Pinned post
Published!
Working on a post regarding AI safety in China, which should be up tomorrowView quoted post on X
@JacquesThibs
Pinned post
Working on a post regarding AI safety in China, which should be up tomorrowView quoted post on X
Kimi K3 is obviously good and a big accomplishment, but I am indeed of the view that:View quoted post on X
- There’s lots to pull out of pretraining via great post-training
- Using traces from other better models allows you to hit the ceiling much faster
- Therefore, an exceptional post-training process allows you to leapfrog stronger base models with weaker post-training (as Ryan effectively shares an example of)
My guess is that Google is much more careful about post-training on OpenAI/Anthropic traces, which is one reason they are behind the Chinese labs as of now.
I think it’s also true to say the Chinese labs are currently paying less attention and effort to safety-related stuff. Google has much more to lose if they release a model that does/says some crazy shit.
I know, I know, “distillation cope”, but I do want to say I’m fine being proven wrong and I also don’t deny the Chinese labs have strong technical teams with extreme talent density. I’m also not anti-open-source.