Xiaomi is now releasing a SOTA model for global languages.
Artificial Intelligence thread
- Thread starter 9dashline
- Start date
Xiaomi is now releasing a SOTA model for global languages.
So according to this interview of an AI expert with connections to top Chinese labs, Chinese chips like Ascend chips can already substitute nvidia for pre-training for top labs including DeepSeek. DeepSeek v4 was likely trained with a mix of ascend and nvidia chips then.
Interestingly, he says that post-training RL still heavy relies on nvidia, but domestic inference chips can be used in certain steps. And also very ironically, CUDA's ecosystem advantage has been eroded by AI assisted coding which makes adapting CUDA kernels to domestic chips much easier.
Interestingly, he says that post-training RL still heavy relies on nvidia, but domestic inference chips can be used in certain steps. And also very ironically, CUDA's ecosystem advantage has been eroded by AI assisted coding which makes adapting CUDA kernels to domestic chips much easier.
Misanthropic and ClosedAI are democratic models that steal user data to bring freedom and justice to the world. Chad-Xi-PT and Kimchi are authoritarian models that steal user data to brainwash and mislead people around the world. I would much rather pay 250 million dollars to support the freedom models.
This reminds me of the week before Feb 28 when they were trying to hype up the idea of OpenAI/Anthropoic being used in US military planning. They certainly got real silent when that war went sideways almost instantly, in no small part because US "hallunicated" a fantasital Iranian non-response scenerio.
But then again I guess in a way that did bring freedom and justice to the world.
Are they not banned from buying Nvidia chips ? There is no way they can smuggle such vast amount of chips required for such a huge project without US authorities being unaware . So I don’t get how they will smuggle those chips
The US authorities were fully aware of this.
Yes, they were fully aware.
They couldn't completely prevent Chinese manufacturers from obtaining Nvidia GPU; enforcement costs are too high if done this way. They occasionally arrest local smugglers as a warning to others, using a small cost to maintain the high compliance costs of Chinese companies.
I understand small scale smuggling can go through underground . However , I am not sure such a large scale purchase can be done under the radar?
sibce Bytedance will need billions of dollars worth of Nvidia AI chips to train their models which wither will running in this massive data centers they are investing in overseas
Why would they need to smuggle those chips? The chips are in Thailand and bought by Thai companies. They are just being used by ByteDance. And ByteDance is not just using it to train models but also run inference.
partnership between China Mobile and Volcano Engine to put Doubao/seed large model on China Mobile Cloud, including Seedance 2.0 and Seed 2.0. MaaS service, Agent development platform and tools to provide reliable AI service. HiAgent and Arkclaw tools
So we finally have a METR estimate of Mythos.
FYI, Mythos seems to be trained to be specifically good at software. In non-software benchmarks it didn't do much better than GPT5.5 except in HLE (likely due to Mythos being much bigger in size).
Dario himself said that he expect Chinese labs to reach this capability in 6-12 months. So the rest of us shouldn't have to wait very long. There is also continual algorithmic improvements, so it's not clear you really need a model as big as mythos to reach the same capability a year from now.
FYI, Mythos seems to be trained to be specifically good at software. In non-software benchmarks it didn't do much better than GPT5.5 except in HLE (likely due to Mythos being much bigger in size).
Dario himself said that he expect Chinese labs to reach this capability in 6-12 months. So the rest of us shouldn't have to wait very long. There is also continual algorithmic improvements, so it's not clear you really need a model as big as mythos to reach the same capability a year from now.