Artificial Intelligence thread
- Thread starter 9dashline
- Start date
Just wanted to share this article here, because I've been paying attention to open source projects in Github lately and there's a lot of LLM AIs popping up lately. There's a possibility that open source LLM will be good enough or even better than OpenAI's ChatGPT in the future, allowing consumers and business to have similar capabilities without the data privacy risk and lower cost.
Huawei unveiled updates to fully domestic Ascend platform (w/ Atlas-900 AI cluster) including MindSpore2.0 ML framework, Ascend C language & Atlas 200I developer kit
Ascend platform supports 30+ big models, representing > 50% of native models in China. It supports 25 stable days training with 200 billion parameters. This is quite the market share
Currently, 25 smart cities have been built using Ascend w/ 14 of them launched & fully operational According to HW, more than 3.5m Kunpeng/Ascend developers, > 5600 partners & 15500 certficiations
Things to consider about Kunpeng & Ascend. They really built this up over a few years. They have a fully ecology behind it with hardware Atlas-900 which was state of art back in 2020 when it first came out, but is now even better hardware wise with greater computation, better supporting infrastructure and software integration and such.
Kunpeng is used everywhere now with wide spread software support.
There are also things like this that by using liquid cooling, they can 4 to 8x computation density and decrease size of datacenter space by 70%
One example of big model is 紫东太初（Omni-Perception Pre-Trainer）by CAS. Can learn from text, pictures, voice, video, 3D point cloud, and sensor signals,
Huawei unveiled Atlas 200i DK A2 Ascend AI developer's kit valued at just 1999 RMB, offering 8 TOPS * 4GB storage with RJ45 network port, HDMI & USB 3.0. It uses 4 core CPU
Looks like a great way to train developers to the real thing which is Atlas 900
Looks like a great way to train developers to the real thing which is Atlas 900
Even better article here on what was unveiled at the Ascend/Kunpeng developer conference
My guess is that the developer kit uses some nerfed version of Ascend-310 that probably can be manufactured by SMIC's 12/14nm process and then paired that up with a 4-core Kunpeng-920 CPU
openGauss database and mindspore AI framework all sound pretty well developed now. That's kind of HW's advantage over other Chinese tech firms. They've just been doing this for longer and also have more R&D resources to through at it.
My guess is that the developer kit uses some nerfed version of Ascend-310 that probably can be manufactured by SMIC's 12/14nm process and then paired that up with a 4-core Kunpeng-920 CPU
openGauss database and mindspore AI framework all sound pretty well developed now. That's kind of HW's advantage over other Chinese tech firms. They've just been doing this for longer and also have more R&D resources to through at it.
There are plenty of countries in the world that do not feel comfortable with a (politically) free weelding AI Chatbot. If China can create a 'politically reliable' AI Chatbot it will be a huge growth industry. The only way for the Americans to compete against that is to censor their own versions for the export market.
Why AI ghostwriting is another doomed fad:
- Media outlets based on advertising need clicks to sell.
- But clicks are in short supply, and a few huge web platforms (especially Google) control most of them
- This forces journalism platforms to embrace gimmicky clickbait strategies that are doomed to failure—and are unlikely to deliver quality writing.
- That’s because gimmicks are just that—and they eventually get exposed. But as one clickbait strategy dies, another one replaces it—in an unending cycle.
- The latest (and probably the final act) in this farce is the AI-generated article, which reduces the cost of clickbait articles to almost zero. This is the endgame—the most brutal trick of them all.
- It will lead to a proliferation of garbage articles like you’ve never seen before. Media platforms will churn them out as fast as they can to grab those last remaining clicks from easily deceived readers.
- But this is a fool’s game. Meanwhile, there’s a better strategy in town—namely the subscription model for quality journalism.
- All the success stories in media right now are funded by subscriptions. The business model changed while the experts weren’t looking, and went back to basics. We don’t need to worry about gimmicks—we’ve returned to selling writing, not ads.
- As final proof of this, note how many of these success stories (especially on Substack, but elsewhere too) are built on individual writers, not media brands. This alone tells you how central writing has now become in the growing sector of American journalism.
Feels like a mistake, instead of trying to squeeze themselves into the market, they could have tried to partner up with all the current big chinese players.
Not only might it actually try to stiffle their AI developments, but also avoid in having to engage in actual competition, not to mention competition in phone making, which Baidu should be new compared to the already big players.
btw, from that wsj article again
i must stress again, that 1999 RMB developer kit is a big deal. Great & cheap way for new developers to setup and train on their own GPU hardware. Training software developers is more critical than anything else. HW appears to have a leg up here with over 3.5 million developers in its community already. Also probably why all the independent smart city projects use HW products. Because that's what all the developers are using
i must stress again, that 1999 RMB developer kit is a big deal. Great & cheap way for new developers to setup and train on their own GPU hardware. Training software developers is more critical than anything else. HW appears to have a leg up here with over 3.5 million developers in its community already. Also probably why all the independent smart city projects use HW products. Because that's what all the developers are using