How did the US not get first movers advantage? Apparently the new chinese models are like 99% there but cheaper token costs.
The Chinese video gen models are groundbreaking. OpenAI couldn't compete with them so they just gave up and discontinued Sora entirely.
>>109549366thanks for reminding me so I can laugh about that again
>>109549101Because China has less GPUs but 3x the amount of AI researchers, and third of US AI researchers come from China anyway so they can easily just steal shit. The chip advantage is the only advantage US has, once China has more chips it will be 102% over for US with 2% margin of error.
>>109549101>cheaper token costs??https://artificialanalysis.ai/articles/gemini-3-7-time-frontier
>>109549840??kimi k3 is 2.8 trillion parameters.grok 4.6 is 1.5 trillion parameters.which model do you think is cheaper to run?
>>109549101The ones that solve important math problems are the american ones.
>>109549951according to american news outlets
>>109549951>openai hires mathematicians to use their llm for math>llm finds some proof>anthropic hires mathematicians to use their llm for math>llm finds some proofguess whats needed for chinese models?
>>109549975no, according to american-chinese mathematicians:https://teorth.github.io/tao-web/slides/age-of-ai-icm-2026.pdfhttps://x.com/henryquantum/status/2083623695436623915
>>109549844>>109549919Their claim to fame is outclassed by 5.6 luna which is the free model that OpenAI offers. Don't expect coherent responses.This is the current landscape:Speed and accuracy are priority = Gemini flashCost is priority = Luna (but also Gemini as well now)Intelligence is priority = Fable, Opus, Sol, GrokLocal is priority = Chine- oh wait I can't afford a $40,000 server to run some 3 terabyte Chinese model
>>109550068I don’t get it. Do you think china lacks mathematicians or something? You think universities in china don’t get free access to chinese AI?Reminder that random people not affiliated to openai also found proofs with AI. American AI is just better.
>>109550155let's read up on anthrophic's secret ingredienthttps://www.anthropic.com/research/riemann-zeta>Jarred Sumner, an Anthropic staff member (and non-mathematician), prompted Claude to “take a real stab” at the hypothesis itself, leaving the mathematical choices from there up to the model. Initially, Claude generated and tried 650 ideas, none of which worked. Jarred prompted Claude to try again, and it spent a day and a half coordinating about 60 Claude subagents, which this time went much deeper: between them, they ran 2,400 shell commands and wrote hundreds of Python scripts.1 The subagents ran thousands of numerical checks against known zeta zeros and refereed one another’s work. Throughout this process, Jarred's input was mostly limited to sending Claude messages of encouragement (mostly variants of “keep going” or “believe in yourself”).2 This seems to have helped Claude overcome some initial skepticism that it could make meaningful progress.do you think universities in china let you spend $100k on ai with those silly prompts?
>>109550189>$100kOpenAI mogged the entire Chinese AI industry with $2,000 worth of tokens. What makes you think Anthropic used $100,000?
>>109550131>>109550267American Ai is stupid, i ask GPT how to fix second display crashes and only told me common issues that I already tried. Gemmi couldn’t understand the question and lied about stuff that doesn’t exist. Only Kimi k3 knew the answer, it was opengl fault for memory leak. Told me how to fix it, second display doesn’t crash anymore and leak is contained.American Ai failed, the future is Chinese Ai that run on gaming PCs
>>109549366Also the Chinese music gens are far better.Multimedia is harder than text, there is no way their coding and text based stuff is not overwhelmingly superior. They are keeping those secret to deceive morons like this one: >>109549951.
>>109549844There's no way the bubble labs would have charged that low without open competition. Chinese models are cartel busters.
>>109549101>US tech companies import Chinese workers because they're cheaper>Chinese visa workers send data back to their homeland>China gets a huge and continuous boost from it citizens in US tech companies>US tech companies still would rather hire cheap Chinese than train AmericansMarx was right that capitalists will sell you the rope you use to hang them.
>>109550267>OpenAI mogged the entire Chinese AI industry with $2,000 worth of tokensfakehttps://openai.com/index/ten-advances-in-mathematics/«The results were achieved by an internal version of Astra, our next major model. The total number of tokens needed to find solutions to these problems would cost roughly $2,000 at Sol API rates.»Astra ≠ Sol
>>109550682Is English your second language? Were you talking to ChatGPT and Gemini in English and Kimi K3 in Chinese? That's might explain your issue.
>>109549101openai and anthropic have to pay the extremely bloated salaries of all of their employees and managers, not to mention all of their government bribes. if you controlled for that, token costs would come down and kimi k3 would just be "that model that isn't quite as good as the other good ones"also the chinese rely heavily on distilling the american frontier models. so "first movers advantage" literally means paying more to get your good model out so other people can cheat off your work, it ends up being a disadvantage.
>>109549101Second mover advantage is a thing too.
>>109550796Marxists will sell you communism then send you to the gulag later.
>>109551094Cry about it, why not ask it the same question and make it discover the memory leak itself. I can easy bet 1 million dollars it won’t ever discover the error. It’s not smart enough, it’s Trillions of ram wasted and devalue American currency will be the end of American economy.
>>109551111>tHeY StEAliNgNone of the American models are open source. Also deepseek and Kimi don’t act like those lame models. Whatever claims you say just prove the American programmer sucks at building Ai. The American’s can’t manage resources like China. Billions wasted just for corporate to install local AI chatbox and make memes.
>>109549101It's not just the Chinese anymore. Koreans have a new model that's less than 9 month behind the frontier labs. In 3 or 4 years, it's going to be every country that can afford more than 10,000 GPUs having top models that are more or less the same.
>>109550131>Speed and accuracy are priority = Gemini flashGemini flash will just prefer to hallucinate your answer instead of looking it up a quarter of the time. It's especially bad if you ask it something niche which it doesn't know, but it isn't obvious to the model that it wouldn't know it.Flash 3.6 is still outperformed by 3.1 Pro for doing basically any generative task from my experience.
>>109550682Did you prompt ChatGPT a second time and clarify that you already checked all of the common issues? I doubt that your third prompt on the issue with K3 was the same as your very first prompt with GPT.
>>109551320How the model acts is easily changed by using a different system prompt.Also, distillation doesn't require the original model to be open-weight. That doesn't even have anything to do with distillation. The procedure is the same whether the original model is self-hosted or prompted with an API.>The American’s can’t manage resources like China.Perhaps, but they are doing pretty well in AI so far.
>>109551526we're on 3.7 flash anon. It has a higher omniscience index than 5.6 sol max which isn't that crazy by itself but the speed and cost at which it's doing it is absurd.
>>109551677Oh wow, thanks for the correction, I had no idea. Everything is so fast now..
>>109549101The chinese variants aint the best, but holy shit they're cost effective for most people's use case. If there are scientists needing AI, the US models are still top performers.
>>109551111>>109551111>have to pay the extremely bloated salariesWhy do they "have to"? Couldn't they just offer low salaries and meme resumemaxxing retards and the "I am working on building le future of le humanity!" niggers into working for them?It's not like the employees they hired right now are the brightest around. Look at the slavnigger they hired to make Claude code for an example of an extremely incompetent employee.
>>109551111>>109552489Also, what do you need "managers" for unironically?
>>109552489question your priors.try to actually interview for openai/anthropic.there are no managers and pay is just normal for the bay. that stock options are currently extremely valuable doesn't cost the company anything.or at least learn how to read a balance sheet.payroll isn't where oopenai/anthropic spend most of their money.
>>109551595Yet you can’t prove China took it, stop the nonsense propaganda.
maybe if china a higher level of information freedom they wouldnt have such super retarded and obvious astoturfing on /g/ and /pol/token cost isnt cheaper and the models are shittierwhats even the point of trying to lie about this? might work in bugland but not here where googling doesnt get you thrown in a dungeon for a decadeching chong bing bong
>>109554499It’s literally $0.03 cents for API and application. Not a subscription, cost per word and not locked down with no app. Took you Americans 7 years to add ai apps into Desktop while Kimi k3 and Deepseek had it ready at release day. American models can’t exist without government subsidies, the Chinese models are from companies who benefit from sharing.If your country doesn’t offer good Ai services then you’re losing money and time on the dumbest Ai services.
>>109556941>captain chang of chinese intel online division waited til page 9 to self bumpcool but no one uses k3 to develop. sorry bud. we won. you will copy it down the road like you do everything else.k3 is being sued because it illegally used claude to train their models. what else is new?
>>109557099Again no proof of Moonshot or Deepseek getting sued. No AI has knowledge of it either. Did you believe in some hallucination from some random AI.
>>109557145>Anthropic said in a blog post on Monday that DeepSeek, Moonshot and MiniMax — three prominent Chinese start-ups — had used about 24,000 fraudulent accounts to generate over 16 million conversations with its Claude chatbot that could be used to teach skills to their own chatbots.oh right theyre not being sued they just copied and got away with it because china is pretty adept at stealing and avoiding consequenceswhat happens when you cant steal from others? will we still be seeing any progress in any science? LMAORike confucious say... Stear from peopo and dont pay for work
>>109557099Bro get on with the timesThe normal dev flow now is stacking an expensive thinking model for the plan + cheap one for implementation, and the combination is K3 + Deepseek flash
>>109557195the company i work for and the companies ive talked to about their ai usage, and the companies that other devs i know dont do that. go eat some crickets chang.
>>109557167They say it but no court papers or where they plan to set the law for AI stealing. If it’s China homeland, Americans lost by default because the model is open source and there is no evidence of their AI. If it’s American land, Anthropic has no evidence because it’s open source and Anthropic needs to pin point the claim with their own.So ya, they don’t have a case nor do they have evidence.>>109557218Rumors are spreading about American companies hiring programmers to download Deepseek or other free open source AI. The NDA says if someone ask which AI they are installing for the company, the programmer lies to keep investors happy. This isn’t about you, it’s about the investors.
>>109557319>So ya, they don’t have a case nor do they have evidence.ant wants to get chinese models banned in the US. this can be done by executive order or congress. no court is needed.
>>109557319>The NDA says if someone ask which AI they are installing for the company, the programmer lies to keep investors happy.this is security fraud. very easy court case for investors to win.
>>109550131>Speed and accuracy are priority = GeminiLMAO
>>109557814do name the model that you think beats gemini 3.7 flash on speed and accuracy
>>109557846there is a reason nobody uses gemini for anything
>>109557861fake newshttps://arstechnica.com/ai/2026/08/google-says-gemini-has-reached-1b-users-faster-than-any-other-google-product/
>US does everything wrong>China does everything rightJust a few years ago I would have laughed at this statement and dismissed anyone saying it as a CCP shill.
>>109549844As much as I do like Gemini as a model to pretend it is K3 or Fable level is laughable.It is a competitor to models like M3 and maybe GLM5.2 and is priced accordingly. Just look at Gemini 1 and 2 pricing to see where we'd be right now if the Chinks didn't force FAGMAN to lower prices drastically you disingenuous faggot.
>>109549101American AI prioritizes marketing for maximum profit.Chinese AI prioritizes getting shit done despite limitations.American AI is also intentionally made dumber for "alignment" reasons, that reduces the gap too.
>>109557901I'm not going to read this but I assume they are counting when whatever gemini cut down retard version of it summarizes the first reddit post it can find at the top of your google resultsThere are not 1 billion people going to gemini.google.com
>>109557921>>109557921chinese big tech isn't any different than american big tech.chinese startups hate them both.https://github.com/demo-zexuan/liang-wenfeng-investor-meeting-2026-7-22/blob/15c6504be51b884a0adc5d77e4dba41f94431454/%E6%A2%81%E6%96%87%E9%94%8B%E6%8A%95%E8%B5%84%E8%80%85%E4%BA%A4%E6%B5%81%E4%BC%9A-%E6%96%87%E5%AD%97%E7%A8%BF_1_18_translate_20260723201651.pdf>We don't compete for users or pursue profit-driven goals; instead, we strive diligently to provide excellent user services. We never harbor the ambition of creating the next "super app," competing with others, or becoming the next ByteDance or Tencent—such aspirations are entirely absent from our mindset. While we could have pursued that approach, we chose not to. In my view, this reflects a fundamental principle of restraint.>Don't expect to profit from everything. Once you've acquired users, it seems like you can create the next ByteDance-like company and then dominate the market. I believe this approach is commercially viable and feasible. Last year, we invested heavily competing with ByteDance for users – that was one strategy. However, we chose a more restrained approach: we decided not to compete directly, because what lies ahead may be a "watermelon" (a massive opportunity), while what comes before might merely be "sesame seeds" (small but valuable opportunities). We shouldn't try to capture every single sesame seed.
Are we still pretending that K3 is good?
>>109557974no, your assumption is wrong. you could have used a fast and accurate llm to not look retarded. google.com has obv a lot more users than a mere billion.
>>109549366The chink models are worse than US video models, but they're released as open weight. I'm genning fetish porn right now!
>>109558094Seedance 2.5 is the gold standard of video gen.
>>109554249Where did I say that? The anon I replied to was making some inaccurate statements, and I corrected them. I didn't claim whether China's recent models are distilled or not. You shouldn't read into people's posts too far beyond what they're actually saying.
>>109550131>$40,000 server to run some 3 terabyte Chinese model$40k you wish lmaoif you legitimately want to run a 3T chink model you are looking more at $400k + an industrial sized chiller + another circuit to your house
>>109558380dgx station gb300 with 768gb ram exists for $100k. it fits on your desk.https://www.nvidia.com/en-us/products/workstations/dgx-station/https://www.atcyrus.com/stories/dgx-station-local-frontier-ai-memorythe best chinese model, glm-5.3, should run their unquantified since it's only 744B A40B.
>>109558380>if you legitimately want to run a 3T chink model you are looking more at $400k + an industrial sized chiller + another circuit to your housewrong, buy usedhttps://www.ebay.com/itm/157742745616
>>109557801Investors win no matter what, win win anon except for you.
>>109549101They knew it was coming too. Now look at them, Google hasn't put out a frontier model in half a year. Fable is overpriced and nobody is using it on a corporate level. GPT is GPT.
>>109558905That's not going to run K3 beyond Q2. Also you'll be stuck running off cpu with llama.cpp which is always going to be an unoptimized toy for homelabs and not much more.
>>109559779??a single blade gives you 256GB vram. $40k gives you 6 blades, so 6x 256GB = 1536GB. that's the size of unquantizied K3.
>>109556941>the Chinese government doesn't subsidize Chinese AI companies!Why is chink propaganda so fucking dumb?
>>109559859You need context per session. That also requires a lot of ram. This is not enough to reliably run the model.
>>109549101>just plug your business into the chinese ai dude, it's cheaper
>>109549951america's real advantage in technology is they have a handful of the best people in any given field on earth making breakthroughs, but again it's only a handful of very smart people responsible for the majority of advancementschina has swarms of scientists and engineers who come up with tons of clever techniques that americans miss, and that manifests as a different competitive advantage
>>109559859I know it was probably a shitpost anyway, but it's not going to work broski. First of all this shit is going to draw at least 15KW of power. Volta V100 is ancient at this point, no FP8/BF16, there is no way to run Nvlink between them so you havce to fall back to ethernet which is orders of magnitude slower. V100 memory bandwidth itself is like 10% of Hopper or Blackwell. No Flash Attention support.If you can get this working at all, I'd be surprised to see more than .1tk/s
>>109549101>cheaper token costsyou think they're saints? doing this out of the goodness of their hearts?that is as long as they remained the underdog trying to gain market dominance
>>109561928what's your opinion on >>109558880
>west has better gpus!>except western gpus were smuggled then later openly exported to china en mass in exchange for petty political bribes>china is committed to building their own native ai gpu hardware and are catching up rapidly>west has insufficient grid capacity to properly power AI development without building noisy expensive polluting gas microturbines on site at data centres triggering huge public backlash>oh and china actually has way more math phds to figure shit out.yeah I don't think time is on the USA's side here.
>>109562041rumor is oai and ant are already doing rsi. it's over for china.
>>109562016It's obviously much better if you can get one. They only do B2B quotes and already have 2-3 month lead times. I'd doubt you actually get this for $100k.B300 servers are on backorder for half a year or more now and prices doubled/tripled over the last year. Not sure where the capacity for these workstations is supposed to come in.And while this can run GLM-5.2, it still cannot reasonably run Kimi K3.It would be a baller DSv4 Flash setup though.
>>109549101I hope we get cheap Chinese PC parts on the market in a few years.I'd gladly buy a Chinese GPU with 128GB or more VRAM and use it to gen with superior Chinese AI models.
>>109562093yeah you will be able to get cheap chinese parts but the government will ban them for government employees and any company with a government contract and there will likely be high tariffs on them
>>109562093delusionaldeepseek internal leaks:>Huawei's hyper nodes, specifically the Huawei 950 hyper nodes, can fully replace NVIDIA's GB200 and GB300 in terms of performance and price.>The cost is certainly higher, but only marginally so.>Whether it's 50%,100%, or even 200% more expensive doesn't matter much.>For instance, a price increase of 100% would already qualify them as viable substitutes in terms of cost.>They are also interchangeable in terms of tasks; all tasks that a GB300 can perform, a Huawei hypernode can do as well, with identical latency performance.>The only cost is that four Huawei cards are equivalent to one NVIDIA card, and they lag behind by two years in performance.>The equivalence of four cards to one is understandable.>The "two-year lag" means that four Huawei Huawei 950 cards can match the performance of one GB300 card, while the "two-year lag" refers to a temporal disadvantage of two years.>The Huawei 950 supernode will be available in Q3 or Q4 this year, while the NVIDIA GB200 was released in Q3 two years ago—there's a two-year gap.>NVIDIA may have already introduced its next-generation model by Q3 this year.>Therefore, regarding the gap between us and the U.S. in chip technology, I believe there won't be any further disparity at the ecosystem level, but the gap in chip technology remains four times larger with an additional two-year delay.https://github.com/demo-zexuan/liang-wenfeng-investor-meeting-2026-7-22
>>109561645Nothing has changed in American technological development. EV cars were Chinese cars first. Solar technology has advanced in China and Europe. Mexico and Canada train travel is better than America. Medical science stopped in USA while everyone in the world is still fighting cancer.Don’t you see American anon, you’re stuck in the past with gas powered vehicles. Phones are still from 2010 with less features. Homes build with easy to burn wood. Awful road infrastructure. The only thing Americans created pass everyone else is mass surveillance technology.
