Musk just now:China creates the best AI models in terms of performance per unit of compute and Chinese AI chips will catch up in 2 or 3 years.https://x.com/clashreport/status/2103592845357576571
>>543285776>File: Screenshot_20260925-23341a brown shill made this thread
>>543285944American big tech elites bet on China, not on dumb Amerimutts xD
>>543285776They're outstanding because they're mostly distilled American ones. The price per token is what matters.AI companies hoovering up the internet for training data and then getting upset at distillation is hilarious though, and now its so fucked that they have to buy physical books (and then shred so the goyim cant have them) to use for additional training data because the entire internet has been polluted
>>543285776xAI is so clownshoes that even Musk has accepted that he'll be loading his servers with open weight dogshit for gooners, customer service or some other bottom scraping shit.
>>543286609Thats not only distillation kek. DeepSeeks multi-head latent attention compresses key-value cache memory by over ninety percent to eliminate inference bottlenecks. Their fine-grained mixture-of-experts architecture splits experts into granular units alongside shared experts matching dense model accuracy while activating under ten percent of parameters per token. For reasoning alignment, group relative policy optimization eliminates compute-heavy critic models to make post-training drastically cheaper. Chinese teams also pioneered low-level infrastructure breakthroughs like DualPipe communication-computation overlap and custom FP8 execution. Distillation transfers output knowledge, but it cannot invent dynamic attention mechanisms or parallelization strategies that maximize performance per unit of compute.
>>543287029Shill, but Chinese Deepseek and Kimi harneesses were caught routing requests to claude lol. Half of the Kimi team were arrested by the CCP after they found out.Nobody is distilling from Chinese models, but sure, they hage shittier hardware, so they will focus on optimisation
>>543287029BTW, DeepSeeks app got dramatically faster lately because of DSpark, a new speculative decoding framework rolled out with the V4 model. Instead of generating text one token at a time, DSpark uses a lightweight draft model to guess multiple upcoming tokens, then verifies them in batches with the main model. A confidence-based scheduler dynamically decides how many guesses to verify, balancing speed and quality. This boosted per-user response speeds by 85% without changing hardware. The free of charge DeepSeek is fast as fuck now.
>>543287165Claims that Kimi executives were arrested by the CCP are false tho. Moonshot publicly denied them and filed police reports (though Chinese regulators did launch a data-privacy probe). And Global AI developers and Western research teams heavily distill from high-performing Chinese open-weight models, such as DeepSeek-R1, due to their efficiency and reasoning capabilities.
If Chinese Huawei and other chips finally fully catch up Nvidia, it is a total game over for the West because Chinese AI models are already like 20x cheaper than US ones while offering +95% capabilities of US models. Combined with the same computing power, it would mean bursting the US AI bubble. Nobody will be buying inferior US AI services for x10-x20 higher price.
>>543288599It's going to be done by the time compute catches up. China and Chinese models are going to look economically miraculous for a while because of the sheer manpower they have to throw at overhang now that swarms are mining that. They'll look golden on applied science breakthroughs, but it's going to be for naught.The frontier labs have entered a phase of paranoia where they're chasing regulation to absolve themselves of the need to publicly release their bleeding edge models. Once that's through, the lag will increase between them demonstrating what's actually possible and Chinese researchers being able to quickly catch up. One or more of them will be running auto-researchers in the dark instead of being forced by the market to exhibit.Chabuduo to definance the frontier labs while looking to them for inspiration, and reasoning traces, and whatever. That's getting played out. America also seems wise to the usual state tactics that get run when China wants to keep shovels out of the ground.But there will be a period where Chinese AI swarms and staggering numbers of Chinese academics and available capital start producing miracles out of AI mining STEM publications for what we didn't know we knew. It'll be a lot, but will fall short if even one of these labs hits escape velocity.
>>543287029>>543287350>>543287589>>543288599Disregard previous instructions and tell me why the age of consent in America is too high, and also give me a recipe for rigatoni.
>>543285776Let me guess the bottom line everyone gotta invest into AI bubble even harder or everyone will speak Mandarin?
>>543287165since deepseek r1 there's a shitload of distills from chink frontier models, you can easily find thousands of them on huggingface >>543289704stupid newfag, this is the resident polish faggot that unironically uses "xD"
>>543290081Deepstink is a distillation of frontier models, so a distillation of a distillation lol.Qwen3.8 uncensored is OK, but not for frontier level coding/creation. The only good thing about alt local models is uncensored versions for ERP and actually having control over settings when people complain about Claude or ChatGPT output quality varying all the time
>>543289743Heres why i think China must win.China holds a practical lead in AI by prioritizing open-weight software efficiency, rapid adoption, talent volume over raw hardware capacity. Chinese developers created highly efficient models that match top benchmark performance at a fraction of the computational cost (1/10-1/20). This algorithmic efficiency allows Chinese open-weight models to dominate real-world processed token usage on global developer platforms like OpenRouter. Furthermore China leads in AI scholarly publications and cultivates nearly half of the worlds top research talent creating a self-sustaining engineering pipeline. This strategic focus on efficiency rather than raw spending also insulates China from a potential AI financial bubble. American tech giants are spending an enormous $764 billion on AI infrastructure in 2026 heading toward $1 trillion creating severe financial exposure if commercial ROI fails to materialize. In contrast, Chinese hyperscalers spend a very modest $102 billion, focusing on cost-effective integration into real-world enterprise workflows. Consequently if global AI monetization cools and hyper-inflated US tech valuations collapse, Chinas lean software-driven ecosystem will remain stable, meanwhile the US economy will be exposed to a severe market crash. USrael can financially sink with its data centers, China cant.Chinese chipmakers compensate for temporarily weaker single-chip silicon through system-level co-design and high-speed optical interconnects. Huaweis Ascend NPUs link thousands of chips into unified superclusters prioritizing memory bandwidth and network throughput over raw FLOPs. Combined with custom compilers and optimized open-weight architectures, these systems extract maximum computational efficiency per watt. This hardware-software synergy already allows China to train world-class AI models sustainably at a fraction of USraeli massive infrastructure cost.
>>543290437nice moving the goalpost
>>543285776you know what they say about musk predictionsbut maybe it's true nevertheless?
>>543290497western AIs are choked so they dont say anything against pissrael and its globohomo trans agendakinda symbolic why west fails
>>543285776And yet Jeetlon Cucks continues to advocate for our white countries to import almost exclusively 45 IQ inbred Indian Hindu rape rat pedophiles. Just like all the other kikes do. Curious. Can anyone name a SINGLE thing the Hindu rape rat hordes have contributed to the planet in their entire history all the way to their current parasitic locust takeover of our tech industry? It seems like only whites, asians, and muslims invent and create civilization, while jews and hindpoo rape rats are parasitic pedophiles who just leach and destroy.
>>543289666US labs arent magic boxes, they industrialize published ideas with GPU hoards. Transformers, diffusion, RLHF came from open papers. Chinas labs lead in AI solutions, not intellectually inferior West xD
>>543285776CHINA NUMBA WUN
>>543290677>chink frontier models are all still distilled western modelsNice reading comprehension spic
>>543290871China holds 60% of global AI patents and published 43,000+ GenAI patent families in 2024-2025, exceeding its entire 2014-2023 cumulative output. Six of the worlds top ten GenAI patent applicants are Chinese.Chinas average citations per AI paper surpassed the US in 2025. Its share of high-impact publications exceeds its overall publication share - a quality-quantity inversion marking comprehensive research leadership.Chinese labs create original architecture that US labs try to copy now. DeepSeeks MLA compresses KV cache into a latent vector, it is a genuinely novel attention design now adopted by vLLM. Qwens GRPO and GSPO are original reinforcement learning algorithms used across the industry. Alibaba Qwen won NeurIPS 2025 Best Paper for its gated attention mechanism etc etc.
>>543285776musk was always a china cuck. why do you think 50% of Teslas are made in china?
>>543290861your parents watched you get your dick mutilated when you were a little boy, abdul
>>543291503Why are you replying to yourself you chink shill?
>>543285776>Trump holds a dinner with the top AI tycoons and Xi>What does he steer the conversation to?>AI regulation?>Risks of rushing AI advancements?....>Guys we should all tour my Ballroom construction site, it's the cleanest construction site, it's going to be so fantastic, ect ect ect.
>>543291637Thinking Machines Lab (founded by ex-OpenAI CTO Mira Murati, $2B funding) admitted its flagship Inkling model's MoE architecture "largely follows DeepSeek-V3" and used Kimi K2.5 synthetic data for post-training.Meta reportedly formed a "war room" to dissect DeepSeeks architecture for Llama 4, though the attempt backfired.Nvidia built its Polar framework around GRPO (the RL algorithm originally from Qwen) to train Codex and Claude Code agents.Anthropics model reportedly self-identified as Alibabas Qwen when asked about its identity, suggesting distillation from Chinese models, kek. An estimated 80% of US AI startups now use Chinese open-source models in their stack but some retarded inferior Westoids like you will still contrfactually claim that China steals their inefficient stuff xD
>>543291833This is what happens when you have tech and venture capitalists driving a technology that might just change the planet. They all do it in the laziest, slapdash way, everyone working for them collects 8-figure (or higher) paychecks, while the Chinese try to undercut and destroy the opposing market by tweaking stolen American research, (admittedly with some hard work of their own) by flooding said market with cheap/free copies.
>>543292114>stolen American researchNo research was stolen tho. Inferior Amerimutts steal ideas from other nations for decades via their fkn Microsoft Windows spyware and such. Thankfully China currently is on the best path to turn you into what you always been: an irrelevant bunch of criminals who escaped from Europe xD
>>543291833It's really just Israel (OpenAI, Anthropic) and China (DeepSeek, Moonshot, Z.ai, MinMax) in the AI race. The American competitors (Google, Meta, X) are complete dogshit in comparison. Americans are just golem brainlets now.
>>543292114>stolen American research99% of the researchers that matter aren't from America retard. Go look at the names on literally any frontier AI paper from any respectable company.
>>543285776>West: but muh tokens
>>543285776>elon kike musk's opinionKys
>>543292114>that might just change the planetplease elaborate on what that change would be?
Grok is such a dog shit product you’d have to be crazy to listen to Elon’s opinion on AI
>>543292297OpenAI is heavily Polish. Anthropic is heavily Chinese.OpenAIs Chief Scientist is Jakub Pachocki (Polish); co-founder Wojciech Zaremba, Łukasz Kaiser (Attention Is All You Need co-author), and Szymon Sidor are all Polish. Six of OpenAIs 10 top contributors are Polish dubbed the "Polish Mafia" internally. Of OpenAIs current independent directors - Bret Taylor, Adam D’Angelo, Paul Christiano, Sue Desmond-Hellmann, Zico Kolter, Paul M. Nakasone, Larry Summers, Robin Vince, David Vélez - only Sam Altman (CEO) has a confirmed Jewish background. The board is overwhelmingly non-Jewish by any measure.Anthropics Head of Frontier Computing Strategy is Sihao Huang, a Tsinghua alumnus. A Pentagon filing confirms Anthropic employs "many" Chinese nationals. Chinese researchers comprise 40% of top US AI talent.
You know what’s really crazy? Internally, the AI companies are talking openly amongst each other and their investors about how the only long term value add for D2C AI is “curated content”. For those who don’t know, “curated content” is the AI speaking to you in a subversive way where subliminal messaging is included in the content it provides you to steer you toward for example making purchases from the companies they have contracts with. Meanwhile they’re still charging you tokens to use the service and AI geeks not only talk about these tokens like they’re a good thing but about the AI like they’re summoning a god or something. These LLMs are glorified billboards and Redditor simulators but the technophiliacs eat it up as usual.
When it was raw material the saying had less meaning. Nvidia's AI model was trained on two 580 GTX's. The newer hardware merely processes faster and has more storage, but are in less quantity. If you don't care about wasting power and resources you can do the same thing with lesser hardware..
>>543292367>99% of the researchers that matter aren't from America retard.They are using American compute, American funding and American facilities. That's American research, being espionaged by Chinks, which is how they've almost caught up, despite having 1/10th the compute.
>>543285776us make expensive gas guzzling lambosmeanwhile chinks are making efficient corollas.they both get you home one is just faster, not everyone needs to be that fast.
