there is actually a pretty good way for you to verify these claims. You can just test it out on their website. I have recently switched from DeepSeek to Qwen for my work related questions.
Artificial Intelligence thread
- Thread starter 9dashline
- Start date
Important to be skeptical with doomer takes, even if risks are non-trivial.
Hinton has been pretty notorious for years but there is rarely any follow-up. Glad people are taking a more critical stance.
---
I tried the Qwen-3 MAX model. I asked it a very simple question about local traffic here in Sweden. It got basic things badly wrong. This is a 1 trillion parameter model.
I asked Claude Opus 4.1 (non-thinking) without search the same thing and it 1-shotted the task. The Qwen team still has some ways to go.
The task I gave was not advanced in any way. At that size, they should have enough world knowledge to get it right. Even when I turned on search, Qwen 3 MAX couldn't complete the task even if it improved and got some aspects right.
In my casual testing, Deepseek is still the best Chinese lab. Moonshot Kimi takes the number two slot.
Hinton has been pretty notorious for years but there is rarely any follow-up. Glad people are taking a more critical stance.
---
I tried the Qwen-3 MAX model. I asked it a very simple question about local traffic here in Sweden. It got basic things badly wrong. This is a 1 trillion parameter model.
I asked Claude Opus 4.1 (non-thinking) without search the same thing and it 1-shotted the task. The Qwen team still has some ways to go.
The task I gave was not advanced in any way. At that size, they should have enough world knowledge to get it right. Even when I turned on search, Qwen 3 MAX couldn't complete the task even if it improved and got some aspects right.
In my casual testing, Deepseek is still the best Chinese lab. Moonshot Kimi takes the number two slot.
Last edited:
Probably what happens in every field, AI doesn't really kill jobs, it just makes jobs easier and you can do things faster.
I think all this AI killing jobs doom gloom will turn out to be wrong in hindsight.
Yeah, Radiology is not a good comparison since healthcare has to be well regulated.
There’s also the issue of data. There’s only so much labeled radiology data out there and you can’t generate more without radiologists and clinical studies.
You can easily generate more programming data, because you can always just run the code to see if it works, so as long as you have compute, you can generate an infinite amount of training data for programming.
Last edited:
And also privacy concerns. Healthcare data is thoroughly protected.
I haven't gone too deep with these ai tools but they have a very generous free cli tool similar to gemini cli or claude code.
I tend to agree that no Chinese model is yet at the level of Claude at agentic programming; likely there's some ingredient that is still missing.
But at the same time, I don't think Chinese models targeting the domestic market will need to worry about competing with Anthropic in the near future:
Basically ANY company that is >=50% owned by a company based in China is banned from using Anthropic API services (and since Cladue is entirely closed source, there isn't anyway for anyone else to host it). This goes further than China's own ban because you can't bypass it via corporate VPNs, side contracts, offshore hosting, etc. The only way to get around it is having a personal subscription and using a VPN but with how expensive Claude is, I don't think it's going to be a common thing.