Dwarkesh and Ilya Sutskever on What Comes After Scaling

Summary

Podcast Summary: The a16z Show – “Dwarkesh and Ilya Sutskever on What Comes After Scaling”

Date: December 15, 2025
Host: Andreessen Horowitz
Guests: Dwarkesh Patel (C), Ilya Sutskever (A), Narrator (B)

---

Overview

This episode explores the current and future challenges in artificial intelligence (AI) and the path toward artificial general intelligence (AGI), through a rare, in-depth conversation with Ilya Sutskever (co-founder of SSI, former co-founder of OpenAI), hosted by Dwarkesh Patel. The discussion focuses on why AI progress in benchmarks often outpaces real-world impact, the limits of “scaling” compute and data, the role of human-like generalization and continual learning, and how these themes affect both economic transformation and AI safety. Sutskever offers candid reflections on the shifting research paradigms in AI, the analogy between machine and human learning, strategic choices for leading AI labs, and his vision for a responsible superintelligent future.

---

Key Themes and Insights

---

I. The Disconnect Between AI Benchmarks and Real-World Performance

(05:00 – 10:48)

• Performance Gap: AI models excel at benchmarks but struggle with practical, robust applications.

  • "They're doing so well on evals... But the economic impact seems to be dramatically behind... It's very difficult to make sense of." (A, 02:47)

• Potential Causes

  • Overfitting to RL environments inspired by benchmarks rather than true generalization.
  • Pretraining utilizes massive data, masking generalization weaknesses.

• Notable Example:

  • Sutskever highlights models that can alternate between creating and fixing bugs in code, exposing lack of real-world robustness.

Memorable Quote:
"Models are much more like the first student [who overpracticed], but even more... Intuitively, with this level of preparation it will not necessarily generalize to other things." (A, 08:49)

---

II. Comparing Machine and Human Learning

(10:48 – 18:10)

• Data Efficiency:

  • Pretraining uses staggering data volume; humans learn deeply from little data.
  • Pretraining lacks a direct human analogue, as humans build deep, versatile understanding with fraction of the data.

• Evolution vs. Pretraining:

  • Evolution imparts deeply ingrained priors (e.g. dexterity, vision).
  • Human emotional “value functions” may hardwire learning efficiency and decision-making.

• Value Functions and Emotions:

  • In reinforcement learning, value functions should help—but in humans, emotions might modulate them in ways AI lacks.
  • Suggests ML may require emotional analogues for robustness and efficient learning.

Notable Quote:
"Maybe what [brain injuries] suggest is that the value function of humans is modulated by emotions in some important way that's hard coded by evolution, and maybe that is important for people to be effective in the world." (A, 17:01)

---

III. The Plateau of Scaling and Return to Research

(20:10 – 26:22)

• Era Shifts:

  • 2012–2020: “Age of Research”
  • 2020–2025: “Age of Scaling” through pretraining and compute
  • 2025–future: “Back to Age of Research”, with massively scaled compute but diminishing returns from further scaling alone.

• Limits of Scaling:

  • Pretraining will exhaust useful data.
  • Future advances will rely not on bigger models, but on new research (better learning algorithms, generalization, continual learning).

Memorable Quote:
"Now the scale is so big. Like is the belief really that oh, it's so big, but if you had 100x more, everything would Be so different... I don't think that's true. So it's back to the age of research again, just with big computers." (A, 22:56)

---

IV. Generalization, Sample Efficiency, and Continual Learning

(25:45 – 33:36)

• Sample Efficiency Gap:

  • Humans learn tasks (like driving) quickly and robustly, even in unfamiliar domains.
  • AI requires orders of magnitude more data for similar performance.

• Fundamental Bottleneck:

  • Sutskever: “These models somehow just generalize dramatically worse than people. And it's super obvious.” (A, 25:45)

• Continual Learning:

  • Humans continually learn on the job, without explicit reward signals for every action.
  • Sutskever suggests that AI needs to mirror this with robust value functions to self-correct and learn efficiently.

• Sutskever’s Secret Hunch:

  • He hints at a yet-undiscussed principle to achieve human-like continual learning, but withholds details due to competitive sensitivity.

Memorable Exchange:

• C: "How do we need to reconceptualize the way we're training models...?"
• A: "... I have a lot of opinions about. But unfortunately, we live in a world where not all machine learning ideas are discussed freely, and this is one of them." (32:36)

---

V. Changing Dynamics in AI Research and Competition

(34:34 – 40:39)

• From Homogeneity to Creative Research:

  • During scaling era, "all the air in the room" was sucked out by compute-focused approaches.
  • Now, new ideas and research efficiency become the bottleneck, not just compute.

• SSI’s Research Positioning:

  • Despite less funding than OpenAI/Google, enough compute is available for inventive research, as much large-company compute is tied up in inference and product support.

Notable Quote:
"We are in a world where there are more companies than ideas by quite a bit... If ideas are so cheap, how come no one's having any ideas?" (A, 34:46)

---

VI. Strategies and Ethics for Superintelligence and Safety

(41:00 – 64:38)

• SSI’s “Straight Shot” vs. Gradual Release

  • Sutskever open to both direct pursuit of superintelligence and gradual deployment.
  • Gradualism important for safety and society’s adaptation (“showing the thing” matters).

• AGI and Pretraining as Imprints

  • The frame of AGI as a pre-trained, all-knowing mind is an overshoot.
  • Sutskever: Human “AGI” is actually continual learners—deployment should mirror this.

• Societal and Economic Impacts

  • Once AI can match human sample efficiency and diversity, massive economic growth and disruption will ensue.
  • Different nations' regulatory approaches could accelerate divergence.

• Alignment and “Caring for Sentient Life”

  • Sutskever’s increasingly favored vision: AIs aligned to care for all “sentient life”—may be easier and more organic than aligning to “humanity” specifically due to emergent empathy (mirror neurons).
  • Even if achieved, most future sentient beings will be AIs; thus alignment goals must account for vast numbers of non-human minds.

Notable Quotes

• “The whole problem is the power. The whole problem is the power. When the power is really big, what's going to happen?" (A, 52:50)
• "There has been one big idea that actually everyone has been locked into, which is the self improving AI... I maintain that there is something that's better to build... AI that's robustly aligned to care about sentient life." (A, 55:00)

---

VII. Uncertainties, Strategic Convergence, and Forecasts

(73:57 – 80:34)

• SSI’s Distinction:

  • Technical focus: novel approaches to generalization and continual learning, as opposed to business model or compute supremacy.
  • Expects convergence in strategies as true solutions become apparent and AGI capabilities change incentives for labs and governments.

• Market Dynamics:

  • Despite theoretical winner-take-all (first to build a truly human-level continual learner), Sutskever expresses strong intuition that real-world diversity, specialization, and competition among labs will persist.

• Timeline Forecast:

  • Human-efficient, superhuman continual learners: "five to twenty years." (77:46)

---

VIII. Diversity, Self-Play, and the Nature of Research Taste

(84:20 – End)

• LLM Sameness:

  • Current LLMs are similar because of overlapping pretraining; RL and post-training may allow more agentic diversity.
  • Self-play and adversarial settings (debate, prover-verifier) as routes to richer skill development and diversity.

• Research Taste (Philosophy of AI Progress)

  • Sutskever’s guiding philosophy: drawn from beauty, simplicity, and correct inspiration from the brain.
  • “It's just beauty, simplicity, elegance, correct inspiration from the brain... The more they are present, the more confident you can be in a top down belief... That's what sustains you when experiments contradict you.” (A, 88:34-90:52)

---

Notable Quotes (with Timestamps)

• On Robustness Gaps:
  “The models are much more like the first student...but with this level of preparation it will not necessarily generalize to other things.” (08:49)

• On Pretraining:
  “The main strength of pre training is that there is A, so much of it. And B, you don't have to think hard about what data to put into pre training.” (09:40)

• On the Next Era:
  “Now the scale is so big...is the belief that if you just 100x the scale, everything would be transformed? I don't think that's true. So it's back to the age of research again, just with big computers.” (22:56)

• On AI Safety & Alignment:
  “The AI that's robustly aligned to care about sentient life specifically...it will be easier to build an AI that cares about sentient life than an AI that cares about human life alone, because the AI itself will be sentient.” (55:00)

• On Strategic Convergence:
  “I maintain that in the end there will be a convergence of strategies...as AI becomes more powerful, it's going to become more or less clearer to everyone what the strategy should be.” (76:11)

---

Timestamps for Major Segments

• Intro & Framing: 00:00–01:55
• Real-World AI vs Benchmark Gaps: 01:55–09:40
• Human vs Model Learning, Value Functions: 09:40–18:10
• Scaling and the Age of Research: 20:10–26:22
• Generalization & Continual Learning: 25:45–33:36
• Paradigm Shifts in Research/Competition: 34:34–40:39
• SSI’s Strategy, Alignment, and Superintelligence: 41:00–64:38
• Governance, Specialization, Forecasts: 73:57–80:34
• Diversity, Self-play, Research Philosophy: 84:20–90:52

---

Conclusion

This episode captures the excitement and uncertainty at the cutting edge of AI, as leading minds like Sutskever anticipate a transition from brute scaling to deeper insights into learning, generalization, and alignment. The path forward, in his view, will be determined not just by resources, but by original research, principled philosophical grounding, and careful adaptation as AI’s impact on society accelerates.

Transcript

[A]

Now that compute is big, computer is now very big. In some sense we are back to the age of research. We got to the point where we are in a world where there are more companies than ideas by quite a bit. Now there is this Silicon Valley saying that says that ideas are cheap, execution is everything. What is the problem of AI and AGI? The whole problem is the power.

[B]

AI models look incredibly smart on bedroom, yet their real world performance often feels far behind. Why is there such a gap and what does that say about the path to AGI? From the Dwarkesh podcast, here's a rare long form conversation with Ilya Sutskever, co founder of ssi, exploring what's actually slowing down progress toward AGI. Dwarkesh and Ilya dig into the core problems in modern AI systems from why RL and pre training scale so differently to why generalization, reliability and sample efficiency still fall short of human learning. They also explore continual learning, value functions, superintelligence, and what a future economy shaped by AI might look like.

[A]

You know what's crazy? That all of this is real.

[C]

Yeah.

[A]

Don't you think so? Like all this AI stuff and all this Bay Area. Yeah. That it's happened. Like isn't it straight out of science fiction?

[C]

Yeah. Another thing that's crazy is like how normal the slow takeoff feels. The idea that we'd be investing 1% of GDP in AI. I feel like it would have felt like a bigger deal where right now.

[A]

It just feels like we get used to things pretty fast. Turns out.

[C]

Yeah.

[A]

But also it's kind of like it's abstract. Like what does it mean? What? It means that you see it in the news that such and such company announced such and such dollar amount.

[C]

Right.

[A]

That's all you see.

[C]

Right.

[A]

It's not really felt in any other way so far.

[C]

Should we actually begin here? I think this is an interesting discussion.

[A]

Sure.

[C]

I think your point about. Well, from the average person's point of view, nothing is that different will continue being true even into the singularity.

[A]

No, I don't think so.

[C]

Okay, interesting.

[A]

So the thing which I was referring to not feeling different is. Okay, so such and such company announced some difficult to comprehend dollar amount of investment. I don't think anyone knows what to do with that. But I think that the impact of AI is going to be felt. AI is going to be diffused through the economy. There are very strong economic forces for this and I think the impact is going to be felt very strongly.

[C]

When do you expect that impact? I think the models seem smarter than their Economic impact would imply.

[A]

Yeah, this is one of the very confusing things about the models right now. How to reconcile the fact that they are doing so well on evals. And you look at the evals and you go, those are pretty hard evals. They are doing so well. But the economic impact seems to be dramatically behind. And it's almost like it's very difficult to make sense of. How can the model on the one hand do these amazing things and then on the other hand repeat itself twice in some situation in a kind of a. An example would be, let's say you use wipe coding to do something and you go to some place and then you get a bug and then you tell the model, can you please fix the bug? And the model says, oh my God, you're so right, I have a bug, let me go fix that. And it introduces a second bug and then you tell it you have this new second bug and it tells you, oh my God, how could I have done it? You're so right again. And brings back the first bug and you can alternate between those. And it's like, how is that possible? Yeah, it's like, I'm not sure, but it does suggest that something strange is going on. I have two possible explanations. So here this is the more kind of a whimsical explanation is that maybe RL training makes the models a little bit too single minded and narrowly focused, a little bit too, I don't know, unaware. Even though it also makes them aware in some other ways. And because of this they can't do basic things. But there is another explanation which is back when people were doing pre training, the question of what data to train on was answered because that answer was everything. When you do pre training, you need all the data so you don't have to think it's going to be this data or that data. But when people do RL training, they do need to think. They say, okay, we want to have this kind of RL training for this thing and that kind of RL training for that thing. And from what I hear, all the companies have teams that just produce new RL environments and just add it to the training mix. And then the question is, well, what are those? There are so many degrees of freedom. There is such a huge variety of RL environments you could produce. And one of the one thing you could do, and I think that's something that is done inadvertently, is that people take inspiration from the evals. You say, hey, I would love our model to do really well when we release it. I want the Evals to look great. What would be RL training that could help on this task? Right. I think that is something that happens and I think it could explain a lot of what's going on. If you combine this with generalization of the models actually being inadequate, that has the potential to explain a lot of what we are seeing. This disconnect between eval performance and actual real world performance, which is something that we don't today exactly even understand what we mean by that.

[C]

I like this idea that the real reward hacking is the human researchers who are too focused on devals. I think there's two ways to understand or to try to think about what you have just pointed out. One one is, look, if it's the case that simply by becoming superhuman at a coding competition, a model will not automatically become more tasteful and exercise better judgment about how to improve your code base, well then you should expand the suite of environments such that you're not just testing it on having the best performance in a coding competition. It should also be able to make the best kind of application for X thing or Y thing or Z thing. And another maybe this is what you're hinting at is to say why should it be the case in the first place that becoming superhuman at coding competitions doesn't make you a more tasteful programmer more generally? Maybe the thing to do is not to keep stacking up the amount of environments and the diversity of environments to figure out approach which lets you learn from one environment and improve your performance on something else.

[A]

So I have a human analogy which might be helpful. So even the case, let's take the case of competitive programming, since you mentioned that, and suppose you have two students. One of them decided they want to be the best competitive programmer, so they will practice 10,000 hours for that domain. They will solve all the problems, memorize all the proof techniques, and be very, very, you know, be very skilled at quickly and correctly implementing all the algorithms. And by doing, by doing so, they became the best one of the best. Student number two thought, oh, competitive programming is cool. Maybe they practiced for a hundred hours, much, much less. And they also did really well. Which one do you think is going to do better in their career later on?

[C]

The second, right.

[A]

And I think that's basically what's going on. The models are much more like the first student, but even more because then we say, okay, so the model should be good at competitive programming. So let's get every single competitive programming problem ever and then let's do some data augmentation. So we have even More competitive programming problems. And we train on that. And so now you got this great competitive programmer. And with this analogy, I think it's more intuitive. I think it's more intuitive with this analogy that. Yeah, okay, so if it's so well trained, okay, it's like all the different algorithms and all the different proof techniques are like right at its fingertips. And it's more intuitive that with this level of preparation it will not necessarily generalize to other things.

[C]

But then what is the analogy for what the second student is doing before they do the hundred hours of fine tuning?

[A]

I think it's like they have it. I think it's the it factor. Yeah. Right. And like, I know, like when I was an undergrad, I remember there was, there was a student like this that studied with me. So I know it exists.

[C]

I think it's interesting to distinguish it from whatever pre training does. So one way to understand what you just said about we don't have to choose the data in pre training is to say actually it's not dissimilar to the 10,000 hours of practice. It's just that you get that 10,000 hours of practice for free because it's already somewhere in the pre training distribution. But it's like maybe you're suggesting actually there's actually not that much generalization for pre training. There's just so much data in pre training. Everybody's like, it's not necessarily generalizing better than rl.

[A]

Like the main, the main strength of pre training is that there is A, so much of it.

[C]

Yeah.

[A]

And B, you don't have to think hard about what data to put into pre training. And it's a very kind of natural data and it does include in it a lot of what people do.

[C]

Yeah.

[A]

People's thoughts and a lot of the features. You know, it's like the whole world as projected by people onto text.

[C]

Yeah.

[A]

And pre training tries to capture that using a huge amount of data. It's very, the pre training is very difficult to reason about because it's so hard to understand the manner in which the model relies on pre training data. And whenever the model makes a mistake, could it be because something by chance is not as supported by the pre training data? You know, and support by pre training is maybe a loose term. I don't know if I can add anything more useful on this. But I don't think there is a human analog to pre training.

[C]

Here's analogies that people have proposed for what the human analogy to pre training is. And I'm curious to get your thoughts on why they're potentially wrong. One is to think about the first 18 or 15 or 13 years of a person's life when they aren't necessarily economically productive, but they are doing something that is making them understand the world better and so forth. And the other is to think about evolution as doing some kind of search for 3 billion years which then results in a human lifetime instance. And then I'm curious if you think either of these are actually analogous to pre training or how would you think about at least what lifetime human learning is like if not pre training?

[A]

I think there are some similarities between both of these two. Pre training and pre training tries to play the role of both of these. But I think there are some big differences as well. The amount of pre training data is very, very staggering.

[C]

Yes.

[A]

And somehow a human being after even 15 years with a tiny fraction of that pre training data, they know much less. But whatever they do know, they know much more deeply somehow. And the mistakes, like already at that age, you would not make mistakes that Reis make. There is another thing you might say, could it be something like evolution? And the answer is maybe. But in this case, I think evolution might actually have an edge. There is this, I remember reading about this case where some, you know that one thing that neuroscientists do, or rather one way in which neuroscientists can learn about the brain is by studying people with brain damage to different parts of the brain. And so, and some people have the most strange symptoms you could imagine. It's actually really, really interesting. And there was one case that comes to mind that's relevant. I read about this person who had some kind of brain damage that took, took out, I think a stroke or an accident that took out his emotional processing. So he stopped feeling any emotion. And as a result of that, you know, he still remained very articulate and he could solve little puzzles. And on tests he seemed to be just fine, but he felt no emotion. He didn't feel sad, he didn't feel anger, he didn't feel animated. And he became somehow extremely bad at making any decisions at all. It would take him hours to decide on which socks to wear and he would make very bad financial decisions. And that's very. What does it say about the role of our built in emotions in making us like a viable agent, essentially? And I guess to connect to your question about pre training, like maybe pre, like maybe if you are good enough at like getting everything out of pre training, you can get, you, you could get that as well. But that's the kind of thing which seems. Well, it may or may not be possible to get that from pre training.

[C]

What is that? Clearly not just directly emotion. It seems like some almost value function like thing which is giving. Telling you which decision to be like what the end reward for any decision should be. And you think that doesn't sort of implicitly come from.

[A]

I think it could. I'm just saying it's not one. It's not 100% obvious.

[C]

Yeah, but what is that like, how do you think about emotions? What is the ML analogy for emotions?

[A]

It should be some kind of a value function thing. Yeah, but I don't think there is a great ML analogy because right now, value functions don't play a very prominent role in the things people do.

[C]

It might be worth defining for the audience what a value function is if you want to do that.

[A]

I mean, certainly I'll be very happy to do that. Right. So. When people do reinforcement learning, the way reinforcement learning is done right now, how do people train those agents? So you have your neural net and you give it a problem, and then you tell the model, go solve it. The model takes maybe thousands, hundreds of thousands of actions or thoughts or something, and then it produces a solution. The solution is created, and then the score is used to provide a training signal for every single action in your trajectory. So that means that if you are doing something that goes for a long time, if you're training a task that takes a long time to solve, you will do no learning at all until you came up with a proposed solution. That's how reinforcement learning is done. Naively, that's how O1, R1 ostensibly are done. The value function says something like, okay, look, maybe I could sometimes, not always could tell you if you are doing well or badly. The notion of a value function is more useful in some domains than others. So, for example, when you play chess and you lose a piece, you know, I messed up. You don't need to play the whole game to know that what I just did was bad. And therefore whatever preceded it was also bad. So the value function lets you short circuit the wait until the very end. Like, let's suppose that you started to pursue some kind of. Okay, let's suppose that you are doing some kind of a math thing or a programming thing, and you're trying to explore a particular solution direction. And after, let's say after a thousand steps of thinking, you concluded that this direction is unpromising. As soon as you conclude this, you could already get a reward signal. A thousand time steps previously when you decided to pursue down this path, you say, oh, next time I shouldn't pursue this path. In a similar situation, long before you actually came up with a proposed solution.

[C]

This was in the Deepseek R1 paper is that the space of trajectories is so wide that maybe it's hard to learn a mapping from an intermediate trajectory and value. And also given that, you know, in coding, for example, you'll have the wrong idea, then you'll go back, then you'll change something.

[A]

This sounds like such lack of faith in deep learning. Like, I mean sure it might be difficult, but nothing deep learning can do. Yeah, so my expectation is that like value function should be useful and I fully expect that they will be used in the future, if not already, what was I alluding to with the person whose emotional center got damaged is more that maybe what it suggests is that the value function of humans is modulated by emotions in some important way that's hard coded by evolution and maybe that is important for people to be effective in the world.

[C]

That's the thing I was actually planning on asking you. There's something really interesting about emotions of the value function, which is that it's impressive that they have this much utility while still being rather simple to understand.

[A]

So I have two responses. I do agree that compared to the kind of things that we learn and the things that we are talking about, the kind of AIs we are talking about, emotions are relatively simple. They might even be so simple that maybe you could map them out in a human understandable way. I think it would be cool to do. In terms of utility though, I think there is a thing where, you know, there is this complexity, robustness, trade off, where complex things can be very useful, but simple things are very useful in very broad range of situations. And so I think one way to interpret what we are seeing is that we've got these emotions that essentially evolved mostly from our mammal ancestors and then fine tuned a little bit while we were hominids, just a bit. We do have like a decent amount of social emotions though, which mammals may lack. But they're not very sophisticated. And because they're not sophisticated, they serve us so well in this very different world compared to the one that we've been living in. Actually they also make mistakes. For example our emotions, Well, I don't know, does hunger count as an emotion? It's debatable, but I think for example, our intuitive feeling of hunger is not succeeding in guiding us correctly in this world with an abundance of food.

[C]

Yeah, people have been talking about scaling data, scaling parameters, scaling compute. Is there a more general way to think about scaling? What are the other scaling axes?

[A]

So the thing, so here is a perspective, here's a perspective I think might be true. So the way ML used to work is that people would just think with stuff and try to, and try to get interesting results. That's what's been going on in the past. Then the scaling insight arrived, right? Scaling laws, GPT3 and suddenly everyone realized we should scale. And it's just this, this is an example of, of how language affects thought. Scaling is just one word, but it's such a powerful word because it informs people what to do. They say, okay, let's try to scale things. And so you say, okay, so what are we scaling? And pre training was a thing to scale. It was a particular scaling recipe.

[C]

Yes.

[A]

The big breakthrough of pre training is the realization that this recipe is good. So you say, hey, if you mix some compute with some data into a neural net of a certain size, you will get results and you will know that you'll be better if you just scale the recipe up. And this is also great. Companies love this because it gives you a very low risk way of investing your resources. It's much harder to invest your resources in research. Compare that, you know, if you research. So you need to have like go forth researchers and research and come up with something versus get more data, get more compute. You know, you'll get something from pre training. And indeed, you know, it looks like I based on various things people say on some people say on Twitter maybe it appears that Gemini have found a way to get more out of pre training. At some point though, pre training will run out of data. The data is very clearly finite. And so then, okay, what do you do next? Either you do some kind of a souped up pre training, different recipe from the one you've done before, or you're doing RL or maybe something else. But now that compute is big, computers now very big in some sense we are back to the age of research. So maybe here's another way to put it up until 2020. From 2012 to 2020 it was the age of research. Now from 2020 to 2025 it was the age of scaling or maybe plus minus. Let's add error bars to those years because people say this is amazing. You gotta scale more. Keep scaling the one word, scaling. But now the scale is so big. Like is the belief really that oh, it's so big, but if you had 100x more, everything would Be so different. Like, it would be different for sure. But like, is the belief that if you just honeydex the scale, everything would be transformed. I don't think that's true. So it's back to the age of research again, just with big computers.

[C]

That's a very interesting way to put it. But let me ask you the question you just posed then. What are we scaling and what would it mean to have a recipe? Because I guess I'm not aware of a very clean relationship that almost looks like a law of physics which existed in pre training. There was a power law between data or computer parameters and loss. What is the kind of relationship we should be seeking and how should we think about what this new recipe might look like?

[A]

So we've already witnessed a transition from one type of scaling to a different type of scaling from pre training to rl. Now, people are scaling RL now, based on what people say on Twitter, they spend more compute on RL than on pre training at this point. Because RL can actually consume quite a bit of compute. You know, you do very, very long rollouts.

[C]

Yes.

[A]

So it takes a lot of compute to produce those rollouts. And then you get relatively small amount of learning per rollout. So you really can spend a lot of computer. And I could imagine, like, I wouldn't at this, at this stage, it's more like, I wouldn't even call it a scaling. I would say, hey, like, what are you doing? And is the thing you are doing the most productive thing you could be doing? Can you find a more productive way of using your compute? We've discussed the value function business earlier, and maybe once people get good at value functions, they will be using their resources more productively. And if you find a whole other way of training models, you could say, is this scaling or is it just using your resources? I think it becomes a little bit ambiguous in a sense that when people were in the age of research, back then it was like people say, hey, let's try this and this and this, let's try that and that and that. Oh, look, something interesting is happening. And I think there will be a return to that.

[C]

So if we're back in the era of research, stepping back, what is the part of the recipe that we need to think most about? When you say value function, people are already trying the current recipe. But then having LLM as a judge and so forth, you could say, that's a value function. But it sounds like you have something much more fundamental in mind. Do we need to go back to, should we Even rethink pre training at all and not just add more steps to the end of that process.

[A]

Yeah, so the discussion about value function, I think it was interesting. I want to emphasize that I think the value function is something like it's going to make RL more efficient. And I think that makes a difference. But I think that anything you can do with a value function you can do without just more slowly. The thing which I think is the most fundamental is that these models somehow just generalize dramatically worse than people. And it's super obvious. That seems like a very fundamental thing.

[C]

Okay, so this is the crux generalization and there's two sub questions. There's one which is about sample efficiency, which is why should it take so much more data for these models to learn than humans? There's a second about even separate from the amount of data it takes, there's a question of why is it so hard to teach the thing we want to a model than to a human? Which is to say, to a human, we don't necessarily need a verifiable reward to be able to. You're probably mentoring a bunch of researchers right now, and you're talking with them, you're showing them your code and you're showing them how you think, and from that they're picking up your way of thinking and how they should do research. You don't have to set a verifiable reward for them. That's like, okay, this is the next part of the curriculum and now this is the next part of your curriculum. And oh, this training was unstable and we got. There's not this schleppy bespoke process. So perhaps these two issues are actually related in some way. But I'd be curious to explore this second thing which feels more like continual learning and this first thing which feels just like sample efficiency.

[A]

Yeah. So, you know, you could actually wonder. One possible explanation for the human sample efficiency that needs to be considered is evolution. And evolution has given us a small amount of the most useful information possible. And for things like vision, hearing and locomotion, I think there's a pretty strong case that evolution actually has given us a lot. So, for example, human dexterity far exceeds amin. Robots can become dexterous too, if you subject them to a huge amount of training and simulation. But to train a robot in the real world to quickly pick up a new skill like a person does, seems very out of reach. And here you could say, oh yeah, locomotion. All our ancestors needed great locomotion. Squirrels. So locomotion, maybe you've got some unbelievable prior. You could make the same case for vision. I believe Yann Lecan made the point. Oh, like children learn to drive after 16 hours, after 10 hours of practice, which is true, but our vision is so good, at least for me. When I remember myself being 5 year old. I was very excited about cars back then and I'm pretty sure my car recognition was more than adequate for self driving already. As a five year old you don't get to see that much data. As a five year old you spend most of your time in your parents house, so you have very low data diversity. But you could say maybe that's evolution too. But in language and math and coding, probably not.

[C]

It still seems better than models. I mean obviously models are better than the average human at language and math encoding. But are they better at the average human at learning?

[A]

Oh yeah. Oh yeah, absolutely. What I meant to say is that language, math encoding and especially math and coding suggests that whatever it is that makes people good at learning is probably not so much a complicated prior, but something more some fundamental thing.

[C]

Wait, I'm not sure I understood. Why should that be the case?

[A]

So consider a skill that people exhibit, some kind of great reliability or you know, if the skill is one that was very useful to our ancestors for many millions of years, hundreds of millions of years, you could say, you could argue that maybe humans are good at it because of evolution, because we have a prior, an evolutionary prior that's encoded in some very non obvious way that somehow makes us so good at it. But if people exhibit great ability, reliability, robustness, ability to learn in a domain that really did not exist until recently, then this is more an indication that people might have just better machine learning period.

[C]

But then how should we think about what that is? Is it a matter of. Yeah, what is the ML analogy for? There's a couple interesting things about it. It takes fewer samples, it's more unsupervised. You don't have to set a very. Like a child learning to drive a car. Children are not learning to drive a car. A teenager learning how to drive a car is like not exactly getting some pre built verifiable reward. It comes from their interaction with the machine and with the environment. And yeah, it takes much fewer samples. It seems more unsupervised, it seems more robust.

[A]

Much more robust. The robustness of people is really staggering.

[C]

Yeah. Okay. And do you have a unified way of thinking about why are all these things happening at once? What is the ML analogy that could realize something like this?

[A]

So this is where one of the things that you've been asking about is how can the teenage driver kind of self correct and learn from their experience without an external teacher? And the answer is, well, they have their value function, right? They have a general sense, which is also, by the way, extremely robust in people. Like, whatever it is the human value function, whatever the human value function is, with a few exceptions around addiction, it's actually very, very robust. And so for something like a teenager that's learning to drive, they start to drive and they already have a sense of how they're driving immediately, how badly they're unconfident. And then they see, okay. And then, of course, the learning speed of any teenager is so fast, after 10 hours, you're good to go.

[C]

Yeah. It seems like humans have some solution, but I'm curious about, well, how are they doing it and why is it so hard to. How do we need to reconceptualize the way we're training models to make something like this possible?

[A]

You know, that is a great question to ask, and it's a question I have a lot of opinions about. But unfortunately, we live in a world where not all machine learning ideas are discussed freely, and this is one of them. So there's probably a way to do it. I think it can be done. The fact that people are like that, I think it's a proof that it can be done. There may be another blocker though, which is there is a possibility that the human neurons actually do more compute than we think. And if that is true, and if that plays an important role, then things might be more difficult. But regardless, I do think it points to the existence of some machine learning principle that I have opinions on, but unfortunately, circumstances make it hard to discuss.

[C]

In detail, even though nobody listens to this podcast. Ilya.

[A]

Yeah.

[C]

I am curious if you say we are back in an era of research. You were there from 2012 to 2020 and do you have. Yeah. What is now the vibe going to be if we go back to the era of research? For example, even after Alexnet, the amount of compute that was used to run experiments kept increasing and the size of frontier systems kept increasing. And do you think now that this era of research will still require tremendous amounts of compute? Do you think it will require going back into the archives and reading old papers? What is. Maybe what was the vibe of like you were at Google and OpenAI and Stanford, these places, when there was like a. More of a vibe of research. What kind of things should we be expecting in the community?

[A]

So One consequence of the age of scaling is that there was this scaling sucked out all the air in the room.

[C]

Yeah.

[A]

And so because scaling sucked out all the air in the room, everyone started to do the same thing. We got to the point where we are in a world where there are more companies than ideas by quite a bit actually on that. You know, there is this Silicon Valley saying that says that ideas are cheap, execution is everything. And people say that a lot. And there is truth to that. But then I saw someone say on Twitter something like, if ideas are so cheap, how come no one's having any ideas? And I think it's true too. I think if you think about research progress in terms of bottlenecks, there are several bottlenecks if you go back to the and one of them is ideas, and one of them is your ability to bring them to life, which might be compute, but also engineering. So if you go back to the 90s, let's say you had people who had pretty good ideas and if they had much larger computers, maybe they could demonstrate that their ideas were viable. But they could not. So they could only have very, very small demonstration that did not convince anyone. So the bottleneck was compute. Then in the age of scaling, compute has increased a lot. And of course there is a question of how much compute is needed. But compute is large. So compute is large enough such that it's not obvious that you need that much more compute to prove some idea. Like I'll give you an analogy. Alexnet was built on two GPUs. That was the total amount of compute used for it. The transformer was built on 8 to 64 GPUs. No single transformer paper experiment used more than 64 GPUs of 2017, which would be like what, two GPUs of today. So the Resnet. Right. Many like even the you could argue that the like O1 reasoning was not the most compute heavy thing in the world. So there definitely for research you need definitely some amount of compute. But it's far from obvious that you need the absolutely largest amount of compute ever for research. You might argue, and I think it is true, that if you want to build the absolutely best system, if you want to build the absolutely best system, then it helps to have much more compute. And especially if everyone is within the same paradigm, then compute becomes one of the big differentiators.

[C]

Yeah, I guess while it was possible to develop these ideas, I'm asking you for the history, because you were actually there. I'm not sure what actually happened, but it Sounds like it was possible to develop these ideas using minimal amounts of compute. But the transformer didn't immediately become famous. It became the thing everybody started doing and then started experimenting on top of and building on top of. Because it was validated at higher and higher levels of compute.

[A]

Correct.

[C]

And if you at SSI have 50 different ideas, how will you know which one is the next transformer and which one is brittle without having the kinds of compute that other Frontier labs have?

[A]

So I can comment on that. Which is the short comment is that, you know, you mentioned SSI specifically for us. The amount of compute that SSI has for research is really not that small. And I want to explain why, like a simple math can explain why the amount of compute that we have is actually a lot more comparable for research than one might think. I'll explain. So SSI has raised $3 billion, which is like not small, but it's like a lot by any absolute sense. But you could say, but look at the other companies raising much more. But a lot of their compute goes for inference. Like these big numbers, these big loans, it's earmarked for inference. That's number one. Number two, you need, if you want to have a product on which you do inference, you need to have a big staff of engineers, of salespeople. A lot of the research needs to be dedicated for producing all kinds of product related features. So then when you look at what's actually left for research, the difference becomes a lot smaller. Now the other thing is, is that if you are doing something different, do you really need the absolute maximal scale to prove it? I don't think that's true at all. I think that in our case we have sufficient compute to prove, to convince ourselves and anyone else that what we're doing is correct.

[C]

There's been public estimates that, you know, companies like OpenAI spend on the order of 5, 6 billion dollars a year, even just so far on experiments. This is separate from the amount of money they're spending on inference and so forth. So it seems like they're spending more a year running research experiments than you guys have in total funding.

[A]

I think it's a question of what you do with it. It's a question of what you do with it. Like they have like the more. I think in their case, in the case of others, I think there is a lot more demand on the training compute. There's a lot more different work streams, there are different modalities, there is just more stuff. And so it becomes fragmented.

[C]

How will SSI make money?

[A]

You Know, my answer to this question is something like right now we just focus on the research and then the answer to this question will reveal itself. I think there will be lots of possible answers.

[C]

Is SSI's plan still to straight shot superintelligence?

[A]

Maybe. I think that there is merit to it. I think there's a lot of merit because I think that it's very nice to not be affected by the day to day market competition. But I think there are two reasons that may cause us to change the plan. One is pragmatic. If timelines turned out to be long, which they might. And second, I think there is a lot of value in the best and most powerful AI being out there impacting the world. I think this is a meaningfully valuable thing.

[C]

But then, so why is your default plan to straight shot superintelligence? Because it sounds like OpenAI anthropic. All these other companies, their explicit thinking is, look, we have weaker and weaker intelligences that the public can get used to and prepare for. And why is it potentially better to build a super intelligence directly?

[A]

So I'll make the case for and against. The case for is that you are. So one of the challenges that people face when they're in the market is that they have to participate in the rat race. And the rat race is quite difficult in that it exposes you to two to difficult trade offs which you need to make. And there is, it is, it is nice to say we'll insulate ourselves from all this and just focus on the research and come out only when we are ready and not before. But the counterpoint is valid too. And those are, those are opposing forces. The counterpoint is, hey, it is useful for the world to see powerful AI. It is useful for the world to see powerful AI because that's the only way you can communicate it.

[C]

Well, I guess not even just that you can communicate the idea, but communicate.

[A]

The AI, not the idea, communicate the AI.

[C]

What do you mean communicate the AI?

[A]

Okay, so let's suppose you write an essay about AI and the essay says AI is going to be this and AI is going to be that and it's going to be this. And you read it and you say, okay, this is an interesting essay, right? Now suppose you see an AI doing this, an AI doing that. It is incomparable. Like basically, I think that there is a big benefit from AI being in the public and that would be a reason for us to not be quite straight shot.

[C]

Yeah, well, I guess it's not even that, but I do think that is an important part of it. The other big thing is I can't think of another discipline in human engineering and research where the end artifact was made safer, mostly through just thinking about how to make it safe, as opposed to why are airplane crashes per mile so much lower today than they were decades ago? Why is it so much harder to find a bug in Linux than it would have been decades ago? And I think it's mostly because these systems were deployed to the world. You noticed failures, those failures were corrected and the systems became more robust. Now, I'm not sure why AGI and superhuman intelligence would be any different, especially given. And I hope we can talk, we're going to get to this. It seems like the harms of superintelligence are not just about like having some malevolent paperclipper out there, but it's just like this is a really powerful thing and we don't even know how to conceptualize how people interact with it, what people will do with it. And having gradual access to it seems like a better way to maybe spread out the impact of it and to help people prepare for it.

[A]

Well, I think on this point, even in the straight shot scenario, you would still do a gradual release of it, is how I would imagine it. Gradualism would be an inherent, inherent component of any plan. It's just a question of what is the first thing that you get out of the door. That's number one. Number two, I also think, you know, I believe you have advocated for continual learning more than other people. And I actually think that this is an important and correct thing. And here is why. So one of the things. So I'll give you another example of how thinking, how language affects thinking. And in this case it will be two words. Two words that have shaped everyone's thinking. I maintain, first word, AGI, second word, pre training. Let me explain. So the word, the term AGI, why does this term exist? It's a very particular term. Why does it exist? There's a reason. The reason that the term AGI exists is in my opinion, not so much because it's like a very important, essential descriptor of some end state of intelligence, but because it is a reaction to a different term that existed. And the term is narrow AI. If you go back to ancient history of gameplay and AI, of checkers, AI, chess AI, computer games AI, everyone would say, look at this narrow intelligence. Sure, the chess AI can beat Kasparov, but it can't do anything else. It is so narrow, artificial, narrow intelligence. So in response, as a reaction to this, some people said, well, this is not good. It is so narrow. What we need is general AI. General AI, an AI that can just do all the things. The second. And that term just got a lot of traction. The second thing that got a lot of traction is pre training, specifically the recipe of pre training. I think the current, the way people do RL now is maybe is undoing the conceptual imprint of pre training. But pre training had the property, you do more pre training and the model gets better at everything. More or less uniformly, general AI, pre training gives AGI. But the thing that happened with AGI and pre training is that in some sense they overshot the target. Because if you think about the term AGI, you will realize, and especially in the context of pre training, you will realize that a human being is not an AGI. Because a human being, yes, there is definitely a foundation of skills. A human being lacks a huge amount of knowledge. Instead, we rely on continual learning. We rely on continual learning. And so then when you think about, okay, so let's suppose that we achieve success and we produce a safe, super, some kind of safe superintelligence. The question is, but how do you define it? Where on the curve of continual learning is it going to be? I produce a super intelligent 15 year old that's very eager to go and you say, okay, I'm going to. They don't know very much at all. The great student, very eager, you go and be a programmer, you go and be a doctor, go and learn. So you could imagine that the deployment itself will involve some kind of a learning trial and error period. It's a process, as opposed to you drop the finished thing.

[C]

Okay, I see. So you're suggesting that the thing you're pointing out with superintelligence is not some finished mind which knows how to do every single job in the economy. Because the way, say, the original, I think OpenAI charter or whatever defines AGI is like it can do every single job that every single thing a human can do. You're proposing instead a mind which can learn to do every single job. Yes, and that is super intelligence. And then. But once you have the learning algorithm, it gets deployed into the world the same way a human laborer might join an organization. And it seems like one of these two things might happen. Maybe neither of these happens. One, this super efficient learning algorithm becomes superhuman, becomes as good as you, and potentially even better at the task of ML research, and as a result, the algorithm itself becomes more and more superhuman. The other is, even if that doesn't happen, if you have a single model. I mean, this is explicitly your vision. If you have a single model where instances of a model which are deployed through the economy, doing different jobs, learning how to do those jobs, continually learning on the job, picking up all the skills that any human could pick up, but actually picking them all up at the same time and then amalgamating their learnings, you basically have a model which functionally becomes super intelligent even without any sort of recursive self improvement in software. Because you now have one model that can do every single job in the economy. And humans can't merge our minds in the same way. And so do you expect some sort of like intelligence explosion from broad deployment?

[A]

I think that it is likely that we will have rapid economic growth. I think the broad deployment, like there are two arguments you could make which are conflicting. One is that look, if indeed you get once, indeed you get to a point where you have an AI that can learn to do things quickly and you have many of them, then they will, then there will be a strong force to deploy them in the economy. Unless there will be some kind of a regulation that stops it, which by the way, there might be. But I think the idea of very rapid economic growth for some time, I think it's very possible from broad deployment, then the question is how rapid it's going to be. So I think this is hard to know because on the one hand you have this very efficient worker. On the other hand there is. The world is just really big and there's a lot of stuff and that stuff moves at a different speed. But then on the other hand, now the AI could, you know. So I think very rapid economic growth is possible. And we will see all kinds of things like different countries with different rules and the ones which have the friendlier rules, the economic growth will be faster, hard to predict.

[C]

It seems to me that this is a very precarious situation to be in where looking the limit. We know that this should be possible because if you have something that is as good as a human at learning, but which can merge its brains. Merge. There are different instances in a way that humans can't merge already. This seems like a thing that should physically be possible. Humans are possible. Digital computers are possible. You just need both of those combined to produce this thing. And it also seems like this kind of thing is extremely powerful. And economic growth is one way to put it. I mean, Dyson spirit is a lot of economic growth. But another way to put it is just like you will have potentially a very short period of time Because a human on the job can, you know, you're hiring people at SSI in six months, they're like net productive probably. Right. A human learns really fast. And so this thing is becoming smarter and smarter very fast. How do you think about making that go well? And why is SSI positioned to do that? Well, what is SSI's plan there basically is what I'm trying to ask.

[A]

Yeah. So one of the ways in which my thinking has been changing is that I now place more importance on AI being deployed incrementally and in advance. One very difficult thing about AI is that we are talking about systems that don't yet exist and it's hard to imagine them. I think that one of the things that's happening is that in practice it's very hard to feel the AGI. It's very hard to feel the AGI. We can talk about it, but it's like talking about the long. Imagine having a conversation about how is it like to be old when you're old and frail and you can have a conversation, you can try to imagine it, but it's just hard and you come back to reality where that's not the case. And I think that a lot of the issues around AGI and its future power stem from the fact that it's very difficult to imagine future AI is going to be different. It's going to be powerful indeed. The whole problem. What is the problem of AI and AGI? The whole problem is the power. The whole problem is the power. When the power is really big, what's going to happen? And one of the ways in which I've changed my mind over the past year and so that change of mind may back. May. I'll say I'll hedge a little bit, may back propagate into the plans of our company. Is that. So if it's hard to imagine, what do you do? You gotta be showing the thing. You gotta be showing the thing. And I maintain that I think most people who work on AI also can't imagine it because it's too different from what people see on a day to day basis. I do maintain. Here's something which I predict will happen, that's a prediction. I maintain that as AI becomes more powerful, then people will change their behaviors and we will see all kinds of unprecedented things which are not happening right now. And I'll give some examples. I think for better or worse, the frontier companies will play a very important role in what happens, as will the government and the kind of things that I think you'll See which you see the beginnings of companies that are fierce competitors starting to collaborate on AI safety. You may have seen OpenAI and Anthropic doing a first small step, but that did not exist. That's actually something which I predicted in one of my talks about three years ago, that such a thing will happen. I also maintain that as AI continues to become more powerful, more visibly powerful, there will also be a desire from governments and the public to do something. And I think that this is a very important force of showing the AI that's number one. Number two, okay, so then the AI is being built, what needs to be done? So one thing that I maintain that will happen is that right now people who are working on AI are maintained that the AI doesn't feel powerful because of its mistakes. I do think that at some point the AI will start to feel powerful actually. And I think when that happens, we will see a big change in the way all AI companies approach safety. They'll become much more paranoid. I say this as a prediction that we will see happen. We'll see if I'm right. But I think this is something that will happen because they will see the AI becoming more powerful. Everything that's happening right now, I maintain is because people look at today's AI and it's hard to imagine the future AI. And there is a third thing which needs to happen. And I think this is, and I'm talking about it in broader terms, not just from the perspective of ssi, because you asked me about our company, but the question is, okay, so then what should the companies aspire to build? What should they aspire to build? And there has been one big idea that actually everyone has been locked into, which is the self improving AI. And why did it happen? Because there is fewer ideas than companies. But I maintain that there is something that's better to build. And I think that everyone will actually want that. It's like the AI that's robustly aligned to care about sentient life specifically. I think in particular there's a case to be made that it will be easier to build an AI that cares about sentient life than an AI that cares about human life alone, because the AI itself will be sentient. And if you think about things like mirror neurons and human empathy for animals, which is, you know, you might argue it's not big enough, but it exists. I think it's an emergent property from the fact that we model others with the same circuit that we used to model ourselves, because that's the most efficient thing to do.

[C]

So even if you got an AI to hear about sentient beings and it's not actually clear to me that that's what you should try to do. If you solved alignment, it would still be the case that most sentient beings will be AIs. There will be trillions, eventually quadrillions of AIs. Humans will be a very small fraction of sentient beings. So it's not clear to me if the goal is some kind of human control over this future civilization, that this is the best criterion.

[A]

It's true. I think that it's possible it's not the best criterion. I'll say two things. I think that thing, number one, I think that if the so I think that care for sentient life, I think there is merit to it. I think it should be considered. I think that it will be helpful if there was some kind of a short list of ideas that then the companies, when they are in this situation could use. That's number two. Number three, I think it would be really materially helpful if the power of the most powerful superintelligence was somehow capped because it would address a lot of these concerns. The question of how to do it, I'm not sure, but I think that would be materially helpful when you're talking about really, really powerful systems.

[C]

Yeah. Before we continue the alignment discussion, I want to double click on that. How much room is there at the top? How do you think about superintelligence, do you think, I mean using this learning efficiency idea maybe is just extremely fast at learning new skills or new knowledge? And does it just have a bigger pool of strategies? Is there a single cohesive IT in the center that's more powerful or bigger? And if so, do you imagine that this will be sort of godlike in comparison to the rest of human civilization? Or does it just feel like another agent or another cluster of agents?

[A]

So this is an area where different people have different intuitions. I think it will be very powerful for sure. I think that what I think is most likely to happen is that there will be multiple such AIs being created roughly at the same time. I think that if the cluster is big enough, like if the cluster is literally continent sized, that thing could be really powerful indeed. Right. If you literally have a continent sized cluster like those, those AIs can be very powerful. And I like. All I can tell you is that if you're talking about extremely powerful AIs, like truly dramatically powerful, then yeah, it would be nice if they could be restrained in some ways or if there was Some kind of an agreement or something. Because I think that if you are saying, hey, like, if you really like, what is the concern of superintelligence? What is one way to explain the concern? If you imagine a system that is sufficiently powerful, like really sufficiently powerful, and you could say, okay, you need to do something sensible like care for sentient life, let's say in a very single minded way, we might not like the results. That's really what it is. And so maybe, by the way, the answer is that you do not build a single. You do not build an RL agent in the usual sense. And actually I'll point several things out. I think human beings are a semi RL agent. You know, we pursue a reward and then the emotions or whatever make us tire out of the reward. We pursue a different reward. The market is like kind, it's like a very short sighted kind of agent. Evolution is the same. Evolution is very intelligent in some ways, but very dumb in other ways. The government has been designed to be a never ending fight between three parts which has an effect. So I think things like this. Another thing that makes this discussion difficult is that we are talking about systems that don't exist, that we don't know how to build. That's the other thing. And that's actually my belief. I think what people are doing right now will go some distance and then peter out. It will continue to improve, but it will also not be it. So the it we don't know how to build. And I think that a lot hinges on understanding reliable generalization. I'll say another thing which is like one of the things that you could say that cause alignment to be difficult is that. Your ability to learn human values is fragile. Then your ability to optimize them is fragile. You actually learn to optimize them. And then can't you say, are these not all instances of unreliable generalization? Why is it that human beings appear to generalize so much better? What if generalization was much better? What would happen in this case? What would be the effect? But those questions are right now still unanswerable.

[C]

How does one think about what AI going well looks like? Because I think you've scoped out how AI might evolve. We'll have these sort of continual learning agents. AI will be very powerful. Maybe there will be many different AIs. How do you think about lots of continent, compute size intelligences going around? How dangerous is that? How do we make that less dangerous? And how do we do that in a way that protects a equilibrium where there might be misaligned AIs out there and bad actors out there.

[A]

So one reason why I like the AI that cares for sentient life. And we can debate on whether it's good or bad, but if the first N of these dramatic systems actually do care for, you know, love humanity or something, you know, care for sentient life, obviously this also needs to be achieved. This needs to be achieved. So if this is achieved by the first N of those systems, then I can see it go, well, at least for quite some time. And then there is the question of what happens in the long run. What happens in the long run? How do you achieve a long run equilibrium? And I think that there, there is an answer as well, and I don't like this answer, but it needs to be considered. In the long run, you might say, okay, so if you have a world where powerful AIs existence in the short term, you could say, okay, you have universal high income, you have universal high income and we're all doing well, but we know that, what do the Buddhists say? Change is the only constant. And so things change. And there is some kind of government, political structure thing, and it changes because these things have a shelf life. Some new government thing comes up and it functions and then after some time it stops functioning. That's something that you see happening all the time. And so I think that for the long run equilibrium, one approach, you could say, okay, so maybe every person will have an AI that will do their bidding and that's good. And if that could be maintained indefinitely, that's true. But the downside with that is, okay, so then the AI goes and like, earns money for the person and advocates for their needs in the political sphere, and maybe then writes a little report saying, okay, here's what I've done, here's the situation. And the person says, great, keep it up. But the person is no longer a participant. And then you can say, that's a precarious place to be in. But so I'm going to preface by saying I don't like this solution, but it is a solution. And the solution is if people become part AI with some kind of neuralink plus plus, because what will happen as a result is that now the AI understands something and we understand it too, because now the understanding is transmitted wholesale. So now if the AI is in some situation, now it's like you are involved in the situation yourself fully. And I think this is the answer to the equilibrium.

[C]

I wonder if the fact that emotions which were developed millions or in many cases billions of years ago in a totally different environment are still guiding our actions so strongly. Is an example of alignment success. To maybe spell out what I mean. The brainstem has these. I don't know if it's more accurate to call it a value function or reward function. But the brainstem has a directive where it's saying mate with somebody who's more successful. The cortex is the part that understands what does success mean in the modern context. But the brainstem is able to align the cortex and say however you recognize success to be. And I'm not smart enough to understand what that is. You're still going to pursue this directive.

[A]

I think. I think there is. So I think there's a more general point. I think it's actually really mysterious how the brain encodes high level desires. Sorry. How evolution encodes high level desires. Like it's pretty easy to understand how evolution would endow us with the desire for food that smells good because smell is a chemical. And so just pursue that chemical. It's very easy to imagine such evolution doing such a thing. But evolution also has endowed us with all these social desires. Like we really care about being seen positively by society. We care about being in a good standing. All these social intuitions that we have. I feel strongly that they're baked in. And I don't know how evolution did it. Because it's a high level concept. It's represented in the brain like what people think. Like let's say you are like you care about some social thing. It's not like a low level signal like smell. It's not something that for which there is a sensor. Like the brain needs to do a lot of processing to piece together lots of bits of information to understand what's going on socially. And somehow evolution said that's what you should care about.

[C]

Yes.

[A]

How did it do it? And he did it quickly too.

[C]

Yeah.

[A]

Because I think all these sophisticated social things that we care about. I think they evolved pretty recently. So evolution had an easy time hard coding this high level desire. And I maintain. Or you know. At least I'll say I'm unaware of good hypothesis for how it's done. I. I had some ideas I was kicking around. But none of them. None of them are satisfying. Yeah.

[C]

And what's especially impressive is it was a desire that you learned in your lifetime. It kind of makes sense because your brain is intelligent. It makes sense why we'll be able to learn intelligent desires. But your point is that the desire is. Maybe this is not your point. But one way to understand it is the Desire is built into the genome, and the genome is not intelligent. Right. But it's able to. You're somehow able to describe this feature that requires like, it's not even clear how you define that feature. And you can get it into the, you can build it into the genes.

[A]

Yeah, essentially. Or maybe I'll put it differently. If you think about the tools that are available to the genome, it says, okay, here's a recipe for building a brain. And you could say, here is a recipe for connecting the dopamine neurons to like the smell sensor.

[C]

Yeah.

[A]

And if the smell is a certain kind of, you know, good smell, you want to eat that. I could imagine the genome doing that. I'm claiming that it is harder to imagine, it's harder to imagine the genome saying you should care about some complicated computation that your entire brain, that like a big chunk of your brain does. That's all I'm claiming. I can tell you like a speculation. I was wondering how it could be done. And let me offer a speculation and I'll explain why the speculation is probably false. So the speculation is. Okay, so the brain, it's like the brain has those regions, you know, the brain regions, we have our cortex, right?

[C]

Yeah.

[A]

And there's all those brain regions. And the cortex is uniform, but the brain regions and the neurons in the cortex, they kind of speak to their neighbors mostly. And that explains why you get brain regions. Because if you want to do some kind of speech processing, all the neurons that do speech, they talk to each other. And because neurons can only speak to their nearby neighbors, for the most part, it has to be a region. All the regions are mostly located in the same place from person to person. So maybe evolution hard coded literally a location on the brain. So it says, oh, when the GPS of the brain GPS coordinates such and such, when that fires, that's what you should care about. Maybe that's what evolution did, because that would be within the toolkit of evolution.

[C]

Yeah. Although there are examples where, for example, people who are born blind have that area of their cortex adopted by another sense. And I have no idea, but I'd be surprised if the desires or the reward functions which require visual signal no longer worked. People who have different areas of their cortex co opted, for example, if you no longer have vision, can you still feel the sense that I want people around me to like me and so forth, which usually there's also visual cues for.

[A]

So I actually fully agree with that. I, I think there's an even stronger counterargument to this theory, which is like if you think about people, so there are people who get half of their brains removed in childhood and they still have all their brain regions, but they all somehow move to just one hemisphere, which suggests that the brain regions, their location is not fixed. And so that theory is not true. It would have been cool if it was true, but it's not. And so I think that's a mystery. But it's an interesting mystery. Like the fact is, somehow evolution was able to endow us to care about social stuff very, very reliably. And even people who have like all kinds of strange mental conditions and deficiencies and emotional problems tend to care about.

[C]

This also, what is SSI planning on doing differently? So presumably your plan is to be one of the frontier companies when this time arrives and then what is presumably you started SSI because you're like, I think I have a way of approaching how to do this safely in a way that the other companies don't. What is that difference?

[A]

So the way I would describe it as, there are some ideas that I think are promising and I want to investigate them and see if they are indeed promising or not. It's really that simple. It's an attempt, I think that if the ideas turn out to be correct, these ideas that we discussed around understanding generalization, if these ideas turn out to be correct, then I think we will have something worthy. Will they turn out to be correct? We are doing research. We are squarely age of research company. We are making progress. We've actually made quite good progress over the past year, but we need to keep making more progress, more research. And that's how I see it. I see it as an attempt to be. An attempt to be a voice and a participant.

[C]

People have asked your co founder and previous CEO left to go to Meta recently and people have asked, well, if there was a lot of breakthroughs being made, that seems like a thing that should have been unlikely. I wonder how you respond.

[A]

Yeah, so for this I will simply remind a few facts that may have been forgotten and I think these facts which provide the context, I think they explain the situation. So the context was that we were fundraising at a 32 billion valuation and then Meta came in and offered to acquire us and I said no, but my former co founder like in some sense said yes. And as a result he also was able to enjoy from a lot of near term liquidity and he was the only person from SSI to join Meta.

[C]

It sounds like SSI's plan is to be a company that is at the frontier. When you get to this very important period in human history where you have superhuman intelligence and you have these ideas about how to make superhuman intelligence go well. But other companies will be trying their own ideas. What distinguishes SSI's approach to making superintelligence go well?

[A]

The main thing that distinguishes SSI is, is its technical approach. So we have a different technical approach that I think is worthy and we are pursuing it. I maintain that in the end there will be a convergence of strategies. So I think there will be a convergence of strategies where at some point as AI becomes more powerful, it's going to become more or less clearer to everyone what the strategy should be. And it should be something like, yeah, you need to find some way to talk to each other and you want your first actual real super intelligent AI to be aligned and somehow be, you know, care for, sentient life, care for people, democratic, one of those, some combination of thereof. And I think this is the condition that everyone should strive for. And that's what SSI is striving for. And I think that this time, if not already, all the other companies will realize that they're striving towards the same thing. And we'll see. I think that the world will truly change as AI becomes more powerful. And I think a lot of these forecasts will like, I think things will be really different and people will be acting really differently.

[C]

Speaking of forecasts, what are your forecasts to this system you're describing which can learn as well as a human and subsequently as a result becomes superhuman?

[A]

I think like five to 25 to 20 years.

[C]

So I just want to unroll your. How you might see the world coming. It's like we have a couple more years where these other companies are continuing the current approach and it stalls out and stalls out here. Meaning they earn no more than low hundreds of billions in revenue. Or how do you think about what stalling out means?

[A]

Yeah, I think the re. I think it could. I think it could stall out and I think stalling out will look like it will all look very similar. Yeah. Among all the different companies, something like this. I'm not sure because I think, I think, I think even with, I think even. I think even with stolen out, I think these companies could make a stupendous revenue. Maybe not profits because they will be. It will be. They will need to work hard to differentiate each other from themselves. But revenue definitely.

[C]

But there's something in your model implies that when the correct solution does emerge, there will be convergence between all the companies. And I'm curious why you think that's the case.

[A]

Well, I was talking more about convergence on their largest strategies. I think eventual convergence on the technical approach is probably going to happen as well. But I was alluding to convergence to the largest strategy. So what exactly is the thing that should be done?

[C]

I just want to better understand how you see the future unrolling. So currently we have these different companies and you expect their approach to continue generating revenue but not get to this human like learner.

[A]

Yes.

[C]

So now we have these different forks of companies, we have vue, we have thinking machines, there's a bunch of other labs.

[A]

Yes.

[C]

And maybe one of them figures out the correct approach, but then the release of their product makes it clear to other people how to do this thing.

[A]

I think it won't be clear how to do it thing, but it will be clear that something different is possible and that is information. And I think people will then be trying to figure out how that works. I do think though that one of the things that I think, you know, not addressed here, not discussed, is that with each increase in the AI's capabilities, I think there will be some kind of changes, but I don't know exactly which ones in how things are being done. And so like I think it's going to be important yet I can't spell out what that is exactly.

[C]

And how are the by default you would expect the company that has the model, company that has that model to be getting all these gains because they have the model that learning how to do all has the skills and knowledge that it's building up in the world. What is the reason to think that the benefits of that would be widely distributed and not just end up at whatever model company gets this continuous learning loop going first.

[A]

Like I think that empirically what happened. So here, here is what I think is going to happen. Number one, I think empirically, when. Let's look at how things have gone so far with the AIs of the past. So one company produced an advance and the other company scrambled and produced some similar things after some amount of time and they started to compete in the market and push the prices down. And so I think from the market perspective, I think something similar will happen there as well. Even if someone. Okay, we are talking about the good world, by the way, where. What's the good world? What's the good world where we have these powerful human like learners that are also like. And by the way, maybe there's another thing we haven't discussed on the spec of the super intelligent AI that I think is worth considering. Is that you make it narrow, can be useful and narrow at the same time. So you can have lots of narrow, super intelligent AIs. But suppose you have many of them and you have some company that's producing a lot of profits from it, and then you have another company that comes in and starts to compete. And the way the competition is going to work is through specialization. I think what's going to happen is that the way competition loves specialization and you see it in the market, you see it in evolution as well. So you're going to have lots of different niches and you're going to have lots of different companies who are occupying different niches in this kind of world where you might say, yeah, one AI company is really quite a bit better at some area of really complicated economic activity and a different company is better at another area and the third company is really good at litigation and that's the way you want to learn.

[C]

Is this contradicted by what human learning implies? Is that it can learn, it can.

[A]

But you have accumulated learning, you have a big investment, you spent a lot of compute to become really, really, really good, really phenomenal at this thing. And someone else spent a huge amount of compute and a huge amount of experience to get really, really good at some other thing. Right? You apply a lot of human learning to get there. But now like you are at this high point where someone else would say, look like I don't want to start learning what you've learned.

[C]

I guess that would require many different companies to begin at the human like continual learning agent at the same time so that they can start their different research in different branches. But if one company, you know, gets that agent first or gets that learner first, it does then seem like, well, you know, if you just think about every single job in the economy, you just have instance learning, each one seems tractable for a company.

[A]

Yeah, that's a valid argument. My strong intuition is that it's not how it's going to go. My strong intuition is that, yeah, the argument says it will go this way, but my strong intuition is that it will not go this way. That this is the, you know, in theory there is no difference between theory and practice. In practice there is. And I think that's going to be one of those.

[C]

A lot of people's models of recursive self improvement literally explicitly state, we will have a million ilias in a server that are coming up with different ideas and this will lead to a superintelligence emerging very fast. Do you have some intuition about how Parallelizable. The thing you are doing is.

[A]

What.

[C]

Are the gains from making copies of Ilya?

[A]

I don't know. I think there'll definitely be, there'll be diminishing returns because you want, you want people who think differently rather than the same. I think that if they were literal copies of me, I'm not sure how much more incremental value you'd get. I think that. But people who think differently, that's what you want.

[C]

Why is it that it's been. If you look at different models, even released by totally different companies trained on potentially non overlapping data sets, it's actually crazy how similar LLMs are to each other.

[A]

Maybe the data sets are not as non overlapping as it seems, but there's some senses.

[C]

Even if an individual human might be less productive than the future AI, maybe there's something to the fact that human teams have more diversity than teams of AIs might have. But how do we elicit meaningful diversity among AI? So I think just raising the temperature just results in gibberish. I think you want something more like different scientists have different prejudices or different ideas. How do you get that kind of diversity among AI agents?

[A]

So the reason there has been no diversity I believe is because of pre training. All the pre trained models are the same pretty much because they're pre trained on the same data. Now RHEL and post training is where some differentiation starts to emerge because different people come up with different RL training.

[C]

Yeah. And then I've heard you hint in the past about self play as a way to either get data or match agents to other agents with equivalent intelligence to kick off learning. How should we think about why there's no public proposals of this kind of thing? Working with LLMs, I would say there.

[A]

Are two things to say. I would say that the reason why I thought self play flow is interesting is because it offered a way to create models using compute only without data. Right. And if you think that data is the ultimate bottleneck, then using compute only is very interesting. So that's what makes it interesting. Now the, the thing is that self play, at least the way it was done in the past, when you have agents which somehow compete with each other, it's only good for developing a certain set of skills. It is too narrow. It's only good for negotiation, conflict, certain social skills, strategizing, that kind of stuff. And so if you care about those skills, then self play will be useful. Now actually I think that self play did find a home, but just in a different form. In a different form so things like debate prove a verifier. You have some kind of an LLM as a judge, which is also incentivized to find mistakes in your work. You could say this is not exactly self play, but this is a related adversarial setup that people are doing, I believe. And really, self play is an example of, is a special case of more general competition between agents. Right. The response, the natural response to competition is to try to be different. And so if you were to put multiple agents and you tell them you all need to work on some problem and you are an agent and you're inspecting what everyone else is working, you're going to say, well, if they're already taking this approach, it's not clear I should pursue it. I should pursue something differentiated. And so I think that something like this could also create an incentive for a diversity of approaches.

[C]

Yeah, Final question. What is research taste? You're obviously the person in the world who is considered to have the best taste in doing research in AI. You were the co author on many of the biggest, the biggest things that have happened in the history of deep learning, From Alexnet to GPT3 to so on. What is it that. How do you characterize how you come up with these ideas?

[A]

So I can comment on this for myself? I think different people do it differently. But one thing that guides me personally is an aesthetic of how AI should be by thinking about how people are, but thinking correctly. It's very easy to think about how people are incorrectly, but what does it mean to think about people correctly? So I'll give you some examples. The idea of the artificial neuron is directly inspired by the brain, and it's a great idea. Why? Because you say, sure, the brain has all these different organs, it has the folds, but the folds probably don't matter. Why do we think that the neurons matter? Because there is many of them. It kind of feels right. So you want the neuron, you want some kind of local learning rule that will change the connections. You want some local learning rule that will change the connections between the neurons. Right. It feels plausible that the brain does it. The idea of the distributed representation, the idea that the brain, you know, the brain responds to experience. Our neural nature should learn from experience, not response. The brain learns from experience, the neural naturally learn from experience. And you kind of ask yourself, is something fundamental or not fundamental, how things should be? Yeah, and I think that's been guiding me a fair bit, kind of thinking from multiple angles and looking for almost beauty. Beauty, simplicity, ugliness. There's no room for ugliness. It's just beauty, simplicity, elegance, correct inspiration from the brain. And all of those things need to be present at the same time. And the more they are present, the more confident you can be in a top down belief. And then the top down belief is the thing that sustains you when the experiments contradict you. Because if you just trust the data all the time, well, sometimes you can be doing a correct thing. But there's a bug. But you don't know that there is a bug. How can you tell that there is a bug? How do you know if you should keep debugging or you conclude it's the wrong direction? Well, it's the top down. Well, how should you can say the things have to be this way. Something like this has to work, therefore you got to keep going. That's the top down. And it's based on this like multifaceted beauty and inspiration by the brain.

[C]

All right, we'll leave it there.

[A]

Thank you so much.

[C]

Thank you so much.

[A]

All right. Appreciate it.

[C]

That was great.

[A]

Yeah, I enjoyed it.

[C]

Yes, me too.

[B]

Thanks for listening to this episode of the A16Z podcast. If you like this episode, be sure to like, comment, subscribe, leave us a rating or review and share it with your friends and family. For more episodes, go to YouTube, Apple Podcasts and Spotify. Follow us on X16Z and subscribe to our substack@A16Z substack.com thanks again for listening and I'll see you in the next episode. This information is for educational purposes only and is not a recommendation to buy, hold, or sell any investment or financial product. This podcast has been produced by a third party and may include paid promotional advertisements, other company references, and individuals unaffiliated with A16Z. Such advertisements, companies and individuals are not involved. Endorsed by AH Capital Management, LLC, A16Z or any of its affiliates. Information is from sources deemed reliable on the date of publication, but A16Z does not guarantee its accuracy.