I don't like AI but if it is to be this century's arms race, then i am glad that China is pulling ahead, first with DeepSeek now Alibaba Cloud. Turns out if you do it right you don't need to build massive GPU farms that drive up your country's electricity prices to a level unaffordable by the average citizen.
this post was submitted on 20 Oct 2025
53 points (100.0% liked)
Technology
1290 readers
31 users here now
A tech news sub for communists
founded 3 years ago
MODERATORS
What I really appreciate with China's AI approach is their focus on open source models and frameworks, and their emphasis on practical AI applications. That's why almost all of the top open source models (Deepseek, Qwen, GLM, Kimi), are all from China, which means anyone with enough hardware can run it on their own devices.
But if your AI stack uses only one GPU instead of ten, how are five rich men going to promise each other to build 100 gigawatts worth of datacenters and speculatively prop up your dying economy that has a constantly declining industrial output?
US economy is just 5 rich men in a trenchcoat
I was just having a conversation with a relative yesterday about open-source; he's a real blue-collar guy. I told him he should open up any manual for any device he has in the house and flip to the back to look at the open-source disclosures inside. That open-source has already beat the corporate private system because not a single piece of technology gets produced without some kind of open-source software. By the end of it, he really got the potential of open-source as a system for building functional and practical shit. Through the lens of the John Deer tractor right to repair struggle. China shows us that open-source coupled with a socialist worldview outperforms capitalist "competition breeds innovation" any day of the week. You can't tell me there isn't a John Deer farmer who wouldn't pick up the IDE if it meant that they could contribute to the maintenance, functionality, and performance of the $10,000 combine they use every single day. That kind of sector-wide multidisciplinary skill set doesn't exist currently, but it absolutely could.
It's so unfortunate that so much of the software we rely on is unmodifiable, inscrutable code running on some company's servers. It really does not inspire confidence in their reliability and long term dependability, especially how they can just axe the project if they don't feel it's profitable enough. Open source should be the way to go. Ideally, it should be the bare minimum.
It also contributes to a kind of illiteracy of code in laypeople. I see this sometimes in the way video game players talk about a game's code, being like "the spaghetti code teehee" etc., as if this explains every bug and design problem in a long-standing codebase. Since the code is hidden, it's hard to correct them. Maybe it really is spaghetti code. But odds are its problems are more complex than that. And if modding communities tell us anything in games where modding is feasible, it's that players are perfectly capable of fixing or creating workarounds for some bugs, even when constrained by the limitations of modding. And the company is just kinda mediocre at actually addressing issues. This is one of the annoying consequences of it - the NDA'd nature of everything means companies can hide behind "development is hard" narratives to excuse their institutional failures.
Open source not only furthers knowledge, but also enables some degree of accountability.
I wish I had 2TB of ram to run Kimi
Excellent share; thank you
Thanks for the share; the PRC makes it so much easier to have revolutionary optimism.
One of the things that keeps me going as I organize with my comrades locally is that China exists and someone is out here doing things that will benefit humanity in the long run.
The future is Chinese. Always was.
Every News I get like this, I wish the USSR was around and also doing stuff like this. Those brave Soviets gave everything they had to make a better future for the ones that will come after. It's heartbreaking to see it today
I appreciate the sentiment.
Socialism is humanity's achievement; as you have alluded to the PRC is the synthesis of the contradictions (in the Hegelian sense) of the USSR including its successes and tribulations - we must not let the borders drawn in blood by the bourgoisie dampen our revolutionary optimism. If the west doesn't cause a nuclear holocaust then socialism, once again, is inevitable for Russia.
Short on details but if you could deploy this to a single GPU you could instantly jump from a 12B model to potentially 80B or more, depending how much space the model takes up.
It's pretty clear China is leading (if not the only country interested in) model optimization. I also saw previously that Huawei managed to quantize models in a way that retained similar performances but quantized them to tiny values, making them take up a lot less Vram to run.
I thought this was more related to running multiple models at a time on the same hardware versus increased sizes and whatnot. I might be mistaken but it seemed like it was finding a way to use the same hardware to run more things at the same time or better scheduling queues, allowing a GPU to process per-token instead of pinning a model to a particular GPU.
The article mentions "Packing multiple models per GPU", but also "using a token-level autoscaler to dynamically allocate compute as output is generated, rather than reserving resources at the request level" which I'm not sure what that means but may hint that there are ways to scale this down, possibly.
If not Alibaba then other researchers will eventually get to it.
Holy mother of technology, this sounds too good to be true but there is a paper on it so...
I still can't believe it lol, well now we know for a fact that China is ahead in the AI race.