Meta is axing 600 roles across its AI division

Surely winning the AI race is finding secret techniques that allow development of superior models, with it not being apparent that anyone has anything special enough that he actually is winning?

I think there's some firms with special knowledge: Google, possibly OpenAI/Anthropic, possibly the Chinese firms, possibly Mistral too, but no one has enough unique stuff to really stand out.

The biggest things were those six months before people figured out how O1 worked and the short time before people figured out how Google and possibly OpenAI solved 5/6 of the 2025 IMO problems.

I think that depends on how optimistic/pessimistic one is on how much more superior the models are going to get. If you're really pessimistic then there isn't all too much one company could do to be 2x or more ahead already. If you're really optimistic then it doesn't matter what anyone is doing today because it's about who finds the next 100x leap.

I don't think it does.

The models have increased greatly in capabilities, but the competitors have simply kept up, and it's not apparently that they won't continue to do that. Furthermore, the breakthroughs-- i.e. fundamentally better models, can happen anywhere where people can and do try out new architectures, and that can happen in surprisingly small places.

It's mostly about culture and being willing to experiment on something which is often very thankless since most radical ideas do not give an improvement.

Which is why getting rid of friction is a good idea.

This is R&D. You want a skunkworks culture where you have the best people in the world trying as many new things as possible, and failure is fine as long as it's interesting failure.

Not a culture where every development requires a permission slip from ten other teams, and/or everyone is worried if they'll still have a job a month from now.

Yes, definitely.