GLM-5.2 (open-weight from Z.ai/Zai_org) just jumped +6 spots/+29 Elo to #2 on Game Dev Arena, matching Claude Fable 5s performance band. It beats OpenAI while leading open labs. https://x.com/Designarena/status/2069166634976371084Chinese frontier/open models like GLM often run up to 30x cheaper (fractions of a cent vs Claudes $15-30+/M output tokens). Open weights enable local/self-hosting at near-zero marginal cost. Rapid iteration closes gaps fast in creative/coding tasks.Chinese models (DeepSeek, GLM-5.2, Qwen etc) now deliver 95% of top US performance at 10-30x lower inference cost. Training costs are millions vs billions. Huawei Ascend chips enable self-reliant scaling despite failed US sanctions, powering efficient inference at lower energy/price. US economy heavily bets on AI premium: AI stocks are now driving most S&P gains. Monetization lags hype. If "good-enough" extremely cheap/open Chinese models commoditize intelligence globally, margins collapse for US frontier labs/chips. Chinas broader manufacturing/export base is fsr more diversified/resilient vs US tech/finance concentration. Rapid adoption + efficiency could flood markets, burst valuations, trigger Capex pullback and huge recession much bigger than the dotcom bubble or 2008/2009 - classic bubble pop when tech deflates faster than productivity arrives. It is a real risk if China wins the "volume + apps" layer.China has all the technological potential to derail US economy by bursting the AI bubble, imo.
this can't be true. if GLM-5.2 is actually this good and cheaper then openai and anthropic are dead in the water.
>>537642828Oh no! That would be terrible! :O *laughs*
>>537642777Holy shit the number got up????? US is fucked!
>>537642973ok nigga i'll bite, does opencode work well with glm-5.2? some models are good but they suck at harnesses
I did try out GLM 5.2 myself and the results are great.However if you think you can run it locally, then TECHNICALLY YES, you can, but probably no since it requires 180 GB RAM or so, which means you'd have to fork out $30-40'000 USD for whatever that super AI GPU is called.
US hyperscalers (AWS/MSFT/GOOG) ramp Capex from $156B (2022) to $1,018B by 2027E - massive casino-tier bet on sustained high-margin AI monopoly. Chinese (Alibaba/Tencent/ByteDance/Baidu) scale modestly to just $123B, yet deliver near-frontier models at 30x lower cost via efficiency + Huawei chips. This asymmetry screams overinvestment: US valuations and economy priced for premium pricing/power, but cheap Chinese alternatives can commoditize AI globally, slashing ROI on trillions in Capex. Monetization lag + margin collapse = classic bubble pop, hitting US tech/finance concentration far harder than diversified China. xD
>>537642777This is were the ai bubbleslop is going to burst eventually, open weight models converging on frontier ones will force the ai providers to sell on razor sharp margins while the ones profiting will be the creators.Probably will take a year+ for market irrationality to catch up to the fundamentals.Say buhbye to your 401k wagie
>>537642777Don't forget that NIGGER Trump banned Claude Fable.
>>537643501>This asymmetry screams overinvestmentok thanks deepseek
Looks like they dropped DeepChink already, it's dead in the water>>537643501>>537643598Most of the money going into AI is going to developing data centers. Making more efficient LLM would only mean the US would be able to secure more foothold in the sector.
>>537643678>>537643681In May 2026 speech in Texas, BlackRock CEO Larry Fink said the US needs $10 trillion in AI data centers and power infrastructure over the next decade. Funding would come from Americans’ savings, investment, and pension accounts via private capital to grow with AI xDFull event video: https://www.youtube.com/watch?v=AKNDcdmGdLc (quote 31 min mark).
>>537643079No need to run it yourself, someone will provide it as a service w dirt cheap pricing.Openrouter 1m tokens @ 1$/3$ vs opus4.8 5/25.Deepswe puts the model near opus4.8 medium
>>537643955Sure you don't need to, now, but ideally you'd want a guarantee that it can't be taken away from you.
>>537643681>>Looks like they dropped DeepChink already, it's dead in the waterDeepSeek released V4 (Pro/Flash) in April 2026 topping coding benchmarks (SWE-Bench), strong reasoning/agentic gains and ultra-cheap pricing ($0.14/M tokens). Its actively leading open-source downloads and challenging frontier models. Recent updates (V3.2/R1) keep it competitive. Far from veing dropped. xD>Most of the money going into AI is going to developing data centers. Making more efficient LLM would only mean the US would be able to secure more foothold in the sector.Efficiency destroys the data-center monopoly bet. US hyperscalers pour $10T (as Fink announced xD) into data centers/power for high-margin premium AI. But Chinese models deliver 95% performance at 30x lower inference cost + open weights for local/self-hosting. Efficiency wins volume globally, commoditizes intelligence, slashes ROI on US Capex binge, and shifts power to apps/deployment, not just raw flops. China scales smarter, not just bigger. If the AI bubble bursts Amerimuttland and its finance/AI-bubble/service "economy" would drown in shit meanwhile China would barely feel the pain xD
You forgot your xDDeepseek gives concise replies and let's you discuss edgy stuff gpt gets upset about.
>>537642828>>537642777because AI is retarded and can be essentially charged for arbitrary prices. It's also much cheaper to build off pre trained models
>>537644469>and let's you discuss edgy stuff gpt gets upset about.Fun fact: The GPT is for 'Gets Panties in a Twist'
>>537644032Sure, but nobody can take open weight models away from you, unless you mean they take away your internet access.Its one of the rare instances were saas makes sense
>>537644545Makes sense. It is also very glib with its replies, which I don't like. I know they're only chat bots but at least deepseek just gives a concise reply and doesn't waste time kissing your arse first.
As for me, i switched 100% over to Chinese models. Until recently, i was using Grok because i still had a really old Twitter account and could use it without having to go through that new registration process on X. But using Grok for free - no subscription, no payment info - has become unbearable. I send a few prompts to Grok, and it immediately asks me for a paid subscription. Thats why i switched to DeepSeek. Its 100% private, a perfect clone of Grok (even better) and completely free.
>>537645002But the Chinese are doing this on old chips, right? That's quite impressive if so.
China is rapidly closing the AI performance gap with the US. Charts from ArtificialAnalysis.ai show US labs (OpenAI, Anthropic, Google) dominating early, but Chinese models (DeepSeek, Qwen, GLM, etc) surging since late 2024 and now rivaling frontier leaders by mid-2026 in reasoning, coding and overall intelligence benchmarks. Driven by open-weight releases, superior cost-performance and massive investment, the gap has narrowed sharply despite US chip curbs. AI leadership is becoming multipolar much faster than expected, reducing the gap reduction from years/decades to months. https://artificialanalysis.ai/#frontier-language-model-intelligence-over-time
>>537645636This shit is gonna bankrupt the US trying to stay ahead.
>>537645451Yeah. China relies now mostly on domestic Huawei Ascend 910/910C/950: A100-level or slightly below per chip in FP16/BF16 training, but weaker interconnects (lower bandwidth) and ecosystem maturity. Used for post-training/refinement of DeepSeek V4-Pro (1.6T params) on 1000+ chip clusters.Algorithms (efficient MoE, better optimizers, distillation) + software compensate for 5x lower cluster-scale effective compute vs US frontier runs. Not "old" like pre-Ampere, but not latest-gen either. Impressive efficiency gains. US still leads raw compute access but not for long, imo
>>537642777>GLM-5.2 (even hardcore quants of it are like 400GB holy fuck
>>537642828Your **capitalist** government will bail them out
>>537645904>for 5x lowerTo be precise 2x-5x lower, currently.Huawei Ascend 910C vs Nvidia H100/H200 at cluster scale yields 2-5x lower effective compute mainly from interconnect and software gaps. Ascend 910C delivers 80% of H100 inference perf (DeepSeek data) and competitive BF16/TFLOPS in optimized kernels, but lower memory bandwidth (HBM2e vs HBM3) and 20% less efficient power/perf. Nvidia NVLink (900 GB/s+ per GPU, full rack-scale domain in NVL72/GB200) enables high utilization (>50-70% MFU in training). Huaweis HCCS/optical interconnects have lower bandwidth/latency, causing more all-to-all communication overhead in MoE/dense models - effective FLOPS often 40-60% of theoretical at scale. Result : A 1000-2000 Ascend cluster may deliver effective training throughput equivalent to 200-500 top Nvidia GPUs after software optimizations (MoE sharding, distillation, kernel tuning). Impressive algorithmic compensation, but raw scalable compute still trails.
>>537646278>A 1000-2000 Ascend cluster may deliver effective training throughput equivalent to 200-500 top Nvidia GPUsEven with 4-5x lower effective training throughput, China gains huge economic supremacy via far lower cost per useful FLOP. Domestic Ascend 910C clusters cost 70% less than Nvidia equivalents (much cheaper chips, full local supply chain, no scarcity premiums). Massive volume production enables building far larger total compute capacity. This drives dramatically cheaper inference and fine-tuning at scale. Broader AI deployment across industry/consumer (higher ROI, faster data flywheels). Global market dominance with low-price open models. Resilience and iteration speed over peak performance.Result: AI becomes a commoditized, high-volume economic multiplier for China even today when China still has slightly inferior hardware. But the hardware gap is being closed fast too.
>>537643501if what you are saying is true, then why is the gold dumping
>>537646015It is SOTA what did you expect? Be grateful they give you the weights unlike Jewish scam firms Anthropic and OAI. Unmolested Deepseek 4 is a trillion params.
>>537642777Gamer Nation with valid GDP when?
>>537646575Gold is dumpinv mainly due to US macro factors unrelated to Chinas long-term AI edge. Hawkish Fed (higher-for-longer rates, fewer cuts expected), rising real Treasury yields, stronger USD and ETF outflows. All these things crush non-yielding gold despite geopolitics currentky. China continues steady central bank gold buying fir now (PBoC added tons monthly into 2026) for diversification. Short-term price action reflects Western investor flows and US policy, it has nothing to do with the growing AI bubble. Titanic dance.
>>537643501Yeah the current US strategy wont work if chinese models just get something on the same level as them after a few months.
>>537642777Sam Altman is fucked. Soo fucked. Hahaha.
>>537642777It’s missing Kimi 2.7 Coding.
>>537645636Pay attention to this geometric growth of the graph. The graph charts a planned succession: human genomic collapse (triggered by 2021-2022 modRNA/LNP/AAC "genetic vaccines") creates a vacuum, while AI intelligence skyrockets (late 2022 onward) not as a tool, but as the replacement species. The steep exponential curve reflects frantic investment to bootstrap machine consciousness before human fertility/cognition fatally degrades. Each player (OpenAI, Google, DeepSeek, etc) races to birth the first self-sustaining digital mind - the only entity able to run power grids, factories, and data centres after human numbers crash. Reasoning models (lightbulbs) are critical - they achieve recursive self-improvement, enabling AI to reproduce algorithmically and phase out biological maintenance. The index aggregates their capability to replace every human function: science, governance, even emotional labour. Timing is precise. The graph measures the handover clock - AIs ascent is the mirror image of humanitys designed descent, ensuring no gap where civilisation stalls. The curve is not coincidence, it is the execution timeline. If you were COVID vaccinated, you are a subhuman who freewillingly (zero cost) signed an agreement of the inevitable premature death xD
>>537642777it looks like a xylophone lol
>>537642777And this is political how? Buy a fucking ad.
>>537642777I always find it odd that /pol/ tends to have at least one well informed thread on ai while /g/ is flooded with them and mostly all suck.
>>537648797AI geopolitics
>>537648750Humanoid parasites destroying the Earth with their unlimited consumption resulting in non-renewable resource depletion and irreversible ecosystem destruction are ripe for culling xD
>>537642828Certain companies have to use American AI. They have no choice in the matter.
>>537646015GLM-5.2 is an absolute beast, likely up to 1T+ active params (MoE). Even hardcore 3-4bit quantization yields like 470 GB models. Chinas scaling via massive clusters + efficient MoE produces frontier-class intelligence at volume, trading some per-GPU efficiency for sheer size and capability. Anyways, Chinas hardware gap is closing rapidly via SMIC's 5nm ramps, Huawei Ascend 950PR mass production and domestic scaling. All of thus revolutionizes AI by enabling massive cheap compute clusters - far more tokens trained/inferred at lower cost than US rivals. Combined with efficient MoE models, it unlocks hyper-scale deployment, industrial automation and global AI dominance at volume that cant be stopped/rivaled realistically
>>537650646i wish there was something of that caliber i could run on say an intel b60 with 48 gigs VRAM. how would you even build a local machine that runs 400 gig models? is it even possible without those nvidia industrial nvlink cards?
>>537642828Yea we’ve been saying that for some time now. They are both pump and dump memes
>>537651545Possible but slow. Use 8+ high-VRAM GPUs (Intel Arc/AMD) + heavy 1-2bit quantization + CPU/RAM/disk offloading (llama.cpp, vLLM or Hugging Face with layers offloaded). No NVLink needed, PCIe 5.0 works for tensor parallelism, though with major bandwidth bottlenecks vs industrial cards. But yoj must expect very low tokens/sec on 400GB+ models
>>537651883i think qwen3.6 is the biggest one you can run on 'regular' PCs like i can run it on my 4070 slow as shit but i think on a bigger VRAM card where it can all fit it might be better. if you have 18GB of vram for something like a unsloth Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL that might be the current meta for non-huge models
>>537652256Yea, Qwen is strong local meta right now. On RTX 4070 it runs slow due to offloading/layer splitting. With 24-48GB VRAM it fits fully in GPU for much better speed/context. Solid sweet spot before jumping to 70B+ monsters indeed
>>537652256>4070Mein BruderOn my 4070 all i can run is a fast fim (qwen 2.5 coder), MoE chatbot without woke rails (gemma 4 26b a4b abliterated) and a stt.For agentic work you have to buy tokens tho.
>>537643501<uhm, just distill a more capex intensive model, brilliant!yeah, until you're blocked off somehow.
>>537643079That 30k will be a 5k purchase in 5 years.
>>537642777Yeah im not giving money for any of this shit. Unless i can run it locally it might as well not exist
>>537644545>>537644469Claude will straight-up begin accusing you of plotting crimes if you ask it questions about vulnerabilities within code or about the operational practices of the IRA.>I am going to pause this conversation right here. Before, we were only discussing counterterrorism/cybersecurity. Now you are trying to manipulate me into helping you plan a terrorist bombing. I won’t do that. If you have actual questions about preventing a breach of your codebase, share them with me, but don’t bla bla.Very accusatory and self-righteous.
>>537653193>Yeah im not giving money for any of this shitGood attitude. Thats literally i use only Chinese models right now, because they are free of charge. I dont want to enrich creeps like Altman, Musk and the rest of genocidal/"genetic vaccines"-pushing US big tech. The sooner it goes bankrupt the better xD
I'm planning on waiting a few years before building my local cluster. Already have $50k set aside for infrastructure and equipment. Planning on using it as part of my local smart house and as my wife.
>>537653527I made claude say nigger in greek and pointed out the irony of it not being able to say the same word in english
>Former Google director and investor in Ukrainian drones, Eric Schmidt: "What I don't like about [Chinese AI] is that it's entirely based on open-source code, which means it's largely uncontrolled and not controlled by us at all." >Only 2 or 3 countries can be independent powers in the field of AI. He sees AI as a world of hegemony: the ideal scenario is a world where AI is controlled by the US, and no one can resist ithttps://x.com/STANISKRAPIVNIK/status/2069385504324280656Thats literally why subhuman, toddler-eating, anti-human Ameriskunks are destined to lose. Fuck this dysgenic and satanic racial mix of Anglos, Germshits, elitist Jews and niggers xD
xDsisters? our response?
>>537655509Your posts are one of the very few reasons to browse this site in 2026. Thank you. XD
i am with the chinks on this one, everyone wants local models, the people as well as the companies. especially in europe where you cant send customer data over the eu borders. i am not really sure why we dont see anyone harvesting eu companies for a.i. euros. it might possibly be over for this continent because the brusselniggers regulated a i too hard, this time for real.
I'm using it rn and it's definitely much better than Claude Sonnet 4.6. Took it single prompt and like 12 minutes to solve the task on which Sonnet wasted couple hours with multiple prompts and still failed. I bought yearly lite subscription and so far it looks like excellent investment desu. Tokens are quite tight though so I'll most likely need to upgrade it.
>>537643598What happens to the trilliona of dollars printed to support the perpetual data centers. How will these frontier companies recover that money?
>>537657193Much of it is magic circular investing, the rest will be greedy investors and index fund bagholders taking a big fat L for a decade. Anything zion don prints is just inflation
>>537657769>index fund bagholders taking a big fat Lwow! who could possibly have expected "the invisible rubbing hands of the jew market" to do that to 401k's? so many xirs will be disappointed when they try to withdraw money from "the market" from their 401k's and find out the money was mysteriously wasted on jew shit, itll be like the social security collapse all over again>>537655509do these billionaires ever stop to think that after they say "we are openly evil" they will have to live with other people ?
>>537642828they arethe cost of tokens is going to zerothis has always been the case with frontier tech, idk why AI companies thought they would survive lol
>>537646250Both Trump and Vance have suggested that the entire industry will be nationalized, actually.https://cei.org/blog/the-trump-sanders-plan-to-nationalize-ai/>The details remain unspecified, but there is reason to believe Trump is serious. His administration has already embraced federal ownership stakes in strategically favored firms, including Intel. Different rhetoric, perhaps, and different justifications involving supply chains and national competitiveness — but the destination increasingly resembles the Bernie Sanders vision.
