Artificial Intelligence thread

[sunnymaxi]

An Ai token factory is planned to be built in Wuxi, China.

Under the goal of "domestic chips, domestic models," it is planned to start with an initial deployment of four Huawei CloudMatrix 384 systems. This will be the largest scale among token factories in China.

Please, Log in or Register to view URLs content!

 

[siegecrossbow]

Please, Log in or Register to view URLs content!

 

[tphuang]

https://twitter.com/i/web/status/2056169826062622924

huge GW class underwater AI Data center with 260 EFLOPS of compute planned using 95% green energy (basically offshore wind)

 

[Shadow_Whomel]

https://twitter.com/i/web/status/2055609308922909117

Huawei and China Unicom in Great Bay Area putting out collaboration to improve speed of AI token generation for mobile end users.

Hi tphuang，did I miss something? The screenshot only shows the difference in upload attachment speed, not the token generation difference.

 

[siegecrossbow]

Please, Log in or Register to view URLs content!

Please, Log in or Register to view URLs content!

I’m seeing a trend here. Maybe stop gloating about AI taking young people’s jobs???

 

[tphuang]

https://twitter.com/i/web/status/2056854679464595588

Another part of AI supply chain that Chinese producers are really gaining is the PCBs and the CCLs used in making them.

 

[tphuang]

https://twitter.com/i/web/status/2056839005518397731

by the way, I was involved in this project for a while and I can say for sure that this is a problem facing US AI industry as a whole. China practically has a monopoly on affordable embodied AI chips.

 

[tphuang]

https://twitter.com/i/web/status/2057081318081482818

interesting news today from Alibaba across the board on Qwen 3.7, the full AI hardware stack and its Bailian platform. Note the language here

在百炼推理平台，阿里巴巴构建了大规模GPU资源集群，并通过并池调度实现资源高效利用。
在此基础上，百炼通过上下文缓存消除重复计算，借助吞吐弹性调度机制应对流量波峰波谷与负载波动。在效果优化方面，百炼引入了Agentic RL，基于Agent执行反馈的强化学习机制，驱动模型持续迭代。此外，百炼内建了安全治理能力，确保自主运行的Agent始终不越界

It really emphasized its own GPU resource cluster which can be used efficiently and introduced Agentic RL which can likely be used in improving various models through post training. But no mention of DeepSeek in its selected partners. They included everyone else.

 

[tphuang]

https://twitter.com/i/web/status/2057001117393785325

I couldn't find this yet on the artificial analysis website but here is claim that Qwen 3.7 scores quite well, even better than kimi 2.6. Along the line of GPT-5.4

 

[mossen]

I've been harsh on Qwen's big models in the past because they were pretty bad, but 3.7 Max seems like a genuinely good release. Finally their big models is approaching the much better performance of their smaller models.

Their blogpost is up here:

Please, Log in or Register to view URLs content!

I look a lot at long-context and hallucination rates. Artificial Analysis does test for hallucination rates and we also have stand-alone benchmarks like Halluhard. For long context, Qwen 3.7-Max is best-in-class at MRCR-v2 (128k).