出口管制下中国开源AI的战略逻辑

I should write a longer piece on this, but: it's really funny how export control bros are confused by Chinese open sourcing. In fact they're being very courteous; the CCP *should* open source cyberweapons. Consider: 1) you have a fixed compute budget for R&D, training, inference, and special defensive projects like Glasswing 2) you won't sell them chips, tools, or even high-end tokens. You already try your best to render their state irrelevant roadkill in the next technological paradigm. 3) you are explicitly building a cyberweapon in the form of Mythos, Dario is salivating at the thought of how it'll make democracies invincible and end the CCP 4) In China, compute is scarce, overpriced and accounted for, internet is largely isolated from the outside, densely surveilled, full of circuit breakers, providers have Party minders, and virtually no private citizen has a realistic chance of doing much damage (at least to China) with a 3T model no matter how good it is 5) the world outside China is quite chaotic, has plenty of misaligned actors, and much easier time standing up a node of 64 GPUs 6) do you see where I'm going with this?

They lose very little commercially and they gain strategically by proliferating frontier capabilities they don't have resources to inference at scale. They want Glasswing and other safety measures to consume your compute advantage. The more you have to dedicate to defense, the slower your takeoff, the narrower the gap, the worse your negotiating position.

Of course, this logic only holds so long as you're irredeemably hostile locusts, have a vast compute overmatch, and stronger models intended for offensive use against China anyway (so that enabling smaller non-aligned actors basically does no marginal damage to their security). In this model, your best hope to negotiate the end of Chinese open souring is offering them semiconductor tool relief (GPUs won't be enough, I"m afraid).

This is not policy advice. I think you can proceed. Just one thing to keep in mind.