In all the past great gold rushes, it was usually the "merchants, supply shop owners and salon operators" that laughed at the last. Looks like history is about to repeat itself again.
Artificial Intelligence thread
- Thread starter 9dashline
- Start date
'
They scared of others catching up or something?
They scared of others catching up or something?
Students at SLAI, a pilot AI educational institute in Shenzhen launched just last year, successfully performed SFT post-training of DeepSeek v4 Pro on a cluster of 1k Ascend 910C chips. They achieved an MFU of 34.9% with 3k training set of operations research and mathematical modeling questions, improving performance on select benchmarks
Although small in scale, this is a data point showing the maturity of the Ascend tech stack. With newer versions of Ascend chips at scale, complete pre-training and post-training will be possible.
Projects like this also give students real world AI experience using domestic AI chips and LLMs, strengthening the ecosystem. This will help diffuse knowledge and expertise when they enter industry.
Although small in scale, this is a data point showing the maturity of the Ascend tech stack. With newer versions of Ascend chips at scale, complete pre-training and post-training will be possible.
Projects like this also give students real world AI experience using domestic AI chips and LLMs, strengthening the ecosystem. This will help diffuse knowledge and expertise when they enter industry.
yes, I posted about this also. I think this is more to just show it can be used for post training. Although this is still not optimal compared to Atlas-950, which is far more powerful. Just 1 box is 8192 chips.
Last edited:
That's certainly one way to announce they're building bioweapons for the US government
They're scared of running out of money trying to compete with superior Chinese AI with brute force compute, they want other brute force American labs to slow down because they have enough braincells to know sooner or later people will stop paying $10k a month for something they can get for practically free from Kimi/DS
Huawei Cloud announces its Token factory with 200 EFLOPS compute and supporting 20 local partners, including all the big name open source ones (other than Xiaomi)
Using Claude to manage inboxes, meetings, and calendars is completely pointless. I'm surprised they've managed to stay in business and are only now switching to Deepseek.
They were most likely using Haiku or Sonnet without thinking before; their quality, speed, and price were absolutely crushed by V4 Flash.