3 ms·
we really could save a lot of time if everyone just actually remembered that LLMs are designed to generate plausible output. it doesn't "know" anything, it doe
by bananapub 3y ago
we really could save a lot of time if everyone just actually remembered that LLMs are designed to generate plausible output. it doesn't "know" anything, it doesn't "reason", it doesn't "think", it doesn't have opinions, it just takes a text prefix/seed/whatever and generates output, based on training of things humans actually produced.
everything becomes pretty clear then.
should you trust the answers it gave you? obviously not, it just generated some crap.
can it be useful for brainstorming? sure, you an consider the output and take it / be inspired by it / etc
can you use it to generate useful code? of course, as long as you read it and it's reviewed by tooling and humans to see that it does what you thought it did.
should you replace your employees with it? obviously not, a text suffix generator isn't able to deal with complex problems nor be responsible for decisions.
etc etc
- danielbln 3y agoOnce again we parade the "it just predicts the next token" around as if it was some deep insight. The magic of it all is how it does it, which we don't quite now. There might be some world model involved, there might not be, who knows. All we know is that "predicting the next token" is a simple objective that requires a lot of complexity inside the black box to work as well as it does.