Is Anthropic shitting the bed because they don't have access to enough compute (not enough data centers around to support the demand) or is it because inference is not profitable and they don't want to burn all their cash on said demand and the actually readily available compute?What's the end game here?
The problem is that there will never ever be enough compute because if a certian task takes 1 minute to complete, then you need to double the compute to complete the same task in 30 seconds, then 15 seconds, 7.5 seconds, it's Zeno's Paradox. The solution is not increasing compute, it's training the customer to improve their prompt rather than hoping they get a good slot roll on their generation, to use fewer agents to complete the same task even if it takes longer or the quality is slightly worse, to spend time fixing buggy outputs rather than throwing the entire thing away and starting over. This is a hard problem since it's always easier to punish people than it is to train them.
>>108660874The problem is they're using AI to write their AI but AI is retarded so it just gets worse.
>>108660874OpenAI did the same thing. They are all making negative money with a great product, but need to pull back costs and freebies somehow to show a positive trend
>>108660874Musk just hijacked their biggest customer. So, that should save them money.
>>108660874They're not profitable. They can't deliver AGI or most enterprise work processes with sufficient reliability.And while the product would be good to produce profitable porn where some flaws are ok, they stopped that.
>>108661011interesting. and eventually all the freebies will be taken away and they will not be able to make the line keep going up without making their subscription prices go through the roof
>>108661136I think the idea was, AI was s'posed to keep advancing to a level that would make it's use compulsory, regardless cost. So much for plans.
>>108661136Give people a taste, get them hooked, certainly have people overuse the free tokens you give them and reel it back in until the poor plebs cant afford it anymore.Grok did the same thing. You use the poor people to test and tweak things for enterprise
>>108661010Singularity bros, not like this.
>>108661103>They can't deliver AGINo one can deliver AGI. You cannot turn LLM's into an AGI.And to answer OP, yes, inference is not profitable, it takes an insane amount of compute power to run LLM's that are good enough to do all this shit. We're at the infancy of the technology but every corposcum wants to force it upon everyone as if it's already here, even though it costs way more to run these data centers than the profit they make by these models doing the work. The only reason it's still ongoing is because investors are dumb as bricks, believe every hype lie they're told, and truly believe this shit'll bring them profit, and not that they're being played like chumps by all these corpos so that they can amass both hardware and money to do whatever the fuck they want to do.The moment all investors mutually realize this simple fact, and stop gaslighting themselves that that's the case, that's when the global economy will come crashing down and we'll be able to finally start figuring out how to use this technology well.
>>108661391>We're at the infancy of the technologyBut if it all comes tumbling down now, wouldn't investors and middle maniggers never touch it again and label anything "ai" as investor trap? If this happens we'll remain at infancy forever and might even extend to already useful applications like computer vision or whatever google used in their protein folding breakthrough
>>108660874boththey dont have enough resources and loses tons of money. we know that they lose money on api costs, and if you max your subscription, you are using around 10 times more compute per dollar
>>108661391But this data center crap is starting to feel like a paperclip optimizer where we are going to run out of resouces building data centers when we should have been saving those resouces for what actually comes next. It's like if computing technology stopped in the 1970's and computers still took up entire warehouses and the solution was to just keep building warehouse computers.
>>108661460Then you'll need a hail mary like Nintendo releasing the NES after the video game market crash to reignite it all.>>108661495You could say we're at a history circle reset point here. Attention Is All You Need gave us the template on how to optimize these models to be more effective, but we are hitting a hard wall in terms of compute power and code efficiency to scale it in any other way than more hardware.We basically are at the infancy stage because it'll take years of heavy work and optimization to be able to do more with the same hardware, and the reason you have the current datacenter craze because this revolutionary technology fell victim to an economic bubble. Much like the invention of the World Wide Web and the dot-net buble that followed.
>>108660943But nobody is trying to complete the task as fast as possible. Pretty much everyone just compares coding AIs based on intelligence.Literally nobody says>I used claude and it took 43 seconds to answer my prompt, but I used codex and it took 4 minutes 19 seconds, how slow!Nobody is measuring that. What they are measuring is how "smart" the model appears, how well it comes up with solutions, how much it misses useful things in the code and reinvents the wheel or hallucinates non-existent stuff, and how often can you say "AI, implement X" and come back to a working feature vs. coming back to a shitty mess.
>>108661589https://www.reddit.com/search/?q=claude+slow+response
>>108660874no, they're shitting the bed because they have hit the ceiling, it's time for bullshit openai style.
>>108661460That didn't happen after the railroad bubble, air travel bubble, or the dotcom bubble. There's no reason to think that AI investment will totally stop, it's just that AI investment will have to actually make sense instead of throwing buzzwords at people and threatening to replace every human worker.
>>108661625If you weren't a retard you'd understand nuance like "my car used to easily go 60 mph and now it can barely hit 30 mph" in the context of this conversation
>>108662866>railroad bubble, air travel bubble, or the dotcom bubbleAt least the dotcom bubble is really different. People keep repeating that the bubble burst and the internet is here but they ignore that the internet (as infrastructure) was the hard part and the bubble and hype at the time took care of it. After that the ways to generate revenue very easily led to profit as costs to companies using the already built infrastructure were very low.Datacenters have 4-5 years of actual life in their gpus and 2-3 years if they want to keep up with competition, and around 60% of a datacenter's cost is the gpus. The cost and revenue difference between AI (whether subscriptions or enterprise api usage) compared to the web with ads and ecommerce which were almost free money, is huge.Without labs and investments they need, or without keeping the managers in companies like Google and Amazon interested for long enough, there will be no research going to improve the llm architecture and we'll keep waiting for miracles like the attention is all you need paper to happen again.Also, "The bubble will burst and useful AI will stay" might as well just mean things like computer vision, machine translation and other cheap things to run, not the LLMs.
>>108663018>"The bubble will burst and useful AI will stay" might as well just mean things like computer vision, machine translation and other cheap things to run, not the LLMs.Good. Those are the useful parts of AI.
>>108660874their problem is their ceo is a fucking idiot who wants to be elon musk but doesn't have a trillion dollar company and a western space monopoly. retard literally thought he could tell the dod what it could and could not target (effectively making himself the president and circumventing the constitution) and has basically gone from retard strength to retard strength from there.
>>108660874https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropics-model-context-protocol-has-critical-security-flaw-exposed
>>108661460The people actually interested in it will keep refining the tech in their labs.
>>108664288the biggest improvement in gpt happened when a paper showed scaling doesn't lead to overfitting and instead makes the model keep improving.Research about scale like this requires investor money not just interest.Having fewer labs working and the unavailable cash needed for compute mean it might as well be another ai winter.
>>108665336It's almost as if AI only moves forward one death at a time.
