Artificial Intelligence thread

[Bellum_Romanum]

that's nonsense. you are giving free data to Anthropic to make their model better.

All this is bullshit. Don't try to make it sound like you support China's AI effort, because you are doing none of that.

If China's AI is good enough, then you should use it. Otherwise, there is a gap and you are helping to make it wider.


if we are operating under the assumption that Chinese models are a lot worse, then okay. But if the Chinese models are good enough (which they are), then these people are clearly helping the other side. Of course, they probably don't think about it that way. But now that US govt is shutting off access to them, good riddance.


DeepSeek really hasn't changed the game.


again, if Zai is actually good enough (and this article is talking about GLM-5.2), then you should use it. Stop telling other people to use it if you are not willing to use it.

I actually used it so much this week, that I hit the limit on my coding plan, so I'm back to Kimi 2.7.

Until you are willing to put your coding and money where your mouth is, keep quiet about how good the Chinese models are. You don't know because you don't use them.

I've used DeepSeek V4 and it's really not up to par for coding. Minimax, I tried that too and it's bad. Chinese models really only got good at coding with GLM-5.1 and Kimi 2.6.

I’m still using pretty much everything you mentioned, including Anthropic until the end of this month. The main reason is simple: until recently, most Chinese open-source models just weren’t that strong at coding. Kimi 2.6 and 2.7, along with GLM-5.2, were the first ones that actually felt competitive. I was pretty underwhelmed by DeepSeek V4, but I’m looking forward to its next update.
On the paid side, I used to subscribe to MiniMax, Kimi, and GLM, but I’ve since consolidated down to just Kimi and GLM. For the free Chinese models, I’ve tried DeepSeek, Stepfun, Yuanbao, and Doubao. On the Western side, I was paying for Grok, Perplexity Pro, and Anthropic. I also have access to Gemini Pro through my cloud bundle, and I’ve been using NotebookLM quite a bit. Work pays for the premium version of Microsoft 365 Copilot, so I use that as well.
On the agent side, I installed OpenClaw and then added Hermes Agent. I also have a subscription through OpenRouter and Nous Research.

 

[tphuang]

“Training runs on physical supernodes, with each supernode having a maximum of 48 machines. The nodes are fully interconnected with high bandwidth, and the nodes communicate with each other via RoCE networks.”

100+ Atlas 900 A3 SuperPoD, only this one supports exactly 48 servers forming a node.

A3 SuperPoD uses Ascend 910C chips, which have significantly larger than H800's memory, so that does not fit their description in there. I guess it's possible they are just pointing to 910B for the memory part, but each chip operate as 2 dies + 128GB of memory, so that doesn't make sense for me.

 

[HighGround]

Please, Log in or Register to view URLs content!

Reports are just today, Z.ai finished trained a model that meets or exceeds Fable/Mythos on cybersec.....

I am not sure if it will be opensourced but basically chatgpt literally also announced a cyber version of their 5.6 just a day ago

It probably wont even take China but a few more months to get to Mythos tier, broadly

Also Deepseek released a spec today that makes inferencing even more efficent up to 500%

Please, Log in or Register to view URLs content!

Writing is on the wall

Tired of this "Mythos tier" nonsense. It's just marketing.

 

[bsdnf]

A3 SuperPoD uses Ascend 910C chips, which have significantly larger than H800's memory, so that does not fit their description in there. I guess it's possible they are just pointing to 910B for the memory part, but each chip operate as 2 dies + 128GB of memory, so that doesn't make sense for me.

910C is essentially two 910B chips on the same board. Each 910B chip has 64GB HBM and a 200 Gbps RDMA network, the latter being directly mentioned in Meituan's technical report.

Adding 48 servers as a supernode, with two completely identical parameters plus one partially matching parameter, the confidence level is high enough as there are no other closer options.