They are overall only about 1-2 years behind Nvidia's rate of development. The only problem is that the relevant software support is still not so great compared to Nvidia (Similar to what you would expect trying to develop AI using AMD chips).
Artificial Intelligence thread
- Thread starter 9dashline
- Start date
Are there any objective comparisons for these chips, based on some of the more common metrics?
It's extremely difficult for one to compare AI chips using common metrics. Sure, one can just lay out the floating point computing speeds and memory bandwidth, and these are good indicators of single chip performances, and Chinese chips is still 2-3 years behind in these metrics, but once you start trying building a supercomputer to develop state of the art models, the traffic architecture between the chips becomes much more important than raw numbers and Ascend chips are good at scaling like this. As I said, the main problem is not with computing power, its that these chips are still much more difficult to learn and use than Nvidia chips which gets the best of everything in terms of support.
Scam Altman now delays their open sourced model lol
I'm going to have to test this out if coding is as good as they say it is.
Right now, I use deepseek R1 to generate simple python coding, but it takes a couple of minutes to generate result for simple requests. Also, the deep research I tried on Qwen wasn't great. I have kimi on my phone already, so easy place to get started.
tbh, cursor (which uses claude IIRC) is already pretty good at filling out code that I want to type. I'd be curious if I can get an IDE that uses one of the Chinese models. I'm paying monthly fees for cursor.
Right now, I use deepseek R1 to generate simple python coding, but it takes a couple of minutes to generate result for simple requests. Also, the deep research I tried on Qwen wasn't great. I have kimi on my phone already, so easy place to get started.
tbh, cursor (which uses claude IIRC) is already pretty good at filling out code that I want to type. I'd be curious if I can get an IDE that uses one of the Chinese models. I'm paying monthly fees for cursor.
He's right, you know.
Open Source AI is basically a Chinese game right now. Meta was the only Western company that could do it at a decent level in the past, but they fell behind. Hopefully their new team will fix things. Mistral is an also-ran. Seems obvious sama delayed because of K2.
Open Source AI is basically a Chinese game right now. Meta was the only Western company that could do it at a decent level in the past, but they fell behind. Hopefully their new team will fix things. Mistral is an also-ran. Seems obvious sama delayed because of K2.
How good is Deepseek compared to OpenAI now? Or is there a different Chinese model on top
The American AI companies pay lip service to open source but it is pure PR. In reality they are 99.999999% focused on building up and maintaining walled gardens and captive markets based on proprietary models. Chinese open source ecosystem and low AI costs (energy or otherwise) are the ultimate weapon to eventually win dominant market share globally. Western companies cannot even help but to outsource decent white collar work to India just to save a little money in the grand scheme of things. Are we really supposed to believe they can universally and forever stick with extremely expensive American AI? Especially with their energy generation bottleneck, cost of energy infrastructure, slow build times, high energy costs, etc.
I think their latest version is as good as OpenAI latest models.
Last edited:
Any whispers on when the next Deepseek model is being released? Was supposed to be May, but still waiting. Trying to time buying put options on NVDA and semiconductors. Maybe just look for large put option contracts for NVDA since it may be the hedge fund side of Deepseek making a killing again.