Quote:
Originally Posted by loramin
[You must be logged in to view images. Log in or Register.]
Well power is certainly a factor, but there's something more fundamental to the tech.
You have to understand "AI' is just math. Everything you tell the AI (ie. all context) becomes part of its calculation. There are fundamental limits to how much context it can process. Throwing more processing power (which requires electrical power) can increase the "context window" some, but only so far.
It's a diminishing returns thing with how the math works: the thousandth context token takes more power than the first, and the ten thousandth takes way more, and so on ... much like how level 60 requires way more effort than level 1.
You might pay a neighborhood kid to level to 2, but it will cost wayyyy more (more than 60x) to pay him to get to 60. Even if you inherit $100 more, it probably won't get you to 60.
|
5 years ago you could have said the exact same thing about LLMs just not being mathematically possible to get to the level they were 3 years ago.
We're handicapped right now not
with how much power it can use, or how much power it can process etc.
We're handicapped with:
It wants more power, it can use more power, but that power is just too expensive to be worth it commercially right now.