>everyone is going to abandon frontier models and agent harnesses so that they can save a few bucks on tokensNot sure what's with this Chinese propaganda. You don't see companies hosting on some chink cloud, or buying their employees Xiaomi phones. The West generally cares about reliability and SLAs and you don't get that with chinktech.The CCP has been running a surveillance state for decades now and despite having a camera on my every street corner, they never figured out LLMs, transformers, or agentic loops. They should have been ahead of the fucking game but they will always copy first. They don't believe in themselves.
>>109252854openai released oss in 120b and 6b, it's not functionally the same as gpt-5.whatever but it has the harness and thinking parts in it so you can probably finnagle it to get something appreciable to 5.x and it will fit on a single H100. I doubt everyone is going to switch to chink models this is the same concept as when everyone was saying cloud bills are insane everyone will switch to on prem and then only a few companies who could afford it were switching to a "light hybrid" setup. What's old is new.
>>109252854>so that they can save a few bucks on tokensSo they can save hundreds of thousands if not millions monthly on tokens.
>>109252854China COPE
>>109252854local inference is also MUCH faster than what apifags can get if you don't have a poorfag gpu.and it's good enough 95% of the time.
>despite having a camera on my every street corner, they never figured out LLMsLLMs don’t just happen when you have a lot of data.
>>109252854I pay openai / antrophic for bleeding edge model access for planning, which I then give to glm 5.2 or deepsneed v4 pro (depending on complexity) to implementI usually burn through ~30 million tokens a day and this setup cuts the spend by at least 80% while still keeping good quality.corpos are hosed since muh data ownership and they're too stupid to buy their own racks, but there's no reason to avoid chinesium if you don't have to care about nonsense like this
>>109252854Real reason why frontier stuff is dropped - 99% of world AI usage is just spamming/scamming internet with every form of slop imaginable by 3rd worlders (I suppose currently there is massive industry build around that).For that purpose usage cost is massive factor so 3rd worlders can squeeze that few dollars of profit every month. That's why they will use cheapest models possible and use of western ai will be just too expensive.
>>109252854>everyone is going to abandon frontier models and agent harnesses so that they can save a few bucks on tokensIt’s not just that.My company is already considering using a self-hosted GLM 5.2 model after Anthropic refused to provide proof that it neither trains on nor stores our prompts.We did the math, and we could fully meet the needs of 1k employees for around 100k euros per month. That is still about four times more than what we currently pay Anthropic, so we will not switch for now.But, by around mid-2027, we will almost certainly make the switch, and so will many mid to large companies.
>>109252854>frontier modelslollmao even
>>109252854I think gemma 4 is pretty good tbqh
>>109252854>everyone is going to abandon commercial Unix operating systems so they can save a few bucks with a hobbyist kernel like LinuxOh, I'm laffin', 1996 bros! I'm laffin'!
>>109252854>The West generally cares about reliability and SLAs and you don't get that with chinktech.Not the same when it comes to LLMs which many companies are already adopting Chinese open source models that compete at a similar levelYou sound more like Peter zeihan than an actual AI researcher, engineer or user so I'm gonna ignore your opinion
>>109252854The biggest cloud providers constantly update with the latest open weight Chinese models but according to local /g/ user there isn't any market demand
GLM 5.2 mogs GPT 5.5 and shows open weight coding models are at the frontierhttps://youtu.be/xwC9MwfZLn0?is=h2qrw4tlZmvpWH65
>pls install this CCP AI to control all your company Lmao
US propaganda against china is to distract from their own survellience state.
>>109252854Grok imagine is already built on top of Flux 1 with their own fine tunes and loras based on it.
>>109258047glm 5.2 is just benchmaxxing
>>109256062>My company is already considering using a self-hosted GLM 5.2 model after Anthropic refused to provide proof that it neither trains on nor stores our prompts.It's comedy gold to hear that companies finally decided to consider this. They do train on your prompts and this is the major reason why programming jobs that require a security clearance don't use AI still. That and the fact that you don't want to transmit all of your classified code over the internet, even encrypted. If Anthropic were really smart they would just deploy the copies of every major website that they already have saved, since you cannot own LLM generated code anyway.
>>109256062>proof that it neither trains on nor stores our promptsNot possible.
>>109252854its chinamen and lefty "muh drumpf bad muh peter thiel technofeudalism 2 more weeks" cope to try and pretend the US models arent the top of the game and china isnt dependent on copying them
>>109252854If it's at the point of "good enough" for half the cost then corps will adopt it
Anthropic models have gotten really really bad, we're hitting diminishing returns and the jews keep lobotomizing the models to steer them to their worldview.Chinese models are cheap, but they also rely on US frontier models to steal and train.>>109258714They already are, DeepSeek V4 is extremely cheap and can do 90% of what the frontier models do. That's good enough for a long while.Even Gemma models are now decent enough, with a few tricks here and there and connecting to external APIs they can probably do 90% of a frontier model does.US companies should switch their business model to selling models/licenses to their models because the grift will be over soon enough.
>>109254450But they certainly don't happen without it.
>>109252854Maybe your company wouldn't be rapidly heading towards bankruptcy if you didn't spend your money on shilling your shit to terminal NEETs.
>>109258675Homomorphic encryption or simply deploying an on site version of their model like they already do for the US govt anyways.>NOOO THEY CAN'T DO THAT FOR... REASONS!Well enjoy losing these customers forever then. Any company investing into an on site GPU cluster will not consider your services for the next ~5 years and by that time all Anthropic and OpenAI will both be bankrupt.
>>109252854little miss bug eye isn't wrong. NVidia has spend shit tons of effort on Nemotron and its vllm ecosystem. the new DGX station is going to be the craziest selling platform.Manufacturing isn't going to put up with token bills
>>109258885>>NOOO THEY CAN'T DO THAT FOR... REASONS!They can't do that because homomorphic encryption is orders of magnitude slower than normal computation and LLMs are already dogshit slow? Have you ever looked at the current HE options that exist and what the drawbacks are?>Any company investing into an on site GPU cluster will not consider your services Of course they wouldn't, they're investing in a local cluster.
>>109258152they distill claude and benchmaxx itthat said, op overstates private companies concern for OpenAI and anthropic being their provider. they just want the best product.
>>109252854>Company spending a fortune on Anthropic tokens.>Owners start complaining.>Tech sector suggests using DeepSeek.>Owners say they’ll sing the Chinese national anthem if the AI is cheaper.In the end, she’s right. Cheap APIs offer the same service as expensive APIs for at least 80–90 percent of real-world applications on the market today.Only the tech and financial sectors will likely end up using front-line models down the road. Everything else will use the “good enough” model.
>>109258047>20 minutes of straight up soilennial yappingts could have been 3 minutesi hate modern youtube
>>109252854if the open source chink models are 80% as good per token, but you can use twice as many tokens per task and still spend 50% less it makes more sense to use the chink models for all but a fairly small number of jobs where either cost isn't a concern or tokenmaxxing can't make up for the lower performance.
>>109259573Same way it's cheaper to buy a BYD EV and import it from Canada than it is to buy a Rivian or Tesla, right? Or the same way it's cheaper to hire Cognizant to get an Indian to manage a DB2 database every 6 months than it is to hire some US college grad who took a class on COBOL.
>>109252854For all its faults, Meta under the Zuck had understood this years ago.Unfortunately, they suck shit at this.
>>109259596Exactly.I’d like to own a BYD, handing my money over to the Communist Party or to some ultra-rich person who lives in his own bubble and treats me like a slave is the same shit. I’m the one footing the bill for an elite that literally invented the global model and has started to lose it.The crumbs that fell from the table of this elite were excellent 20 years ago, but they aren't anymore.
>>109252854>You don't see companies hosting on some chink cloud, or buying their employees Xiaomi phones.No, but you see companies hosting on Azure, buying their employees win11 laptops and giving them C*pilot licenses precisely for the same reasons listed on op's pic (or in europe's case, they are now starting to use foss alternatives even if they are shit).I don't think it will necessarily be chink AI taking over the enterprise market, but most businesses won't be using cutting edge AI products. They will buy something decent and cheap that they are familiar with and that's it. Extremely few businesses need any sort of cutting edge tech aside from the specific tech that's core to their business.
>>109259798Most European companies are more concerned about data governance and not OSS. We self-host a ton of crap (including AD and whatnot) but a lot of it is not OSS.What would shake up the European market would be a bunch of blade servers/racks you can buy that come predeployed with some OpenAI/Anthropic model. You are contractually not allowed to touch or modify it, BUT you can airgap it, put it in your own server rack and ensure it doesn't phone home to OpenAI or Anthropic.They would make money hand over fist with this shit. I think Mistral is already doing something like that (their problem is that their models are shit)
>>109259823And little by little, this is pushing its competitiveness into the abyss.There are two types of protectionism: one that guarantees a bright future namely, the kind that ensures the development of your domestic industry so protectionism that promotes technological advancement, as China has done, is incredibly effective.Now, protectionism aimed simply at slowing your decline, as the rest of the world does whether in Europe, Latin America, and so on is simply a disaster in both the short and long term.
>>109259899I don't get it either, most Anthropic engineers make more money than our CEO does and are not interested in stealing our shitty app and competing in our tiny market.Nevertheless here we are. I was merely suggesting a way for OpenAI/Anthropic to make some quick bux in Europe.Your post is mostly true though
>>109259735>crumbs at the beginning of a meal are superior to the crumbs at the end of itcattle.
>>109260076Anon, anthropic used the data they gained from watching how people used their models and the retards who "tokenmaxxed" to create services which the sold at a premium. They are in fact using you retards to train their models and learn things they can roll into the LLM and sell back to people. Why WOULDNT they ice you completely out of the fucking market? Dey guud boyz dey dindu nuffin? Dey makes nuff munney dey don need no mo'. You naive dumbass. OpenAI is going to fed court now over corporate espionage from Apple. "AI" models being sued by major companies and nations over IP theft is happening now and it's not stopping.If they start having concerns about national data leaks and national espionage...it's over. They will go out of their way to block openai and anthropic from using any of their data and use opensource models they keep inhose to do everything they want. The open hightrust internet is completely over now. It's dead. It's done. And the major LLMs NEED an open unaware internet and users to train from constantly. They do not have that anymore and there will be people injecting poison as well. People already know W11 is a data theft and LLM training tool masquerading as an operating system. Google is even worse in that regard. I think we should all understand that the party is over. The free open high trust internet is dead now. Nothing is trustworthy, no open organic sites are findable or safe anymore. It's done. And the "AI" companies that NEED this are fucked forever because every avenue to keeping these models sane and correct are closing with many getting ready to go to war agaisnt them.
>>109260084Not him, but just the truth. I don't care if billionaires are billionaires and exploit us; my father had a great life, but I don't have that chance. So the crumbs from the first meal were actually much better and fostered social cohesion more than they do today.
>>109252854It's not some nebulous propaganda, it's just common sense. American AI is much more dangerous to sovereignty and freedom of any state and no sane state will implement any American made or backed AI. American anything is just a liability, a danger, a trojan horse, which is why everyone is moving away from American tech as a whole. America is a proactively harmful imperialist police state.Chinese AI is open source and cheap, and can be adapted to anyone's needs with no danger whatsoever. It's that simple.
>>109260339Bro, the foundational americans are sickened, poisoned, dying, partially lobotomized slave race who don't know they are all in fact fucking poor. The tech-oligarchs, oligarchs, and political class are not america. Nor are the illegals they are flooding into the country and displacing foundational americans with. America is a dying state now. We have no control over any of this and claiming we are all part of this debacle is frankly insulting.
>>109252854>a few buckskek
