Remix.run Logo
▲ threatripper 10 hours ago

Maybe we don't need real intelligence or consciousness to be effective.

▲pitched 27 minutes ago | parent | next [-]

Consciousness is a trick that our brains play on themselves because it tends to make their humans more successful. Like opposable thumbs, if it didn’t make us more effective at some point in history, it wouldn’t be there. So it must make us more effective in some way. There’s no reason to think that trick couldn’t be reproduced with a system prompt though.

▲pseidemann 8 hours ago | parent | prev [-]

Nobody knows what intelligence actually is. Psychology and neuroscience have no answer, and especially not computer science.

▲whattheheckheck 6 hours ago | parent | next [-]

The most important behavior is doing the right thing at the right time to satisfy the four Fs

▲morpheos137 6 hours ago | parent | prev [-]

Intelligence like consciousness is a word to describe patterns of processes tending to optimize toward some goal set under the constraints of environment and with sensory feedback. Humans goal is to reproduce. Llms goal is to predict best next token. Human feedback is sensory mediated. Llm feedback is prompting on outputs including from tools such as lean. If RH was provable it is not surprising either such system could prove it in principle. If a sound proof required an intricate path or vast contextual memory to assemble it is not surprising a machine proved it before a human. However I don't think this is generalizable. I don't think intelligence is a scalar in that the proof shows the machine is more intelligent than the human. Rather it is qualitative. The machine was structured so that the path was attainable to it. This is intelligence in the particular case but not general intelligence if such a thing exists. Now what would point more toward general intelligence is a machine that can compress proofs or truths better than humans. If we start seeing short proofs of FLT or RH then that will be interesting. I am skeptical with the current technology. The tendency of LLMs seems to be to explode complexity not condense things to essential elements or minimal description lengths.