Remember the early days of Uber and Lyft, when rides were dirt cheap because the companies were operating at a loss in order to capture the minds/wallets of the masses?
-
Remember the early days of Uber and Lyft, when rides were dirt cheap because the companies were operating at a loss in order to capture the minds/wallets of the masses?
The rug pull in the AI/LLM world when the companies adjust pricing to actually make a profit is going to be spectacular. Especially when you consider the numbers of people / orgs that are addicted to or dependent on such technology.
@wdormann
I'm just using API pricing up front to make sure the economics still work out.
The $20/month plans will go away and the $200/month plans will be scaled back, probably. -
@wdormann Do you think companies that start installing LLMs directly on users' machines may have an edge in this war? Considering they're offloading the price of energy onto the end user?
https://www.thatprivacyguy.com/blog/chrome-silent-nano-install/
@mdm @wdormann Yes. That will reduce the blood loss. You still have to train it, but you can train it once for all users, you can even distribute the training for the next model among your users too. The problem is efficiency, cloud providers *could* get efficiency through scale, use renewable energy, reuse cooling water. They typically dont, they do whatever is cheapest which is usually to freeload on the local town's resources.
-
@wdormann while I am critical about LLMs (but also use them) and it is clear that companies are currently trying to capture/create the market through losses, there is another scenario where LLM training/inference gets much cheaper through technological advances. But I am not sure of how probable this is.
@aaronkurz @wdormann A current trend among people really into AI is that they try to compensate for the inaccuracy by running the same text through multiple AI products, then asking one of them to combine and summarize the results. It seems plausible to me that resource waste will expand to compensate for technological advances, just like it has with RAM and disk space.
e.g. https://github.com/karpathy/llm-council -
@wdormann
I think the difference might be that Uber and Lyft provide actual services, even if they are leeches and screw their employees to death. Whereas AI doesn't provide anything of any real value anyway.Quasit Most employers don’t want anything of value. They want to feel powerful, have a sense of control, and of being right all of the time.
LLMs provide all of that.
-
Quasit Most employers don’t want anything of value. They want to feel powerful, have a sense of control, and of being right all of the time.
LLMs provide all of that.
@kichae
Last I heard some companies had to start rehiring humans because AI couldn't do the work it was supposed to. -
Remember the early days of Uber and Lyft, when rides were dirt cheap because the companies were operating at a loss in order to capture the minds/wallets of the masses?
The rug pull in the AI/LLM world when the companies adjust pricing to actually make a profit is going to be spectacular. Especially when you consider the numbers of people / orgs that are addicted to or dependent on such technology.
Ditto streaming services.
-
@bmoreinis @sleet01 @wdormann @aspeed
Okay, fine, he's MY bard anyway!
-
@wdormann what if it took 5 or 10 years maybe until there were cost-reflective pricing for llms?
-
@wdormann I think we're at the point where it's an opiate of the masses.
I gotta say I'm looking forward to seeing people go cold turkey, especially those who have the habit of taking LLM output and pretending it to be their own acumen at work. Or those who use ChatGPT to craft forum rebuttals and such - when this tide goes out, we'll truly see who brought their swim trunks and who never had them
@resonancewright @wdormann thank you for this metaphor! 🤩
-
@mycotropic @aaronkurz @wdormann Chinese research already went two order of magnitude in improvement in energy efficiency. The point is you spike inference like the brain. And the find out you get 99% Energy consumption reduction. But there's a law of economics that states the more efficient you figure out how to use a resource the faster you deplete it. But every pleb flies now and every pleb will code, Ryanair coding, that is.
I haven't read it but i add it as ref
https://arxiv.org/html/2509.05276v4 -
@mycotropic @aaronkurz @wdormann Chinese research already went two order of magnitude in improvement in energy efficiency. The point is you spike inference like the brain. And the find out you get 99% Energy consumption reduction. But there's a law of economics that states the more efficient you figure out how to use a resource the faster you deplete it. But every pleb flies now and every pleb will code, Ryanair coding, that is.
I haven't read it but i add it as ref
https://arxiv.org/html/2509.05276v4@splinux @mycotropic @wdormann interesting, thank you!
-
T tokeriis@helvede.net shared this topic
