China Lagging in AI Is a 'Fairy Tale,' Mistral CEO Says (msn.com) 57
Claims that Chinese technology for AI lags the US are a "fairy tale," Arthur Mensch, the chief executive officer of Mistral, said. From a report: "China is not behind the West," Mensch said in an interview on Bloomberg Television at the World Economic Forum in Davos, Switzerland on Thursday. The capabilities of China's open-source technology is "probably stressing the CEOs in the US."
The remarks from the boss of one of Europe's leading AI companies diverge from other tech leaders at Davos, who reassured lawmakers and business chiefs that China is behind the cutting edge by months or years.
The remarks from the boss of one of Europe's leading AI companies diverge from other tech leaders at Davos, who reassured lawmakers and business chiefs that China is behind the cutting edge by months or years.
Who thought they were? (Score:2)
Re:Who thought they were? (Score:4, Insightful)
The consensus I have seen is that Chinese efforts are absolutely on par with what is going on in the US and that its not yet clear if one or the other (or neither) will race out ahead and "win" the AI race.
And what does "win" mean?
Re: (Score:1)
>And what does "win" mean?
Market dominance, succeeding where others fail, competing to such a degree that most users cancel their western AI subscriptions (gemini, claude, grok, chatgpt, etc. if china wins)
Re: (Score:2)
And what does "win" mean?
I think the win state is being the last one standing and establishing the type of dominant monopoly status that Google, Amazon, Nvidia, or TSMC have in their domains. But, even bigger. Is it possible? I dont know, but I think thats the dream of these companies/nations/whatever.
Re: Who thought they were? (Score:2)
For the CEO of this French AI company, it means getting the EU to give him money to restore French pride while claiming it's all about the European economy. See also Quaero, which had the same goal, aiming to replace Google back when search engines were all the rage. Why these guys think playing the "also ran" game is going to get them ahead at anything is anybody's guess.
Re: (Score:3)
Artificial General Intelligence, I would imagine.
Re: (Score:2)
Being able to do significantly more calculations, which will win or lose battles, especially if your AI can handle a lot more in a smaller duration of time.
Re: (Score:3)
Re: (Score:2)
The Chinese are spending much less on it, and on top of that, they have a lot more to spend. It has been debated whether it was the US/SU arms race that broke the SU's back, but here we almost certainly have one that will break the US's back. The US has basically put their whole economy all in on the one chip that is AI.
There's far less transparency into Chinese funding and spending for AI. However, it's unreasonable for the Chinese to have both more funding and less spending on a technology that has been identified as critical on a national level. That makes no sense. It's clear that the Chinese are spending lots of money and resources on AI. Most of the Chinese money is coming essentially (directly or indirectly) from the government, in contrast to the US where the funding is essentially private.
Will this spending "b
Re: Who thought they were? (Score:3)
How is none false? Are any of the firms training models actually profitable yet? Last I checked they were all in the red.
Re:Who thought they were? (Score:5, Interesting)
The Chinese are spending a lot less than the US; and are getting essentially the same results, if not better. Compared to the US they also have a lot more money to put into it, if they should want to. There's nothing unreasonable about that.
The difference between the Chinese state driven approach and the US private driven approach, like in any other area, is twofold. First, the Chinese actually have a plan, and the plan is made by an administration where competence and results are the main driver of careers. None of this applies to the US. Second, within the limited semiconductor supply available, the prices the US is paying inflate to suck up all available funding, however much of it should become available. The inflating part of this does not apply to China, and this is a major part of them spending less than the US. But the inflation part applying to the US, that is indeed a good example of the adaptability of the private model. Expect energy prices to adapt next.
And like drinky already drew attention to in his post - none of the US AI companies and endeavours are profitable. The difference between spending and revenue is 100-fold. And there's no way to profitability anywhere in sight. Spending is only going up with no limits, while revenue is hard limited by competition from the labour the AI is hoping to replace. As of now, even the highest paying customers are losing AI companies money on every query.
Re: (Score:2)
The Chinese are spending a lot less than the US; and are getting essentially the same results, if not better.
there is no chinese nvidia; top performing models are not chinese; deepseek was a splash, not a wave; SMIC doesnt compete with TSMC or even lesser foundries.
these are present day results, the sum product of policy/administration/everything, not what-ifs or projections.
if may very well change, but at present, china leading AI is fan fiction.
Re: (Score:3)
The Chinese are directing much less of their effort into garbage AI products like ChatGTP, and more into specific products like AI systems which will sort recycled materials, drive a quadruped robot, or develop new industrial chemicals for specific purposes. This is both less expensive and faster to develop, as well as producing almost immediate returns on investment. For the life of me I can't understand the reasoning behind the process that the Western AI companies are following, unless it's simply to r
Re: (Score:2)
Re: (Score:2)
It was in 1981, before Ronnie Raygun's first budget.
Re: (Score:2)
China switched to optical processors.
Since nearly 3 years already.
100 - 1000 times faster than GPU/TPUs, and for 1/30th of the energy.
160 tera-operations per second per watt (TOPS/W)
See for example: https://www.youtube.com/watch?... [youtube.com]
Google "Tai Chi optical processor" or "Tai Chi II".
In Germany we have a few start ups that do the same.
Re: (Score:2)
>> Chinese efforts are absolutely on par with what is going on in the US
I use these products daily for coding and in my experience the Chinese LLM's are significantly inferior to Anthropic, Google, OpenAI, etc.
Re: (Score:2)
China seems to be trying more things, while what I see in the West is that they will have one AI algorithm and just keep throwing hardware at it to make it work.
For example, lets say AI uses a function of Big-O of (n^2). If China could get O (n*log(n)), then all the hardware thrown at the former Big-O function is worthless.
Remarks diverge from other tech leaders.... (Score:4, Insightful)
More importantly; does it matter? (Score:3)
If anyone still thinks that LLMs are just suddenly going to Singularity and turn into an archotech omnibrain then, yes, it matters a lot who crosses that line first(or it might; also possible that cyber-god will be totally uninterested in who its nominal owners are and have something else in mind; unknowable alien superintelligences are guesswork like that); but if the trajectory is a sort of muddled meandering toward sucking incrementally less; with it being discovered that lazy drop-in integrations actually work really badly and nontrivial fiddling is needed; it seems like a market where the expensive, controlling, probably-stealing-your-data-for-training-or-to-replace-you first movers are going to be less attractive than the runners up just because those guys are hungrier and more willing to give up control in exchange for money.
Would I rather have the 'best' model, as provided by the guy who is just pinkie-swearing that he's not mining my every interaction with the bot to replace me the way Amazon uses 3rd party sellers to determine what categories to slap a house brand competitor on; or a somewhat less good one from some dude who is hungry enough that he'll give or sell me the weights and let me run it locally or on infrastructure that's just someone else's tech swapping dead GPUs, rather than a hyperscaler reading anything they want?
This is where capitalism actually wins. (Score:1)
OMG RLY? (Score:1)
AI executive, talking about China, at the WEF. Three strong reasons to completely disregard everything said here.
Re: (Score:2)
Yep, code for "We've tapped out private money, now we need National Security money".
Why would the CEOS be stressed (Score:2)
The took a western open-source AI model, (Score:1, Insightful)
I'm not downplaying Chinese technical capabilities. In a lot of fields, they're competitive with the US and they're ahead in some areas. But we're not about to be steamrolled by the invincible Chinese AI juggernaut.
Re: (Score:3)
Which one are you talking about? The current SOTA models from China are all architectures developed by the Chinese companies.
Re:The took a western open-source AI model, (Score:4, Interesting)
Fun fact: The latest Mistral large model uses the DeepSeek architecture. Open source is great.
Re: (Score:1)
Sorry, I simply don't buy it. They tweaked a pre-existing model and gave it a different name.
Just to be clear - most western companies are doing exactly the same thing. Now that the general idea and working OS code is out there, everyone is taking it, and twea
Re: (Score:2)
You can look at the code, because it is open source. No need for conspiracy theories.
You can also compare their model with others. Don't forget that Chinese companies pioneered thinking models (i.e. chain of thought baked into the model). We have to credit Meta for starting the open source movement with Llama 1 and providing good models until Llama 3, but the Chinese then continued to provide the best Open Source models (with Mistral being close, but releasing less often).
Re: (Score:1)
At least part of it may come down to what you might call second mover advantage.
After it's been demonstrated that something actually can be done, it's that much easier for someone else to come along and do it too.
Even if he does it in a different way than the first guy, there's value in knowing at the outset that the task you're attempting is not impossible.
Gap too small to be meaningful (Score:2)
Even if the AI CEOs are right, and China is 6 months behind, that only means that the rate of improvement is more important than the questionable benchmarks they cite. And in that regard, China is far ahead.
Re: (Score:2)
China also has a lot of cheap, clean power, and is installing more than the rest of the world combined every year. While US companies are throwing money at fusion and maybe restarting fission reactors one day, China continues to ramp up production of solar panels and wind turbines, and robots to install them in massive arrays, and batteries to store the energy.
China has a holistic approach to development that is driven by the government setting strategic goals and making sure private companies are doing wha
That's an order of magnitude! (Score:2)
China is behind the cutting edge by months or years
Being ahead by months is trivial, and the competition is on your heels and possibly gaining. Being ahead by years gives you some substantial breathing room.
If they can't pin down the size of the gap better than that, maybe they should get their AIs to figure it out for them. /sarc
Give Me Money! (Score:2)
FTFY
Who cares? (Score:2)
about the winner of a slop race.
Why should they be lagging? (Score:5, Informative)
Every week there are new free models from China.
Today Qwen released a collection of state of the art TTS models.
This week GLM release GLM4.7-Flash
About a week ago was the release of the HeartMuLA music generator from a Chinese research group.
Two weeks ago z.ai released GLM-image.
If you go back a bit more you find GLM4.7 (the large model), z-image-turbo, Longcat-Flash-Chat, Longcat-Image, Qwen-Image-Edit-2512, Qwen-Image-2511, YuE, and many more LLM, TTS, Image, and music models.
The best thing: Most Chinese models are MIT or Apache licensed.
Re: (Score:2)
So salty ...
But other than your post, my post has substance: I named concrete models. What are your arguments? ... you didn't even make a point for or against anything. You just posted an ad hominem.
Wait
Benchmarks (Score:2)
IMHO, the "Best" benchmark which has images as part of its input is MMMU (https://mmmu-benchmark.github.io).
Top 5 on MMMU: GPT4o, Claude 3.5 Sonnet, Gemini 1.5 Pro, Claude 3 Opus, and Gemini 1.0 Ultra.
Best Chinese Model is Qwen-2.5-VL-72B which