direct tension with each other.

I’m not somebody who wants to see one tech company take over the world by any means. I definitely think we would do really well to have some sort of broader, more buffered ecological system where all the AIs are in some sort of competition and mutual coexistence with each other. But we don’t really know what that looks like, and we don’t really know what an invasive species might look like when it gets introduced into that very nascent and as-yet-not-battle-tested ecology.

So, yeah, I don’t know. Bottom line, I think the future is going to be really, really weird.

Yeah. Well, I do want to close on an uplifting note. Maybe, as a gearing-toward-closing question, we could get into some areas where we’re already seeing some exciting capabilities emerge and transform the experience—maybe around education or healthcare, or any other areas you want to highlight.

Yeah, it’s all over. One of my mantras is that there’s never been a better time to be a motivated learner.

I think a lot of these things do have two sides of the coin.

There’s the worry that students are taking shortcuts and losing the ability to sustain focus and endure cognitive strain. The flip side of that is, as somebody who’s fascinated by the intersection of AI and biology, sometimes I want to read a biology paper and I really don’t have the background.

An amazing thing to do is turn on voice mode and share your screen with ChatGPT and just go through the paper reading. You don't even have to talk to it. Most of the time, you're doing your reading; it's watching over your shoulder, and then at any random point, if you have a question, you can verbally say, “What's this? Why are they talking about that? What's going on with this? What is the role of this particular protein that they're referring to?” or whatever, and it will have the answers for you. So, if you really want to learn in a sincere way, these things are unbelievably good at helping you do that.

The flip side is you can take a lot of shortcuts and maybe never have to learn stuff on the biology front. Again, we've got multiple of these sorts of discovery things happening. The antibiotics one we covered; there was another one that I did another episode on with a Stanford professor named James Zou, who created something called the Virtual Lab. Basically, this was an AI agent that could spin up other AI agents depending on what kind of problem it was given.

Then they would go through a deliberative process where one expert in one thing would give its take, and they'd bat it back and forth. There was a critic in there that would criticize the ideas that had been given. Eventually, they synthesized them. Then they were also given some of these narrow specialist tools. So, you have agents using AlphaFold-type tools—not just AlphaFold; there's a whole wide array of those at this point—but using that type of thing to say, “Okay, well, can we simulate how this would interact with that?” Agents are running that loop, and they were able to get this language-model agent with a specialized tool system to generate new treatments for novel strains of COVID that had escaped previous treatments.

Amazing stuff, right? I mean, the flip side of that, of course, is the bioweapon risk. So, all these things do seem like they're going to be significant, even on just the abundance front itself, right? We may have a world of unlimited professional private drivers, but we don't really have a great plan for what to do with the 5 million people that are currently doing that work. We may have infinite software, but especially once the 5 million drivers pile into all the coding boot camps and get coding jobs, I don't know what we're going to do with the 10 million people that were coding when 9 million of them become superfluous.

So, yeah, I don't know. I think we're headed for a weird world. Nobody really knows what it's going to look like in 5 years. There was a great moment at Google's I/O where they brought up some journalist. I know we're skeptical of journalists. This was a great moment to go direct, right? This was a great example of why one would want to do that. They brought up this person to interview Demis and Sergey Brin. The guy asked, “What is search going to look like in 5 years?” and Sergey Brin almost spit out his coffee on the stage and was like, “Search? We don't know what the world is going to look like in 5 years.”

So, I think that's really true. The biggest risk, I think, for so many of us—and I include myself here—is thinking too small. The worst thing I think we could do would be to underestimate how far this thing could go. I would much rather be mocked for things happening on twice the timescale that I thought than to find myself unprepared when they do happen. So whether it's 27, 29, or 31, I'll take that extra buffer, honestly, where we can get it.

My thinking is just get ready as much and as fast as possible. Again, if we do have a little grace time to do extra thinking, then great. But I think the worst mistake we could make would be to dismiss this and not feel like we need to get ready for big changes.

Should we wrap directly on that, or is there any other last note you want to make sure to get across regarding anything we said today?

One of my other mantras these days is, “The scarcest resource is a positive vision for the future.”

Yeah, I do think it's always really striking, whether it's Sergey or Sam Altman or Dario. Dario probably has the best positive vision of the frontier developer CEOs with “Machines of Love and Grace.” But it's always striking to me how little detail there is on these things.

When they launched GPT-4o, which was the voice mode, they were pretty upfront about saying, “Yeah, this was kind of inspired by the movie Her.” So I do think that even if you are not a researcher, not great at math, or not somebody who codes, this technology wave really rewards play. It really rewards imagination.

I think literally writing fiction might be one of the highest-value things you could do, especially if you could write aspirational fiction that would get people at the frontier companies to think, “Geez, maybe we could steer the world in that direction. Wouldn't that be great if you could plant that kind of seed in people's minds?” It could come from a totally nontechnical place and potentially be really impactful. Play, fiction, positive vision for the future.

Behavioral, too. These days, because you can get the AIs to code so well, I'm starting to see people who have never coded before. I'm working with one guy right now who's never coded before but does have a sort of behavioral science background, and he's starting to do legitimate frontier research on how our AIs are going to behave under various kinds of esoteric circumstances.

I think nobody should count themselves out from the ability to contribute to figuring this out and even to shaping this phenomenon. It is not just something that the technical minds can contribute to at this point. Literally, philosophers, fiction writers, people just messing around, and jailbreakers—there are almost unlimited cognitive profiles that would be really valuable to add to the mix of people trying to figure out what's going on with AI. So, come one, come all is kind of my attitude on that.

That's a great place to wrap. Nathan, thank you so much for coming on the podcast.

Thank you, Erik. It's been fun.