Analyst Calls Chinese Open Source Frontier AI Models Flimsy And Contingent

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTexTECH

Regardless of how Chinese deliberation on export controls goes, I think everyone must internalize how flimsy the "China open source" thing is. We've been developing a disgustingly complacent culture, where of course Americans will sell you frontier tokens, like those capitalists selling rope to hang them; and if the worst comes to worst, you can wince and fall back on GLM X.Y. that's <8 months behind (despite it, again, being handicapped by Dutch export controls against "unfree countries"). No bro. Between L3-405B (technically obsolete at release, followed by flailing and abandonment of the initiative) and DeepSeek-V3/R1, the idea of frontier open models was quite fanciful, cope even. This is a very contingent arrangement, made possible by generic Chinese catch-up industrial policy, somewhat lazily extrapolated state philosophy of encouraging academic publishing and open source to improve transparency, build global standards and bolster "discourse power", wild talent overproduction in conditions of compute scarcity, and last but not least — *the will* of one outlier individual, namely Liang Wenfeng, who has nudged China towards a specific path-dependent regime of hardcore technical lab competition in the open, and is trying to shape it further, making AGI part of an infrastructure rush, a utility, the abundant access to which harmonizes with the Party vision of Great Rejuvenation and all downstream plans. At least to some extent, namely the extent of making this achievement global with no strings attached, he will definitely fail; they'll sell you transformers and PV panels but they don't subsidize third countries' electricity prices, and there's no reason for them to subsidize third countries' (nevermind *hostile countries'* engaged in a trade war with them) augmentation of human capital. When China makes it to the current American frontier of AI, and even before that, their best move is to use it to entrench their competitiveness in everything they are already strong in, capitalizing on superior supply chains and abundance of trained labor, on costly experimental equipment, proprietary data streams and industrial feedback. We see that the state wants just that. In such matters, the Chinese state tends to get what it wants. What does the European state, to an extent that it's a real entity, want? "Predictability", as von der Leyen says? They'll have it, the predictability of obsolescence and decline; this is the convergent default outcome. Anything else requires more balls and will to power, and the hunger to make the best of what advantages they still hold.

Commentary on X