Fuck AI

Nah that's a bit of an exaggeration they're pushing, yes it's getting more expensive to buy computer hardware and they'll try to keep that racket going too, but you honestly don't need that much computing power to run lots of really capable, specialized LLMs and other neural-net models at a reasonable or at least usable speed, it's not nearly as much as they say. Some decent stuff will even run on a smartphone (not that I recommend using smartphones for anything) or raspberry pi. All this time, the big AI companies are chasing each other to the bottom of the barrel making their AIs bigger and huger than everything ever before in history to do everything all the time all at once, calling it more "dangerous" for marketing purposes, but also making it sound smarter while actually being dumber so it takes longer (and more tokens) to get anything done, and ultimately just making it more expensive so they can charge more for it. Meanwhile amateurs and hobbyists are making it smaller, narrower, more precise models that can run quickly and efficiently, that are in many cases almost as capable as the so-called "frontier" models with a fraction of the size and inefficiency inherent to the expensive, proprietary mega-models.

For the few and far between things that genuinely do need that much computing power and resources, there will always be hardcore prosumer enthusiast assholes like me willing to deliberately and obsessively spend unconscionable amounts of money on ownership of the technologies we need to survive and thrive, and then sneakily, carefully, anonymously share some of those capabilities with everyone because sharing is caring. <3 My server racks and hundreds of terabytes of disk arrays bow to no tyrant but me. People who think they'll price me out of this hobby misunderstand the lengths I have always been willing to go to collect and run this kind of nonsense equipment for far stupider and less useful reasons than this.

Problem is the distilled llms are getting smaller and are requiring less and less power/gpu/cpu processing for use, and there’s now varying types and sizes of models so you can use many very small and very specifically trained models to get more and more done. That trend will continue until some lower limit is hit, kind of the opposite of what happened with the max limit for transistors and higher cpu processing speeds.