Dwarkesh Patel
I’m confused why some people have super-short timelines yet, at the same time, are bullish on scaling up reinforcement learning on top of LLMs. If we’re actually close to a humanlike learner, then this whole approach of training on verifiable outcomes is doomed.
1. What are we scaling?
Currently, the labs are trying to bake in a bunch of skills into these models through mid-training. There’s an entire supply chain of companies that are building RL environments that teach the model how to navigate a web browser or use Excel to build financial models. Now, either these models will soon learn on the job in a self-directed way, which will make all this pre-baking pointless, or they won’t, which means that AGI is not imminent.
Humans don’t have to go through a special training phase where they need to rehearse every single piece of software that they might ever need to use on the job. Baron Millig made an interesting point about this in a recent blog post he wrote. He writes:
“When we see frontier models improving at various benchmarks, we should think not just about the increased scale and the clever ML research ideas, but the billions of dollars that are paid to PhDs, MDs, and other experts to write questions and provide example answers and reasoning targeting these precise capabilities.
“You can see this tension most vividly in robotics. In some fundamental sense, robotics is an algorithms problem, not a hardware or a data problem. With very little training, a human can learn how to teleoperate current hardware to do useful work. So if you actually had a humanlike learner, robotics would be in large part a solved problem.
“But the fact that we don’t have such a learner makes it necessary to go out into a thousand different homes and practice a million times how to pick up dishes or fold laundry.”
Now, one counterargument I’ve heard from the people who think we’re going to have a takeoff within the next 5 years is that we have to do all this clunky RL in service of building a superhuman AI researcher. Then, a million copies of this automated Ilia can go figure out how to solve robust and efficient learning from experience.
This just gives me the vibes of that old joke: “We’re losing money on every sale, but we’ll make it up in volume.” Somehow, this automated researcher is going to figure out the algorithm for AGI, which is a problem that humans have been banging their heads against for the better part of a century, while not having the basic learning capabilities that children have. I find it super implausible.
Besides, even if that’s what you believe, it doesn’t describe how the labs are approaching reinforcement learning from verifiable reward. You don’t need to pre-bake in a consultant’s skill at crafting PowerPoint slides in order to automate Ilya. So clearly, the labs’ actions hint at a worldview where these models will continue to fare poorly at generalization and on-the-job learning, thus making it necessary to build into these models beforehand the skills that we hope will be economically useful.
Another counterargument you can make is that, even if the model could learn these skills on the job, it is just so much more efficient to build in these skills once during training rather than again for each user and each company. And look, it makes a ton of sense to just bake in fluency with common tools like browsers and terminals.
Indeed, one of the key advantages that AGIs will have is this greater capacity to share knowledge across copies. But people are really underrating how much company- and context-specific skill is required to do most jobs. And there just isn’t currently a robust, efficient way for AIs to pick up these skills.
2. The value of human labor
I was recently at a dinner with an AI researcher and a biologist. It turned out the biologist had long timelines, and so we asked why she had these long timelines. Then she said, “One part of work recently in the lab has involved looking at slides and deciding if the dot in that slide is actually a macrophage or just looks like a macrophage.”
The AI researcher, as you might anticipate, responded, “Look, image classification is a textbook deep-learning problem. This is dead-center in the kind of thing that we could train these models to do.”
I thought this was a very interesting exchange because it illustrated a key crux between me and the people who expect transformative economic impact within the next few years. Human workers are valuable precisely because we don’t need to build in special training loops for every single small part of their job.
It’s not net productive to build a custom training pipeline to identify what macrophages look like, given the specific way that this lab prepares slides, and then another training loop for the next lab-specific microtask, and so on. What you actually need is an AI that can learn from semantic feedback or from self-directed experience and then generalize the way a human does.
Every day, you have to do 100 things that require judgment, situational awareness, and skills and context that are learned on the job. These tasks differ not just across different people but even from one day to the next for the same person. It is not possible to automate even a single job by just baking in a predefined set of skills, let alone all the jobs.
In fact, I think people are really underestimating how big a deal actual AI will be because they are just imagining more of this current regime. They’re not thinking about billions of humanlike intelligences on a server that can copy and merge all the learnings.
3. Economic diffusion lag is cope
And to be clear, I expect this, which is to say I expect actual brain-like intelligences within the next decade or 2, which is pretty fucking crazy. Sometimes people will say that the reason AIs are more widely deployed right now across firms and already providing lots of value outside of coding is that technology takes a long time to diffuse.
I think this is cope. I think people are using this cope to gloss over the fact that these models just lack the capabilities that are necessary for broad economic value. If these models actually were like humans on a server, they’d diffuse incredibly quickly.
In fact, they’d be so much easier to integrate and onboard than a normal human employee is. They could read your entire Slack and drive within minutes. And they could immediately distill all the skills that your other AI employees have.
Plus, the hiring market for humans is very much like a lemons market, where it’s hard to tell who the good people are beforehand. Obviously, hiring somebody who turns out to be bad is very costly. This is just not a dynamic that you would have to face or worry about if you’re simply spinning up another instance of a vetted AI model.
So for these reasons, I expect it’s going to be much easier to diffuse AI labor into firms than it is to hire a person. Companies hire people all the time. If the capabilities were actually at AGI level, people would be willing to spend trillions of dollars a year buying tokens that these models produce.
Knowledge workers across the world cumulatively earn tens of trillions of dollars a year in wages. And the reason that labs are orders of magnitude off this figure right now is that the models are nowhere near as capable as human knowledge workers.
4. Goal-post shifting is justified
Now, you might be like, “Look, how can the standard have suddenly become that labs have to earn tens of trillions of dollars in revenue a year? Until recently, people were saying, ‘Can these models reason? Do these models have common sense? Are they just doing pattern recognition?’”
Obviously, AI bulls are right to criticize AI bears for repeatedly moving these goalposts, and this is very often fair. It’s easy to underestimate the progress that AI has made over the last decade. But some amount of goalpost shifting is actually justified.
If you showed me Gemini 3 in 2020, I would have been certain that it could automate half of knowledge work. So we keep solving what we thought were the sufficient bottlenecks to AGI. We have models that have general understanding. They have few-shot learning. They have reasoning. And yet we still don’t have AGI.
So what is a rational response to observing this? I think it’s totally reasonable to look at this and say, “Oh, actually, there’s much more to intelligence and labor than I previously realized.”
While we’re really close, and in many ways have surpassed what I would have previously defined as AGI, the fact that model companies are not making the trillions of dollars in revenue that would be implied by AGI clearly reveals that my previous definition of AGI was too narrow. And I expect this to keep happening into the future.
I expect that by 2030, the labs will have made significant progress on my hobby horse of continual learning, and the models will be earning hundreds of billions of dollars in revenue a year, but they won’t have automated all knowledge work. And I’ll be like, “Look, we made a lot of progress, but we haven’t hit AGI yet. We also need these other capabilities. We need X, Y, and Z capabilities in these models.”
5. RL scaling
Models keep getting more impressive at the rate that the short-timeline people predict, but more useful at the rate that the long-timeline people predict. It’s worth asking: What are we scaling with pretraining?
We had this extremely clean and general trend in improvement in loss across multiple orders of magnitude in compute. That was on a power law, which is as weak as exponential growth is strong. But people are trying to launder the prestige that pretraining scaling has—which is almost as predictable as a physical law of the universe—to justify bullish predictions about reinforcement learning from verifiable reward, for which we have no publicly known trend.
When intrepid researchers do try to piece together the implications from scarce public data points, they get pretty bearish results. For example, Toby Board has a great post where he cleverly connects the dots between the different o-series benchmarks, and this suggested to him that:
“We need something like a 1,000,000× scale-up in total RL compute to give a boost similar to a single GPT-level model.”
6. Broadly deployed intelligence explosion
End quote. People have spent a lot of time talking about the possibility of a software singularity, where AI models will write the code that generates a smarter successor system, or a software-plus-hardware singularity, where AIs also improve their successor's computing hardware. However, all these scenarios neglect what I think will be the main driver of further improvements atop AGI: continual learning. Again, think about how humans become more capable than anything else. It's mostly from experience in the relevant domain.
In our conversation, Baron Miller made this interesting suggestion: The future might look like continual-learning agents who are all going out, doing different jobs, generating value, and then bringing back all their learnings to the hive-mind model, which does some kind of batch distillation on all of these agents. The agents themselves could be quite specialized, containing what Karpathi called the cognitive core, plus knowledge and skills relevant to the job they're being deployed to do.
Solving continual learning won't be a singular, one-and-done achievement. Instead, it will feel like solving in-context learning. GBT3 already demonstrated that in-context learning could be very powerful in 2020. Its in-context-learning capabilities were so remarkable that the title of the GPT3 paper was “Language Models are Few-Shot Learners.” But of course, we didn't solve in-context learning when GPD3 came out. Indeed, there's still plenty of progress that has to be made, from comprehension to context length.
I expect a similar progression with continual learning. Labs will probably release something next year which they call continual learning and which will, in fact, count as progress toward continual learning. But human-level, on-the-job learning may take another 5 to 10 years to iron out. This is why I don't expect runaway gains from the first model that cracks continual learning and gets more and more widely deployed and capable.
If you had fully solved continual learning and it dropped out of nowhere, then sure, it might be game, set, match, as [Speaker?] put it on the podcast when I asked him about this possibility. But that's probably not what's going to happen. Instead, some lab is going to figure out how to get some initial traction on this problem, and then playing around with this feature will make it clear how it was implemented. Other labs will soon replicate the breakthrough and improve it slightly.
Besides, I just have some prior that the competition will stay pretty fierce between all these model companies. This is informed by the observation that all these previous supposed flywheels, whether that's user engagement on ChatGPT or synthetic data or whatever, have done very little to diminish the greater and greater competition between model companies. Every month or so, the big 3 model companies will rotate around the podium, and the other competitors are not that far behind. There seems to be some force—potentially talent poaching, the rumor mill, or just normal reverse engineering—that has so far neutralized any runaway advantage that a single lab might have had.