Futurology

2514 readers

16 users here now

founded 2 years ago

MODERATORS

Outperformed by Chinese Open-Source AI, US firms want their government to ban it. (self.futurology)

submitted 1 month ago by Lugh to c/futurology

12 comments fedilink hide all child comments

Article with overview.

OpenAI & Anthropic have both made calls for Chinese AI models to be banned in the US on national security grounds. While it is true countries have reason to distrust other countries' tech, I doubt this is the real reason they are upset.

Their big problem is that Open-Source AI annihilates their chances of succeeding as businesses. Silicon Valley's model of VC funding is to bet on many small start-ups, hoping one becomes a 'unicorn' - a multi-billion dollar company (like Google, Meta, etc) able to dominate an industry and rake in hundreds of billions of dollars.

Even if they succeed in banning Chinese Open-Source - does this mean they'll become unicorns? I doubt it. The Chinese Open-Source AI models are superior to theirs. Most of the rest of the world will use them, and the real AI innovation will happen in the rest of the world. Meanwhile Americans will make do with the second-best AI, that can only survive when it gets the best banned.

top 12 comments

sorted by: hot top controversial new old

[–] regrub@lemmy.world 15 points 1 month ago (1 children)

Protectionism won't prevent them from failing. They're already struggling to keep the bubble from popping

[–] Lugh 7 points 1 month ago

Exactly, At best, it's a stay of execution.

[–] AfricanExpansionist@lemmy.ml 4 points 1 month ago (1 children)

Deepseek is FOSS, hosted on github. Anybody can download and modify it however they want to, fork the project and give it a new name

[–] heavydust@sh.itjust.works 2 points 1 month ago (1 children)

Stop calling that open source. It’s an insult to what the movement is all about.

Open model or free weight at most.

[–] AfricanExpansionist@lemmy.ml 1 points 1 month ago* (last edited 1 month ago) (1 children)

It's released under the MIT license. Deepseek's marketing literature says "distill and commercialize freely".

The code has been modified already and I have run one of these versions on my aging hardware.

You can go to github and read the code license AND the model license yourself.

You can read more about it at Contractnerds.com, or there are lots of Reddit threads about it. Indeed, Contract Nerds make the point that Facebook's Llama is open weight while Deepseek is much freer to use

So AFAICT, you're just completely wrong.

[–] heavydust@sh.itjust.works 1 points 1 month ago* (last edited 1 month ago) (1 children)

It’s released under the MIT license

There are 5 Python scripts with the MIT license. The huge binary blobs are not included. You know, binary, the thing that is by definition and by its content NOT open-source. Or Maybe Windows is open-source too?

[–] AfricanExpansionist@lemmy.ml 1 points 1 month ago

Yes, you can just run their models on other software, which is also open source.

I guess you're talking about the Deepseek online product? That's NOT FOSS, AFAIK

[–] just_another_person@lemmy.world 3 points 1 month ago

Can't ban "software". Idiots.

[–] A_A@lemmy.world 2 points 1 month ago (1 children)

Yes, + you can install DeepSeek locally

(...) DeepSeek’s open models don’t contain mechanisms that would allow the Chinese government to siphon user data; companies including Microsoft, Perplexity, and Amazon host them on their infrastructure.

Also, Alibaba's model is better

... Qwen 2.5 ... as far a.i.k.

[–] juli@lemmy.world 2 points 1 month ago

Ann both likely do not follow robots.txt standard. Like most AI scrapers.

[–] Etterra@discuss.online 2 points 1 month ago

LOL nah.

[–] sinceasdf@lemmy.world 1 points 1 month ago

I agree with some of the overall premise, open weight models will put immense financial pressure on the ai-only companies doing the foundation training like openai and anthropic. But lmao at the notion that ANY open source models are better than the existing cutting edge ones from the big players. The difference just doesn't matter for many applications.

Chinese or not doesn't matter for this, releasing open weight models is a strategy for those whose cutting edge models are not the absolute best. Metas entire strategy has been open weight since day 1. It's not benevolence. They're propaganda/ad machines and they want people to use them. If you've got to pay someone for that little extra reliability it's still probably going to be anthropic, openai or Google.