Post #2880040

@denschub@mastodon.schub.social thanks. This was a really nice way to explain the “mostly plausible, sometimes accidentally correct” to my kids during breakfast. I like your tennis ball analogy too. I think you mention in it some of what makes us different from (current) AI. We have learned from experience and evolution that you sometimes need to be correct. It matters. It can ultimately be a matter of life and death for us. Of bruises, cuts, hunger, cold, social disapproval. So we can value being correct. We’ve can “feel” it in some sense. We’ve got a stake in the game. Current AI training doesn’t seem to feature that. Or give an ai a universe to reckon with correctness in. I suspect that’s surmountable. Future training approaches might be able to embody that kind of learning process. Reckoning with reality. I thought maybe the chain of thought and reinforcement learning used in more recent LLM training might start to do that, but it doesn’t seem to have.