Teortaxes argues Chinese companies open-source competitive AI models because they face Western API resistance and lack proprietary secrets

@francoisfleuret Why not? They don't have much secrets Americans don't know, and they can't serve all the expected demand (partially because Westerners won't use their APIs, partially because they don't have the compute). It's a surprisingly cheap move that wins goodwill/attracts talent.

Ah "they can't serve all the expected demand" sounds like a very good explanation.

@francoisfleuret it also probably erodes frontier margins and slows down Western labs, but I don't buy that argument much because the frontier still rakes in cash by the billions

@ptremblay You are joking but it would make sense that companies fight in the open-source and close-source space, with a strong differential between the two.

For some reasons the Chinese companies are competing in the open-source space with the heavy stuff.

@tugot17 @francoisfleuret They try Singapore, Malaysia, Japan. Tencent, ByteDance use a lot of overseas compute. They're also the weakest open source players. but ultimately if you're a PRC actor you can get designated a security risk (Zhipu is on the entity list already) and you can be evicted too

@francoisfleuret n-no... you can't just release a better model and make it free! i wanted to suck people's wallets dry with my saas models first! p-please don't publish those weights it will crash my stocks!

@francoisfleuret Releasing huge models = free promo from enthusiast. Mayority of open llm fans do not have the infra to run it “at home” so they spend money on api.

If a chinese provider releases a non-open model right now, they get bad reputation from community and no usage over api (qwen)

@farstpaul https://x.com/search?q=GML-5.2

@francoisfleuret yeh why. why not release weak models? 😆

@francoisfleuret Which one was released while I was asleep ?

@francoisfleuret They want to pave over American moats, to avoid getting locked in Dario’s trunk. Dario and others have been very open about their goal of creating winner-take-all and then using AI to overthrow the Chinese government.

@francoisfleuret Sadly for freedom-loving westerners, the CCP has rather informed and scientific industrial policy that decided a while ago China should lead in AI.

What Zuck could not do, the Chinese have done with enterprise diversity and perhaps government support.

Publishing algorithms too.

@francoisfleuret Chinese companies don't necessarily need to operate to be profit driven like American companies do since they can benefit from government intervention.

The CCP can subsidize open-weights to drive down the profitability of American closed-weight labs over time if desired.

@teortaxesTex @francoisfleuret but why can't exacly xiaomi rent from a US based neo cloud and serve the model through idk corewave on Blackwell to get good margins on the API?

@francoisfleuret So companies don't have moat

@francoisfleuret This assumes that the Chinese open-weights models can match the frontier. Then it becomes a race to the bottom for cheap API costs. Even if China doesn't capture that revenue directly, it cuts down the revenue driver of American labs. I say this as a supporter of open-weights.

@francoisfleuret Open source opens the door to faster more divergent thinking.

It outsources exploration by default. R&D is done by the community.

@francoisfleuret They are an export driven economy, but people won't trust their data to a Chinese data centers, so they can't compete here. By releasing open source models they cut the profitability of other AI labs which limits their ability to build a decisive advantage.

@francoisfleuret So if you are not oai, ant or google

Release closed model = no usage, no good reputation

Release open weights = usage over providers + usage over api/subscriptions.

@francoisfleuret Same philosophy as that of Chinese products. Cheap producers, global reliance. Amodei chose “I have a nuke, and only I can tame it” and got slapped. The Chinese are saying “you want good model? I can make it”.