>Deepseek CEO Liang Wenfeng: Chinese AI models are falling far behind the US>"Open-source" clearly means top Chinese AI models while "closed-source" clearly means top American AI models.https://arxiv.org/html/2512.02556v1Fuck, they need EUV ASAP...
>>108074434>We have to win the AI race or that other country will...>They'll try to...>We'll have to....>WELL I DON'T KNOW WHAT HAPPENS BUT WE JUST HAVE TO WIN ALRIGHT!?
>>108074442Interesting quotes from a recent interview with Liang recently (translated from Chinese). If true we really are in the worst timeline:>The R-squared of progress the frontier AI labs are making and EUV-enabled high end chips is basically 1.>LLM research is trending towards undisclosed techniques ('dark arts'), model tweaks and deployment tricks where a small group of researchers make cutting edge advancements
>it's sigmoi-ACKACCELERATEalso this sounds mostly like an appeal from the deepseek ceo to ccp for access to compute / a nationalised efforti do think they will steal weights at some point (good), but it's too early
>>108074556So is gpt 5.2 actually better than opus 4.5?
>>108074525>LLM research is trending towards undisclosed techniques ('dark arts'), model tweaks and deployment tricks where a small group of researchers make cutting edge advancementsThis isn't really surprising considering it's the same transformer model that's made in the 80s, maybe back propagation is the last innovation.Some of this hype money will trickle down into research for a new AI model or just neuromorphic computing. Although for results to come to production is probably another 10 years.
>>108074598nevermind they released opus 4.6 yesterday which is better and costs the same as 4.5
>>108074525OpenAI and Anthropic want to virtue signal signal to the max. They can't even talk to each other about the fact they know every chink employee is a spy. Also even the non chinks are complete hypocrites (they know they are pirates and fair use is a joke). The moment they go to a different company they leak everything too.There is no moat, since there are no secrets for very long.
>>108074672More interesting questions are raised if we assume there in fact ARE moats.
>>1080746475.3 codex came out yesterday and initial impressions suggest it's a better coding model than 4.6full 5.3 models probably soonboth the new releases seem to be significant upgrades too
>>108074672There's an immigration free for all and national identity in the west is basically meaningless, ofc everything is filled to the brim with spies.
>>108074672>The moment they go to a different company they leak everything too.this is by design in SV. non-competes are not enforceable. as long as they aren't smuggling out actual material, everything in their brains is for sale.
>>108074774They should still not communicate trade secrets, should.
>>108074801they absolutely should.it's bad enough they've stopped publishing research.if the tech is going to turn out to be what they say it is, or even if they get half way there, then that information cannot be allowed to exist only within a single company/statethere is a need for some control, because you don't want any retard building plague 2.0 in their garage, but by and large, we benefit from having lots of players with roughly equal capabilitiesMAD
>>108074774https://www.globalarbitrationnews.com/2021/03/11/california-appellate-court-overturns-arbitration-award-that-violated-plaintiffs-statutory-right-to-workThey can fuck with the leaker but not before he leaks everything
>>108074647>Kimi K2.5 scores 1301>GLM 4.7 scores 1192Both open weight models that cost 10x less than Opus to run. Locally it costs even less.>Gemini 2.5 Pro scores 939>GLM 4.7-Flash scores 888I can run the latter at home on 10 year old hardware at 3 tokens/sec. It's slow, but so is 10 year old hardware.Imagine what a few more years of this will bring to open weight models. They are gonna become more capable and cheaper to run.
>>108074556>progress is when line exponential>don't ask questions about what the line actually measures just l-look at the line IT'S EXPONENTIAL
>>108074556Go ask an AI to figure out a custom tune for your car engine and simply chart out all the changes you need to manually make, dont ask it for an ecu file or copy-pastable tables into an ecu editor.Just ask it, with all the information about your vehicle (including seasonal weather, fuel octane, fuel quality, ethanol content of your area)Then plug those numbers into the ecu.I'll tell you what's going to happen.The car is going to run worse every time, up to and including undrivability failures from a broken VE map reference.You can coach an AI on physics, base principles, known examples of manufacturer tune settings differences for the same engine.It's still going to fail.Not 50% of the time,Every single time.It's a task that a semi-practiced novice can accomplish in about 20 minutes, from scratch and only with knowledge by example.Not 100 minutes not 150 not 200.It's a 20 minute task all of these AIs will fail every single time.These machines cannot understand base principles, well documented physics, and cannot apply them functionally.If AIs can't understand reality how is anything they spit out trustworthy.
>>108075226no one's reading all that, bro
>>108075246AIs can't figure out custom engine tuning based on vehicle details. They'll just make it worse every time. Humans can do it in 20 mins, but AIs will fail every single time. AIs don't understand basic physics and principles, so nothing they say is trustworthy.
it's cap they just want reasons to buy more nvidia
>>108075246>i cant spend 45 seconds readingyou're the people AI is going to replace
>>108075253so what you're saying is that cars interest you?>>108075278mad
>>108074434I don't get it. Are western models really much better? Or is it just that western companies are finding actual applications for them, while the chinese seem to have stalled at competing with one each other for the models that get the most points in tests?
>>108075535This is from DS CEO himself. Yes they are a lot better especially in terms of depth (and the ability to be integrated into the workflows of top devs and scientists) and he's worried the gap is accelerating not shrinking. He's frustrated that people in China don't understand the real reasons for this gap.
>>108075337You're aware that AI has to be manually proof read, yeah? Each word every time.
>>108075689i print all my ai outputs out and proofread them at the cafe
>>108075565>Yes they are a lot better especially in terms of depthWell, how come? Is it a matter of having more memory?Also, deepseek is what motivated other players in China to try to push the limits, isn't it? They should once again lead by example, I guess
Deepseek models are not even open source. They contain clauses which restrict you from using the model for non-lawful purposes. Complying with the law should not be a matter for licenses. It should be a matter for democratically elected governments, under free and fair elections of universal suffrage.
>>108075246>>108075226@grok summarize this guy
>>108075823Better design, better training, more inference, etc. Chinese models are benchmarkmaxxed to hell like grok. There's a noticeable difference in usability when you actually try them
>>108074434I hope deepseek realizes theirs is very good because they didn't fall for anthropic and openais muh knowledge scam. they fake every test result and the exams are rigged. all their shit sucks. deepseek at least cuts you some of the good bits of home grown hacked onto openai breadth slop.the real problem is these companies not deleting reddit from the weights
>>108074525>LLM research is trending towards undisclosed techniques ('dark arts'), model tweaks and deployment tricks where a small group of researchers make cutting edge advancements
>>108074840Problem is that both GLM and Kimi were trained on OpenAI/Anthropic
>>108077630The technological dark age is upon us lmao.
>>108077630>>108078046Praise the Omnissiah!
deepsneed models still makes amerimutt token price seething & dial8'ng so that's good enough for me
>>108074434So basically, they would need to help Huawei or domestic chipmakers on improving equipment.Isn't this exactly what the US intended by withholding top cards from China? Basically ensure China stays out of first place.This is why Chinese AI companies pining for NVIDIA is a bad idea. The more dependency, the damaging the fall.I keep saying it: but Chinese companies need to make a nation-wide Manhattan effort to separate themselves Western chipmakers. The West isn't interested in working with China; they are trying to contain it. Better hope you can come up with something interesting...That being said, being open source is itself a plus. I have no doubt that OpenAI, Google, Anthropic and other popular AI entities have no plan on open-sourcing their models. Like, at all. Governments and the companies will argue that deepfake porn or misinformation warrants restriction. To me, that argument effectively makes closed-source AI useless in a sense. Or may not "open source", but offline usable. If
>>108075226I work in the semiconductor industry and this is why I'm not worried about AI taking my jobthere are so many real world applications that have enough data for trained humans to make good designs but not nearly enough for an LLM to fit to and "understand" the problemeven with writing and programming which have vast datasets of examples available AI struggles to match human perforance without handholdingthere's no way it hell it can learn to tune a car much less design a dry etching process with some shitty skematics and TEM images to go off of
I assumed that when you make a request to claude/openai/gemini these days there's more than just the model inference going on, like a backend of agents
>>108080981China put a block on the importation of Nvidia GPUs
>>108074945lrn 2 read
>>108082359>progress is when AI has a 50% chance to not hallucinate shitkys
