You CAN argue that these capabilities stem from research which isn't getting replicated (because it's compute-intense), and so this will majorly slow down China, plus the capability gap is great enough that catch-up will take a while. But no, they will not be capped at US level.

I guess Anthropic is cynically misconstruing it to get the desired policy action, because if they put it as it is, they'll validate the arms race anxiety and fail to persuade the DC. Rationalists/doomers on here, being racist and tech-illiterate, take it at face value.

Joe mentions it

@teortaxesTex I do think that raw distillation will be less efficient compared to very specific workflows for synthetic data generation with many controllable knobs. Will still be used but less impactful

@Rafa_Schwinger I am generally bearish on cost-efficiency of distillation in this era, and Minimax for me proves its limits. But you can be smart about it, eg like Ant accuses DeepSeek of reverse engineering their reward models.

@teortaxesTex Distillation smells like the input-output era of AI as a black box but as we move towards the workflow era where the bulk of it is run in their servers and not exposed my hunch is that the information ROI will decrease

@teortaxesTex china gets the research done while we celebrate compute flexing. wake up