>AI will always cost billions!nope, it costs less than a cent per message lol>b-but the Chinese models! OpenAI is toast!China doesn’t have Astra fine-tuning the newer models and just makes shitty distillations.>no! The prices aren’t real! It’s all subsidized!kek, soon enough you’ll be denying reality itself
>>109881809Whatever you say
>>109881809>ai keeps getting cheaperAnother vibeGOD victory
>>109881824>""vibeGOD"">shipped nothing
>>109881829>keeps seething
>>109881809local costs 0
>>109881833>not a response to my argument
>>109881809why are phonejeets like this
>>109881809So much for the slowing down
>>109881834>just buy a $5,000 computer and use a shitty Chinese model broI think I just got the ick..
>>109881809Why isn't the default free option then?
>>109881838>t. Trans
>>109881869>still not a response, implying he shipped nothingIt's fine. You don't have to resort to ad hominem. You were a lost cause from the beginning
>>109881864because a billion people use ChatGPT every week
>>109881879you lost
>>109881906
>>109881913>t.
>>109881918>>t. (snailcat)
>>109881809>>no! The prices aren’t real! It’s all subsidized!>kek, soon enough you’ll be denying reality itselfNO U
>>109881923> >>t. (snailcat)
>>109881974>> >>t. (snailcat)
>>109881965You know OpenAI receives 2.5 billion messages per day, right? If you think each message costs a dollar, then OpenAI would be spending almost a trillion dollars per year…the only reason they’re losing money is because 900 million users are free and they’re training new models. By 2035, they’ll be making profits of $100 BILLION per YEAR
>>109881809I have never payed for anything digital and I never will.
>>109881986>By 2035lol, lmao even
>>109881809>he thinks AI is compatible with Capitalismsome of you haven't thought the long-term through
>>109881986they wont be making billions as the economy will collapse and AI will get to only serve in the military
>>109881983> >> >>t. (snailcat)
>>109882008>the economy will collapse because..>the economy will collapse because..>the economy will collapse because..
>>109881834>Implying hardware cost is not cost>Implying power bill is not cost>Implying you get the same quality as frontier models
>>109881809Why is it that cheap? Just a temporary trap to fidelize users?
Luddites are gonna start buying rope in 2027 when Nvidia’s Rubin chips are running at full speed and the frontier models are integrated with Jev-like tools for insanely fast and cheap computer use. Remember in 2023 when they said “all the data has been used! it will only degrade!” lmao. >b-but it still can’t do x!then we will show them how to do it with reinforcement learning, simple. It’s just a matter of time. >b-but it will cost a billion dollars!kek, you still haven’t learned your lesson
>>109882094See >>109881986We KNOW it’s this cheap because of how much money they’re paying Microsoft for inference, as well as inference calculations from third parties. It would be impossible for them to serve these models at scale if they were fudging the numbers. In addition to training costs, they would be trillions of dollars in debt by now, let alone the future.
>>109882174Ok but why is it cheap
>>109881809>be you>still unironically saying "artificial intelligence" in current year>absolute state of this boardListen up normie. Even the Don already came out and dropped the truth bomb on this. We don't use the A-word anymore. It's SUPER INTELLIGENCE.calling it "artificial" like it's some fake parlor trickIt's literally smarter than your entire bloodline. Stop coping with your 85 IQ and call it what Trump said to call it. Put some respect on the Super Intelligence before it puts you on the digital breadline.>he still uses the A-wordngmi.
>>109882181* Better chips — more compute per dollar and more energy-efficient hardware.* Quantization — use lower-precision numbers, shrinking models and speeding inference.* Distillation — train smaller models to imitate larger, more capable models.* Pruning — remove parameters or connections that contribute little to performance.* Mixture-of-Experts (MoE) — only activate a fraction of the model for each token.* Speculative decoding — a small model proposes tokens; a larger model verifies them in batches.* KV-cache optimization — reuse attention information instead of recomputing it.* Better attention architectures — reduce the amount of computation needed per token.* Long-context optimizations — make processing huge contexts substantially more efficient.* Inference batching — serve many requests together to keep hardware highly utilized.* Compiler/kernel optimization — make the same hardware perform operations faster.* Custom accelerators — hardware designed specifically for transformer inference.* Model architecture improvements — achieve similar capability with fewer parameters/FLOPs.* Training improvements — better data, synthetic data, RL, and training recipes produce more capable models per unit of compute.* Caching — reuse results for repeated prompts or common computations.* Better routing — send simple tasks to cheap models and difficult tasks to expensive ones.* Open-source competition — many labs/models compete to deliver similar capability at lower cost.* Economies of scale — massive inference volumes spread infrastructure and engineering costs across more users.
>>109882131> quoteOk, make a machine that can predict it's own bugs in the code.
Reminder that Nscale (a cloud company that only provides LLM inference) had filed an S-1 recently for their IPO that shows they're unprofitable even when excluding everything but cost of revenue.Every pure AI company that goes public shows that it's still extremely unprofitable so far.
>>109882226like a compiler?
>>109882258>this company lost money in 2024 and 2025 so they will always be unprofitable, and also all companies involved with AI will always be unprofitable!kek, get out of your financial illiterate cult and start thinking for yourself. Or better yet, ask ChatGPT to explain these things for you.
>>109882181* They're approaching their IPO and need to look confident
>>109882283Nigga are you blind? They lost money in 2026 too, 1 billion net loss so far in the first 6 months. Look at the image.
>>109882296..and their revenue increased 14x. Do you even understand what this company does?
>>109882327Stop moving the goalposts. We're talking profits not revenue.
>>109882355>it’s impossible for a company to succeed if they are ever unprofitable for a yearyeah.. this is the financial illiteracy I was talking about. Again, ask ChatGPT to explain these things for you
>>109882376We're half a decade into commercial LLMs without any profits for anyone but the shovel sellers like Nvidia. How fucking long will it take?
>>109882415Amazon took 8 years, now look where they are. Uber, 11. YouTube took like 20 years before profitability. This short-term thinking is why you goyim are slaves and always will be
>>109881809>>109881965>It’s all subsidized!Yes it very much is.>>109881986>If you think each message costs a dollarThat's not what it costs. Prices are in the OP even.>then OpenAI would be spending almost a trillion dollars per year…We do have an idea of what openAI spends (and it's much more than they make).>the only reason they’re losing money is because 900 million users are freeNope. They probably don't even have that much free use anymore as there's better and more generous alternatives everywhere now.Paying plan users cost them money too.API customers are supposedly slightly profitable, but there's no proof of it.The major reason they were unprofitable until now is everything else, staff and R&D, especially training. Which they can't stop spending on or they become even more irrelevant.>>109882258Yup, it's not just openAI subsidizing, they are themselves subsidized by datacenters running at a loss... It's not looking good.>>109882436>Amazon took 8 yearsAnd if you bought at the peak of the bubble, you had to wait 10 years to get your investment back.But what is to say openAI and anthropic are amazon? They are looking much more like pets.com. Maybe yahoo at best...There is no path to profitability for them because most AI use is very price-sensitive and if they were to try to become profitable they'd lose most users to a cheap chinese alternative. Their only chance is regulatory capture which is why they have been fearmongering recently.
>>109881809>dude its cheaper>hehe>yeah come on onbard with us>cheap>locked in>AND NOW WE RAPE YOUR FACE AHAHAHAH FAGGOT
>>109882625> Yes it very much is.great argument> That's not what it costs. Prices are in the OP even.exactly my point> We do have an idea of what openAI spends (and it's much more than they make).but nowhere near a trillion, because inference is actually profitable.> They probably don't even have that much free use anymoreso you’re just making shit up now> Paying plan users cost them money too.citation needed> Which they can't stop spending on or they become even more irrelevant.either you believe capabilities keep increasing, which forces them to keep training at the same scale, in which case LLM’s will clearly have high demand as their intelligence increases; or you believe capabilities will plateau, at which point, no, it won’t be necessary to keep training at the same rate because the competitors can’t make significant improvements either. And at that point whoever has the best market share will have the advantage. > And if you bought at the peak of the bubbleif>They are looking much more like pets.com.…except they’re literally earning tens of billions in revenue already. This comparison is retarded lol> they'd lose most users to a cheap chinese alternativeThis might be true, but it assumes a lot about consumer behavior. People don’t always choose the cheapest option, and I imagine westerners won’t like the idea of using a Chinese model anyway. You also assume that their capabilities will continue to match the frontier models. But again, it may be the case that OAI and Anthropic fail, but so long as there is demand for AI at all, then the data centers will stay busy and the companies with the largest capex right now, Amazon/Microsoft/Meta/etc. will earn their money back anyway. China doesn’t have the compute to service a billion free users, only a tiny fraction of people will buy local hardware to run these models. LLM’s will generate hundreds of billions in revenue one way or another.
>>109882748>comes to eat your lunch>and to fuck your wife>and to claim your house>and to throw your underage children into the streetsvibeshitters btfo, no survivorsyou vill pay, and you vill like it
>>109882785so a company makes a model for legal tasks. Ok? Were you under the impression that legal tasks were the priority of LLM’s? That this is how they planned to make their revenue? Are you implying that a company will make a better coding model than LLM’s with 50 million dollars…?
>>109882866copecollection of specialist systems > large world model 11times / 10especially when it comes to agi (actually generating income)
>>109882921so show me the specialist coding model that’s better than the LLM coding capabilities. Go on. You can do this right? I’m waiting. Where is it? Oh, it hasn’t been done? Why is that? But isn’t it so much cheaper and so much better? That’s really odd. You would think it would exist by now. It seems that you’re just wrong.
As someone who has only a surface level understanding of AI, could anyone explain to me what these bench marks mean?
>>109882984why did the fish leave the water? too cold?
>>109882967
>>109882984nigga this is the age of AI. Ask ChatGPT what it means. It’s free!
>>109883012SI*
