Artificial Intelligence thread

[tphuang]

Weekly token share of Chinese models used by US firms.

Please, Log in or Register to view URLs content!

OpenRouter is such a small % of actual AI consumption that quoting it is just not useful

https://twitter.com/i/web/status/2078103148665667648

this is okay, seems reasonable to treat GLM-5.2 as around Opus-4.6

In an interview, the head of MINIMAX research stated that their fable-calss model, the M3, will be released within six months.

Kimi K3 has already been released, and Deepseek V4 Pro, also a fable-class, is about to be officially released. Z.ai's GLM-5.5 is rumored to be released next month as well. "Within 6 months" means they are way behind schedule.

Minimax just sucks. Until they come out with this 2.7T model, it's just all talk to boost stock prices.

It remains to be seen how good V4 official looks to be. I wouldn't expect it to be better than K3. The next GLM should be.

 

[tankphobia]

I don’t understand why it’s such a stretch that China keeps models secret when we know that the U.S. does the same thing.

It's probably secret in the same way that Anthropic has models reserved for the US government that is ahead of the curve. Training AI at scale is basically grabbing a huge pile of money and setting it on fire, so I don't think the Chinese government will be spending so much money and computing reserved for military research when instead they can just co-opt a model from major local AI firms. In exchange they're probably helping out with training data which is not available to the pubic.

 

[tphuang]

so the Saudis met with Moonshot this week. I'm expecting investment. They are also getting investment from ByteDance on Data center. Meituan is investing there for unmanned delivery.