What makes you think this is the best algorithmic innovation from DeepSeek or China, for that matter?
Artificial Intelligence thread
- Thread starter 9dashline
- Start date
The AI industry is moving so fast, there is no scope to "save" your breakthroughs. Its use it or someone else will type of situation.
I’m impressed by Xiaomi, their first try and they outperform many companies who focus only on a that sector, they did same with EVs outcompeting most ChineseEV makers who have been in car business for years . Impressive company overall .
I
If so, maybe it would have been better to wait a few more months to release a more advanced ready made version with atlas 950?
I'm sure painters didn't view photographers very kindly in their early days either.
I think the implication of holding back launches to save algorithmic innovation is answered in your comment. If DS doesn’t launch its model, then someone will eventually use their innovation, like mHC, to innovate upon. There is no use in saving your breakthrough for a better launch; it is best to publish the result and work on the next frontier.
DS founder did claim to be pushing for China’s own AI stack sufficiency, and in doing so, they trained on Huawei hardware to understand all the quirks of training on an independent hardware stack. In all honesty, DS V4 is less about algorithmic innovation and more about the hardware front. They did just that and are probably in the best position to pull off training on their own hardware stack similar to Google (not sure about design, but they were originally quant, so they can't be that bad at designing hardware accelerators).
Also, this has massive implications for other Chinese AI teams if the political situation calls for a migration to Huawei infrastructure. In that case, Huawei will have this wealth of experience working with a frontier lab on how to provide, roll out, and optimize infrastructure for an AI launch. This can help if other teams decide to also move away or be less reliant on Nvidia's hardware in the future.
Last edited:
The way Nvidia has a stranglehold over AI training and the stack is crazy. It so hard to break that without falling out of the race altogether, so it’s a tough one for China , and requires a lot of careful policies not to cut off totally from their stack, even despite the sanctions , since on a normal day China should not even be using Nvidia due to security risks as well
Longsight's 65 GHz sampling oscilloscope came at the right time, just as 1.6T optical transceiver deliveries are stepping up.
there are always things that US labs can learn from Chinese AI labs, but they tend to train larger models. There are definitely things that Chinese AI labs can learn from each other's papers. DeepSeek, Kimi and GLM are all learning from each other. ByteDance and Tencent also benefit from all the stuff open labs are sharing with rest of the world.
AFAIK from what is being rumored around on Chinese internet, V4 still used Nvidia cards for training and only use Ascend cards for inference.
It's very likely still trained on western hardware, just that inference now use Chinese cards which is a big step forward but not as big as what most people are trying to make it out to be.