Why distillation explains less of China's AI progress than assumed

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 9:01

   Yeah, so the Chinese labs are very technically excellent. They are more compute star. Yes. Without a doubt, they have a comparable talent pool to draw on. I think there might be like a slight edge based on who was where when AI really got going. But to say that China is anything but equal in talent, you're going to have a hard time because they just have so many people being trained at such a high level.

2. 9:23

   And what they don't have is compute. You talk to people there and you're like, oh, I heard an average open AI researcher has 5,000 H100 equivalents that they can access at any time. And like their jaw just falls to the floor. They're just like, I can't comprehend. And this is like a top pre-training researcher at one of the labs that you would know. and then you have this interaction a few times. And the compute thing, I think, is the biggest difference.

3. 9:46

   And distillation is a way to effectively spend more money to get what is compute equivalent. Like outputs from a language model are a function of compute. And if you know the outputs you want, it's the same as having compute to generate with a model. And I've seen the stages of distillation and how I see it is that previous post-training was much more reliant on the supervised fine-tuning and other to like this RLHF era, which is where post-training started.

4. 10:13

   And you could get most of your performance just by you do this next token prediction loss on the outputs of a very strong model. It was very established that you take the best model you can, you generate completions, you train on it, and your model's going to be pretty good so long as you have access to the right prompts. Prompts have always been a key to post-training. I remember at AI2, like with Hannah, who I co-led post-training there with, we would always say like prompts are the key unit of post-training.

5. 10:39

   And so you need to get the prompts and you can do different things with them at different stages of training. And now this has morphed a bit into environments where it's RL heavy. But it's just like hard to map this type of post training, which is you get the completions and the score goes up to what is going on in reasoning today. And based on the public literature, it's very hard for me to be like distillation is a major factor.

6. 11:01

   because, for example, in the fully open literature of people trying to do open post-training recipes, all the open model builders are stuck using a really outdated Quen model because you can't figure out how to get the teacher right. My hypothesis is that there's some mid-training and something in the common pre-training data that's open that makes it hard to do this thing, which is get reasoning traces from the absolute best model and do SFT and the numbers go up.

7. 11:26

   So if I were to have access to Claude's reasoning trace when I was building Ulmo, I'm actually like Claude Opus, Claude Mythos, whatever. I'm not sure that it would have actually made our model a lot better. So this is like my evidence grounded in the literature, why I think distillation is hard. And then like, how do you take the outputs of a model and build an RL environment to do a big RL run? Because then you have the Xiaomi's of the world doing this week-long open RL run.

8. 11:52

   They're like, here's what we did. And then they've dropped the model. And it's like, where exactly does distillation go into this? They're clever. They know how to do marketing well. There are some questions of like, did they do this ahead of time? Because they did their RL run and then they dropped the model and then their tech report said they had an extra post-training stage after RL. So there's some questions.

9. 12:11

   But like mapping the established distillation literature to that is very hard. But there are question marks that make me support the labs a bit, which is like these cloud proxies and these model proxies in China where there's a ton of tokens that are used and there's intermediaries and they have tons of accounts. And even cases where Kimmy was serving Claude in their products to users. And I'm just like, don't really understand how this fits into distillation as a common technical tool in post-training.

10. 12:41

    And like, yes, we should watch it too. Like, there's other weird stuff with Chinese companies using the outputs from OpenAI and Anthropics models at a vast scale. So I think we have a lot more questions we need to answer. And I think most reasonable people when they're like, what's the gap from distillation? are like, there's very big error bars. But when I say that distillation helps them by a few months, it's based on this scientific literature that we just, like, that's the information we have.

11. 13:07

    We should do more open research and enable this so that we can study it. But like, I don't know how you can come to a different conclusion that it's like not a huge effect in current post-training based on this. And that's just like, the Chinese are so good at building models as well. I think there's a bit of cope where people are like, only we are good at building models. And it's like, there's so many companies.

12. 13:26

    There's like 20 to 30 companies that are building good models. You can like read.