How are LLMs intelligent? Point me to the intelligence?
Literally it’s a black box of statistics. Tokens in, Next token probabilities out
Literally it’s a black box of statistics. Tokens in, Next token probabilities out
You literally just repeated the exact same mistake, albeit in different words. This is tantamount to speaking of airplanes: "It's not flight, it's just a bunch of fluid mechanics". You're focusing on the internal implementation details, not the expressed behavior.
I don't know if people do this out of some sort of "biological chauvinism" or what, but it strikes me as rather odd.
I can (to pick one random example out of many) ask ChatGPT "Who is the current Prime Minister of Japan and what is the largest prime number smaller than their age?" and get a correct answer. To say that a system that can pull that feat off isn't intelligent boggles the imagination IMO. Token prediction? Who cares. Hallucinations? Humans hallucinate also. Perhaps less frequently, but all that says is that we aren't "all the way there" yet with AI.
But it's not a human intelligence, and it's creepy when its creators dress it up to pretend it is. It's an extremely superficial imitation of a human, and it doesn't behave like a human. It's like when a human furry claims to really be a dog or an otherkin.
is probably the best summary I've read on the topic
Literally they're just meat. Sensory inputs in, electrical signals out.
I don't think that but you're right that I was partially countering a pro-LLM argument that was not there.
It was a bit of a knee jerk reaction due to the number of people overstating current LLM's capabilities here on HN.
In-context learning (Few Shot Prompting) is absolutely demonstrable by LLMs since GPT 3.5 at least.
If you have a challenge (question) and the subject reduced the unknown qualities of that challenge in a deterministic way then some degree of intelligence has been achieved.
You just described an extremely accurate model of people.