Hacker :vf: on Nostr: Process text to do what exactly? Computers are good calculators. You can do ...

Process text to do what exactly? Computers are good calculators. You can do calculations to check if a point in euclidean space is bounded within a known volume. Or do calculations to obtain the probability of something happening.

Statistics have a margin of error. Errors can be compounded. A concatenation of trivial snippets based on the probability of them forming a sequence is error compounding, propagation of uncertainty. You end up with a pile of crap. The only thing you can do to mitigate that is using an enormous number of samples to lower the error up to a limit. That's inefficient.

But LLMs are even worse, because they work at token level. Error propagation is greater, so you need even more samples to keep gibberish from happening too often.