China Welcomes Open Source AI Models Despite Censorship Concerns

This is possible. Limiting knowledge on biowarfare, license changes in particular are almost certain. Copying the US wholesale, especially if they get sold on some "G2 dialogue of good faith", is also a move I can see the CCP making. Then again, recall that we've been through this discussion a few times. In June 2023 (when DeepSeek was founded), Helen Toner (you know her from the OpenAI board drama) argued that not only is Chinese AI progress feeble, but Xi Jinpooh will not tolerate "freewheeling chatbots" in principle. (We're in a race with a rigid, brutal totalitarian state, a slave bug empire aesthetically opposed to human freedom, 'member? That's the premise of the scenario, the rationale for this whole neo-Cold War). Open source was not even considered a factor. 1.5 years later R1 happened, Xi was shaking hands with Wenfeng, and a year later he's *at least* AGI-pilled enough to consider AI to be comparable to "steam engine, electricity or the internet", yet we still see open source releases welcomed by the state. Maybe he's not pilled ENOUGH, after all he doesn't say "nuclear weapons" (Dario's preferred metaphor for what he's building). But he's at least ahead of European philosopher-queens.

Anyway, some more strategic reasons a general crackdown on open source capabilities might still not happen for at least the next 4 update cycles, which is how long I expect them to kinda catch up to Mythos in ways that matter: Tl;DR: Restricting public access to the best Chinese models does not meaningfully improve the CCP's chances of not getting discombobulated. 1) Chinese models are unlikely to get better than Dario's strongest soldier at any given moment, which also constitutes the bulk of the threat to Chinese national security. Making them less accessible mostly increases their own ignorance about the state of affairs. 2) There are few systems in China that can serve a, say, 6T monster at any real throughput, and it'll be easy to account for them given new tools. If, driven by hardware differences, model design moves further towards hyper-sparsity, this is only exacerbated. As their new stock becomes mostly domestic compute (Huawei Ascend/Cambricon/Baidu Kunlunxin/Alibaba T-Head etc), very scarce and sold in bulk to a small number of entities, tracking will become even more trivial, and China can strike fear into the hearts of its businessmen in ways that Trump envies. We will likely see mandatory Fable-style nerfing at deployment on offense-relevant problems on SiliconFlow, Alibaba Cloud and others. New models make it easier to make better censorship, after all. Generally, China will have little fear of domestic rogue actors. Two other major cyberoffense nations (Russia, North Korea lol) are allies and have no compute anyway. Europe is… as you see, Europeans have their own problems, attacking China is a bit ludicrous for them now, and they can always sell out more for a Mythos subscription to do that anyway. 3) As I've said before, distillation is becoming a relatively minor factor, proliferation of narrow capabilities to smaller local models just won't be adding to total offensive superpower. 3) Even though Xi might get more AGI-pilled soon, the logic of prestige between the labs + undermining American proprietary moat and frontier margins + accelerating competitive domestic research will hold. Some labs will go closed like Qwen, which will be very celebrated in the US as a sign of Chyna following the lead. Others like Moonshot and DeepSeek likely won't: recall that they cannot address all their demand and their margins are eroded by default by being in the follower position, so they don't actually lose that much from open sourcing. It's more of a difference in corporate culture than a competitive posture. 4) Crucially, Chyna really needs to accelerate R&D. They are vulnerable now. They need to achieve the stage of defender's advantage in cybersecurity, and they will have their own Glasswing but it won't be enough and there will be trust issues that privilege on-prem deployment of powerful models by major companies seeking to harden their systems. Continued open sourcing is the least friction way there. 5) simple cultural inertia. They have already formed this image of a "nation championing open source and open science", providing means of progress to those further behind. This is a talent pipeline for capable labs (in the way that making generational wealth, going to cnc orgies in the Bay Area and serving the Empire is for American scientists), and it'll take a while to give way to security paranoia.

This is a grab-bag of reasons, but holistic grab-bags are, as I often say, more natural to the Chinese way of thinking than Hajnali game theory. I think we will see an open source Mythos at some point in 2027, and it will have Chinese characteristics. Most likely we'll have to download it from Modelscope, not Huggingface.