# The Real Surprise of the Last 3 Years in AI – Dario Amodei

*Dwarkesh Clips* · https://www.youtube.com/watch?v=UKfmj_zE0VE

## Summary

Dario Amodei reflects on the last three years of AI progress, saying the underlying exponential has developed roughly as he expected, from smart high school student to college student to PhD-level capabilities, with code going beyond that. The biggest surprise for him is the lack of public recognition of how close we are to the end of the exponential. He restates his 2017

## Key points

- Dario Amodei says the underlying technology exponential has gone roughly as he expected over the last 3 years, plus or minus a year or two.
- The most surprising thing to him has been the lack of public recognition of how close we are to the end of the exponential.
- He has held the same hypothesis since 2017, written in a doc called the big blob of compute hypothesis, similar to Rich Sutton's bitter lesson.
- The hypothesis lists about seven things that matter, including raw compute, quantity of data, quality and distribution of data, training duration, and an objective function that can scale to the moon.
- Pre-training scaling laws have continued to give gains, and the same log-linear scaling is now being seen with RL on top of pre-training.
- RL scaling has been shown publicly by other companies on math contests like AIME, and Anthropic sees the same scaling across a wide variety of RL tasks.

## Notes

### Overview
- Dario Amodei discusses what has and hasn't surprised him about AI progress over the last 3 years.
- The technology exponential has gone roughly as expected, with plus or minus a year or two on specific timelines.

### The Biggest Surprise
- The most surprising thing is the lack of public recognition of how close we are to the end of the exponential.
- People inside and outside the AI bubble keep discussing the same tired old hot button political issues.

### The Big Blob of Compute Hypothesis
- Written by Amodei in 2017, around when GPT-1 had just come out.
- It was a general document, not specific to language models — it also covered robotics, separate reasoning work, and RL scaling like AlphaGo, Dota at OpenAI, and AlphaStar at DeepMind.
- Similar to Rich Sutton's bitter lesson, published a couple years later.
- Core idea: clever techniques and new methods don't matter very much; only a few things matter, and he listed seven of them.

### The Seven Factors
- How much raw compute you have.
- The quantity of data.
- The quality and distribution of data — it needs to be a broad distribution.
- How long you train for.
- An objective function that can scale to the moon — the pre-training objective is one; the RL objective of reaching a goal is another, with objective rewards (math, coding) and subjective rewards (RL from human feedback).
- Two factors around normalization or conditioning — numerical stability so the big blob of compute flows in a laminar way.

### Scaling Today
- Amodei still holds this hypothesis and hasn't seen much that contradicts it.
- Pre-training scaling laws have continued, and it's widely reported that pre-training is continuing to give gains.
- The same scaling is now seen with RL on top of pre-training: other companies have published results showing performance on math contests like AIME is log-linear in how long the model is trained.
- Anthropic sees the same RL scaling across a wide variety of tasks, not just math contests.

## Transcript

[00:00] So, we talked 3 years ago. I'm curious in your view, what has been the biggest update of the last 3 years? What has been the biggest difference between what it felt like last 3 years versus now? >> Yeah, I would say actually the underlying technology, like the exponential of the technology, has has gone, broadly speaking, I would say about about as I expected it to go. I mean, there's like plus or minus, you know, a couple. There's plus or minus a year or two here, there's plus or minus a year or two there. I don't know that I would predict the specific direction of code. Um, but but actually when I look at the exponential, it it is roughly what I expected in terms of the march of the models from like, you know, smart

[00:39] high school student to smart college student to like, you know, beginning to do PhD and professional stuff and in the case of code reaching beyond that. So, you know, the frontier is a little bit uneven. It's roughly what I expected. I will tell you though what the most surprising thing has been. The most surprising thing has been the lack of public recognition of how close we are to the end of the exponential. To me, it is absolutely wild that, you know, you have people, you know, within the bubble and outside the bubble. You know, but but you have people talking about these these, you know, just the same tired old hot button political issues and like, you know, or or or around us we're like

[01:19] near the end of the exponential. I I I want to understand what that exponential looks like right now because the first question I asked you when we recorded 3 years ago was, you know, what's up with scaling? How what is it work? Um, I have a similar question now, but I feel like it's a more complicated question because at least from the public's point of view Yes. >> 3 years ago there were these, you know, well-known public trends where across many orders of magnitude of computer you could see how the loss improves. And now we have RL scaling and there's no publicly known scaling law for it. It's not even clear what exactly the story is of is it supposed to be teaching the model skills? Is it supposed to be

[01:54] teaching metalearning? Um, what is the scaling hypothesis at this point? >> Yeah, so so I have actually the same hypothesis that I had even all the way back in 2017. So, in 2017, I think I talked about it last time, but I wrote a doc called the the big blob of compute hypothesis. And and you know, it it wasn't about the scaling of language models in particular. When I when I wrote it, GPT-1 had had just come out, right? So, that was you know, one among many things, right? There was back in those days, there was robotics, people trying to work on reasoning as a separate thing from language models. There was scaling of the kind of RL that happened that you know, kind of happened in AlphaGo, and you know, that that

[02:34] happened at Dota at OpenAI, and you know, people remember StarCraft at DeepMind, you know, the AlphaStar. Um, so it was written as a more general document. And and the specific thing I said was the following, that and you know, it's it's very you know, Rich Sutton put out the bitter lesson a couple years later, but you know, the the hypothesis is basically the same. So, so what it says is all the cleverness, all the techniques, all all the kind of we need a new method to to do something like that, doesn't matter very much. There only a few things that matter, and I think I listed seven of them. One is like how much raw compute you have. The other is the quantity of data that you have. Then the third is kind of the

[03:18] quality and distribution of data, right? It needs to be a broad broad distribution of data. The fourth is I think how long you train for. Um, the fifth is you need an objective function that can scale to the moon. So, the pre-training objective function is one such objective function, right? An another objective function is, you know, the the kind of RL objective function that says like you have a goal, you're going to go out and reach the goal. Within that of course, there's objective rewards like, you know, like you see in math and coding, and there's more subjective rewards like you see in RL from human feedback or kind of higher order higher order versions of that. And and then the sixth and seventh were

[03:58] things around kind of like normalization or conditioning like you know just getting the numerical stability so that kind of the big blob of compute flows in this laminar way instead of instead of running into problems. So that was the hypothesis and this hypothesis I still hold. I I don't think I've seen very much that is not in line with that hypothesis. And so the pre-training scaling laws were one example of what of of of of kind of what we see there. And indeed those have continued going like you know you know I think I think now it's been it's been widely reported like you know we feel good about pre-training like pre-training is continuing to give us gains. What has changed is that now we're also seeing the same thing for RL

[04:44] right so we're seeing a pre-training phase and then we're seeing like an RL phase on top of that. And with RL it's it's actually just the same like you know even even other companies have have published like you know in some of their in some of their releases have published things that say look you know we train the model on math contests you know AIME or or the kind of other things and you know how well how well the model does is log linear and how long we've trained it. And we see that as well and it's not just math contests it's a wide variety of RL tasks. And so we're seeing the same scaling in RL that we saw for pre-training. If you enjoyed this clip you can watch the full episode here and

[05:29] subscribe for more clips. Thanks.