AI Gets WEIRD: LLMs learn reasoning solely by their own internal "sense of confidence"

Self reinforcement through confidence is a great way to entrench all existing biases...

Damn it feels like these videos come out faster than I can keep up now, you’re really on the pulse Wes. But good job for staying on top of it because it helps everyone else do the same. Just crazy to see how quickly these models are developing (now themselves) lol.

sometimes only 10% of people have read the latest papers on a subject while the other 90% still have an outdated information set. so 90% of people would give a confident and incorrect answer based on widespread old understanding.

I gave this a thumbs up on the sole reason it was a Wes Roth video under 20 minutes!

This is quite unsettling when you think about it. Right now, AI models don’t possess an inner world—no true introspection or consciousness. But we may be on the verge of teaching them to simulate it so convincingly that we won’t be able to tell the difference. And once we reach that point, there’s no telling what AI might become. The scariest part, I think, is this: we don’t even know what it really means to be human. For all we know, we might just be incredibly good at faking it too—playing out complex behaviors shaped by evolution, without any deeper awareness. The more we understand about AI, the more we’re forced to confront the mystery of ourselves. What will it mean for us if we can build a machine that looks, thinks, and feels just like us? And what if it turns out that the difference between real and fake was never as clear as we thought?

It was a matter of time for us to realise that those models develop internal intuition (which, in humans, is feelings-based,) and we can exploit this fact by giving *them* the means *to reflect* on those "feelings"!

It would also be good if they use the same low confidence metrics for the models to say "I don't know"

I'm 100% confident we'll get AGI before 2030

What could possibly go right?

I've never envy anyone anything apart from wisdom. I wish to have half of your knowledge...