China Just Erased America's AI Lead
- Reference: 0184496032
- News link: https://slashdot.org/story/26/07/17/1820200/china-just-erased-americas-ai-lead
- Source link:
> Kimi K3, a massive new model by Beijing-based Moonshot AI, threatens the foundations of Americas AI boom. Its release Thursday dazzled developers, jolted Silicon Valley and [2]reset the AI race overnight . Kimi immediately vaulted into the top tier of global AI, beating Anthropic's Fable 5 and OpenAI's GPT-5.6 Sol in front-end coding tests by [3]AI evaluator Arena .
>
> In [4]Arenas broader text ranking , Kimi finished ahead of Anthropics Opus 4.8 -- the company's flagship model until Fable 5 arrived in June -- while costing 40% less. Unlike the premium U.S. models its challenging, Moonshot plans to release Kimi as an open-weight model on July 27 -- allowing companies and governments to customize and run it on their own systems.
>
> Kimi's arrival suggests that cushion may have collapsed far faster than expected. "The entire game has changed. I expect this will trigger some code red for some," AI analyst Kim Isenberg [5]predicted . For companies, governments and developers, a model that performs near the frontier, costs 40% less and can be customized or run in-house may be the more attractive option. Its very existence puts pressure on the pricing power of U.S. labs, the enormous valuations built around their technological edge, and the case for spending hundreds of billions of dollars on ever-larger data centers.
[1] https://slashdot.org/~schwit1
[2] https://www.axios.com/2026/07/17/china-ai-kimi-k3-open-source-anthropic-opus
[3] https://x.com/arena/status/2077824029126504525?s=20
[4] https://arena.ai/leaderboard/text
[5] https://x.com/kimmonismus/status/2077836497739304968?s=20
Re: (Score:2)
> I"m curious...do these Chinese models contain Chinese guard rails?
> I mean, will it censor you if you ask about the Chinese premier looking like Winnie the Pooh or tienaman (sp?) square....will it say it can't answer that...or give answers that only praise china?
> Genuine question with potential genuine concerns....
They remove that kind of data from the training and so model has no idea what you're talking about. They don't add guardrails after the fact.
Anyways, its open weights/open source. You can fine tune it to your liking. Lots of companies do that. Even the company Cursor that SpaceX just bought for $40 billion used a Chinese model, fine tuned it and ran the company on it.
Re: (Score:1)
> I"m curious...do these Chinese models contain Chinese guard rails?
> I mean, will it censor you if you ask about the Chinese premier looking like Winnie the Pooh or tienaman (sp?) square....will it say it can't answer that...or give answers that only praise china?
> Genuine question with potential genuine concerns....
This is pretty funny to me. I do not care about Chinese censorship in my day to day, so censuring Winnie the Pooh or Tiananmen does not matter at all. On the other hand, Grok replacing all correct answers with whatever Elon Musk's opinion is does cause trouble. I can only imagine it is the opposite for anyone living in China as they are unaffected by the opinon-rails on Grok but not whatever is politically sensitive over there. Perhaps the US ends up being the premier customer for Chinese AI and vice versa
Re: (Score:2)
> I"m curious...do these Chinese models contain Chinese guard rails?
> I mean, will it censor you if you ask about the Chinese premier looking like Winnie the Pooh or tienaman (sp?) square....will it say it can't answer that...or give answers that only praise china?
> Genuine question with potential genuine concerns....
I don't think it matters all that much. Uncensored version of previous Kimi models are widely available. Given absurd model size you might have to wait some time for an uncensored K3.
Oh No! (Score:1)
Now China will have to deal with CSAM created 40% cheaper and on more machines!
Re: (Score:2)
They don't really care.
Re: (Score:3)
These aren't image or multi-mode models are they? There are already tons of image models out there (SDXL, Illustrious, Pony, Flux, WAN-video, etc.) and most can easily load on local hardware. Image models are a lot easier to train than coding LLMs, and I don't see a lot of these big companies pouring that much more into them when they can just pull and merge checkpoints from Civitai by people who've done the heavy lifting. After all, OpenAI pulled their Sora product.
Re: (Score:2)
Kimi can't make CSAM. And if it could it would be more like 70% cheaper.
Re: (Score:2)
Indeed that is what seems likely to happen next. But you forgot to mention that they will also urge the US president to pressure "allies" into not using Chinese models as well, to the extent that all-out economic wars are declared on anyone daring to use the better, cheaper product from China.
Component Price Bubble? (Score:2)
When do component prices start dropping. Computer equipment prices are completely fucking ridiculous right now.
Re: (Score:2)
The big issue is GDDR6 memory which is pretty new and there's not a ton of capacity already, and it's all been bought out through the end of 2028. Plants take 3 years to come online and are expensive (Read: risky) to build if the demand falls off
I would love to be able to buy a 4TB GDDR7 machine and never have to use cloud services again, but it's not happening before 2035
This is a good thing (Score:3)
We should abandon the old cold war thinking and move toward cooperation with China.
Re: This is a good thing (Score:1)
Yeah, basic human rights are highly overrated, America should just stop worrying about them.
Re: (Score:2)
> We should abandon the old cold war thinking and move toward cooperation with China.
>> Yeah, basic human rights are highly overrated, America should just stop worrying about them.
Can we cooperate and worry?
Re: This is a good thing (Score:2)
We already have, from a leadership perspective anyway
Re: (Score:2)
Say that again while speaking into a mirror.
Re: (Score:2)
See "Plan A". That will never happen because it assumes the US govt. will make only wise, honest, and insightful decisions, but just try to come up with a better "Plan B" that doesn't start WWIII.
Re: This is a good thing (Score:1)
We should have worked with that nice Mr Hitler and taken advantage of his sources of cheap labour and gold.
Re: This is a good thing (Score:1)
Instead we eventually hired someone who has only read Hitler's book to run our country
Re: (Score:3)
No, usually invoking Godwin's law is a loser because it's so far detached from the topic under discussion. I think in the case of discussing the authoritarian, wannabe dictatorship that has gained power in the USA it is entirely appropriate. That's how I see it as a Brit observer.
So I'm all for ditching cold war nonsense (Score:2)
But I do wish we would keep China more at arm's length because they are a fascist dictatorship. The same goes for Saudi Arabia and basically all of the Middle East.
I mean we're not going to do that but it would be nice if we would call them out on their bullshit and try to move them towards democracy. On the other hand is an American I'm watching my country turn into a fascist hellscape so there we go...
This is getting to be ridiculous (Score:3)
This model has 2.8T parameters /w 50B active. They should focus on repeating layers or something to get the size down. Memory prices are too damn high.
Re: (Score:1)
For those of us who aren't fully up-to-date on local AI prerequisites, what kind of ram and computer requirements would running a model of this type require?
Re: (Score:3)
It depends what you want to do. Minimum requirement is something able to multiply numbers and with (probably) about 3 terabytes of storage space.
You're probably actually going to want something that can at least hold those 50 billion parameters in some sort of reasonably fast memory. That's probably about 50 GB.
Or Slashdot will tell you you need 10 terabytes of the fastest VRAM and a few dozen of the very latest processors from Nvidia.
Give it a bit though. Whenever an open weight model is released there's a
Re: (Score:2)
50B active does not mean you choose 50B when starting an answer, but it are different 50B for each token. That means the computation is fast(er), but you still need to full RAM. In theory you could read weights for each token from SSD, but it would be incredibly slow.
Re: This is getting to be ridiculous (Score:2)
I read 900 GIGAbytes VRAM.
It will come down, but whether it will come down far enough, and still be usable on lowly home systems, will take some time to emerge.
Re: (Score:2)
Quantized to 2 bits you might just about be able to run it with 512GB of RAM and 24GB of VRAM. But it would lose a lot of its intelligence.
However, there are people who've been running big models on consumer hardware by streaming data from an SSD rather than keeping it all in RAM. Significantly slower but doable.
Re: (Score:2)
> Quantized to 2 bits you might just about be able to run it with 512GB of RAM and 24GB of VRAM. But it would lose a lot of its intelligence.
For reference Kimi 2.7 weights in at 1T tokens and 2 bit quants for it are in the 250GB to 350GB range. With 2.8T tokens I don't see it fitting.
> However, there are people who've been running big models on consumer hardware by streaming data from an SSD rather than keeping it all in RAM. Significantly slower but doable.
This seems crazy to me. I guess it depends on how routing works inside of the model. From memory this one has almost 900 experts and 16 of them are selected on each pass. Maybe you can cache the most used layers and hope for the best but realistically I suspect you are still going to have to transfer tens of gigabytes per token and performance will be counted in
Re: (Score:2)
> For reference Kimi 2.7 weights in at 1T tokens and 2 bit quants for it are in the 250GB to 350GB range. With 2.8T tokens I don't see it fitting.
Sorry I meant to say parameters not tokens.
Re: (Score:2)
> For those of us who aren't fully up-to-date on local AI prerequisites, what kind of ram and computer requirements would running a model of this type require?
Would need a TB of RAM to run it effectively. Might be able to get away with a stupidly bit starved quant using as little as a half TB however would expect quality drop to be such you would be better off running a smaller model.
CPU inference isn't a huge deal /w only 50B active parameters unless expecting instant responses. For Intel hardware I think most 4th gen or later Xeons /w AMX are able to max out memory bandwidth. Realistically if you have 1TB of ram you are going to have at least an 8 channel sys
Don't worry (Score:3)
America will retake the lead in AI the same way we did with EVs: Just ban the ones from China. Problem solved.
Leapfrog (Score:5, Interesting)
"The entire game has changed."
No. It is still a game of leap-frog. Every time a company releases a new AI it is the greatest thing evar!1!! ... and then somebody else one-ups them with an even more advanced one.
Same as it ever was.
Frontend? (Score:2)
So they developed a super web designer? Gross
Well it was inevitable (Score:1, Insightful)
They are willing to build data centers at scale with no regard for the environment or citizen's opinions on the matter, subsidize the hardware, reverse engineer the models, and distill the training done in the US. The only thing holding the US back is itself. We honestly shouldn't compete, because the playing field is far from level. Let them spend the money and destroy their resources, and we'll distill THEIR models.
Re: (Score:3)
Oh, until I read the last word I thought you were talking about the US.
Who has the most energy and engineers? (Score:2)
(Did I smell a brain fart?)
Two significant advantages on the China side, though I think the title of the story is clickbait and it's much too soon to see what is going on with AI.
The major advantage is the energy. But what do you do when the wind isn't blowing or the sun isn't shining? You just wait to train your AI because training can be done at your convenience. But I'm also sure that the Chinese are working on prioritization systems for the time-sensitive queries so that when Xi has a question he ALWAYS
Re: (Score:2)
"But what do you do when the wind isn't blowing or the sun isn't shining?"
You run on the battery power you saved during when the sun was shining and the wind was blowing. And there are other ways of storing energy.
Re: (Score:2)
I acknowledge that batteries exist and are significant, but the efficiency is significantly lower. I don't think it makes that much sense to do low priority stuff off of batteries and I regard most of the AI training as relatively low priority. However I have to acknowledge the other side of the inefficiency of building extra computing capacity that spends a lot of time waiting for the power...
Hmm... I just thought of a retired CPA who worked with this kind of stuff. Maybe he could work up the numbers? Or s
Re:Well it was inevitable (Score:4, Interesting)
> They are willing to build data centers at scale with no regard for the environment or citizen's opinions on the matter, subsidize the hardware, reverse engineer the models, and distill the training done in the US. The only thing holding the US back is itself. We honestly shouldn't compete, because the playing field is far from level. Let them spend the money and destroy their resources, and we'll distill THEIR models.
LOL. We bet the farm on this.
Also, we don't have to distill their models. Their models are open-source, open weights.
They did jump ahead a few times and we did use their techniques and learnings to jump past them in the next round.
Re: (Score:2)
There's no guarantee the the models will continue to be open weight.
Also, if it's actually done by distilling, it will be quite fragile outside the predetermined area. (But bet on harnesses to solve that problem.)
Re:Well it was inevitable (Score:4, Insightful)
The point is that there's no technology gap. If you want to make back trillions of dollars of investment on your data centers, you can't have a competitor that's almost as good, 40% cheaper and open source.
There are only 6 billion internet-connected people in the world. They invested $2.5 trillion. Just to make the investment back, they need everyone to pay roughly $400 each. By the way, in many parts of the world, $400 is twice their annual income.
Re: (Score:2)
> By the way, in many parts of the world, $400 is twice their annual income.
And in other parts of the world, that's less than people spend on coffee every month.
Do you really expect that the cost must be distributed evenly across all people worldwide?
Re: (Score:2)
Errrm....so if only half the people must pay, then they will owe $800/person. If only a quarter.....you do see where this is headed, yes?
Re: Well it was inevitable (Score:2)
I think the inevitable thing is the main players getting crushed in the middle. At the high end it is as you said, but at the low end people can run some pretty impressive local models. Why pay monthly for something you can setup locally for a few $k in hardware. Anthropic and such are in the middle, not price competitive at the high end or low end.
Re: (Score:3)
Some months ago we were reading articles here on Slashdot about how we had already reached peak AI by current methods, based on many industry experts and an event where a next-level bump in training data and resources was used and produced a model that was only barely better than the prior generation.
I guess the industry experts were just wrong? Or did we change training methods?
In any case, current evidence suggests that we have yet to hit peak, and so it would be reasonable to assume that some other coun
Re: (Score:2)
1) We keep changing the training methods.
2) We keep improving validation methods.
3) The evidence we've reached peak AI is, at the very best, shaky.
4) The best models at any time probably aren't public either in accessibility or even in knowledge that they exist.
5) John Henry may have driven deeper than the steam engine, but the steam engine drove deep enough. And the strain killed John Henry. You don't see many people driving steel pegs anymore.
etc.
Re: (Score:1, Troll)
I agree that you shouldn't compete. You should worry about getting rid of Trump and his family instead. Priorities people.
Re: (Score:3)
Interesting to hear China blamed for the environment when they lead in solar and the US president beefs with wind.
Re: (Score:2)
"the US president beefs with wind." The alleged U.S. president is nothing but wind.
His latest bright idea is to tariff Canada for the smoke from their wild fires while conveniently not mentioning the ones in Wisconsin and Minnesota.
I hope someone suggests to him he have a wall built to keep out the "Canadian" smoke. That would be sheer enjoyment for me.