Comment by BigHatLogan | Hacker News Reader

Comment by BigHatLogan | Hacker News Reader

Parent Full thread

I may be overanalyzing this entire "phenomenon", but it serves to remind me that our own understanding of ourselves is incredibly limited. There seem to be a number of processes happening under the surface that are completely unbeknownst to us. It made me wonder what else I have missed or misinterpreted in other aspects of my life. How often have I been utterly convinced of something? How often am I so sure that what I'm feeling or thinking is the "correct" way to feel or think about a problem? It amazes me that something as simple as a good night's sleep (the phrase, "Sleep on it", comes to mind) can completely change my perception of an event in the course of a few hours. Whereas in the prior evening it felt as if my world was caving in on me, in the morning things became more manageable. The event itself did not change, only my perception. I liken the human mind to an iceberg. We're cognizant of such a small part of it while the rest remains under the surface and out of our sight.

We are repeatedly finding ways to hack general human perception the same way we can hack AI neural network perceptions. It's all very well pointing and laughing at the way we can construct pathological data: 'look, here's two pictures of a kitty, but the AI thinks the second one is an emu and can't be convinced otherwise!'

And then it turns out we're very much in the same boat.

100%. This is going to be one of the big struggles with AI, especially in an interaction like driving. By and large, I don't think humans realize how much they misperceive things, and make little mistakes. How many times have you momentarily seen a face in a bush when it's dark, misheard a sound as your name, thought that you saw a dear old friend across the room, failed to figure out where the lanes are on the road?

We're really good at adjusting to these little mistakes, covering them up, because they're insignificant. But we do them constantly...and then forget they happened. I think that most of us overestimate humans' abilities in really mundane matters. We do well, not because we don't make mistakes, but because we recover from them quickly. (And when they're done in a social context, we have rules of politeness that usually allow us to forgive others' tiny mistakes.)

That high evaluation of our own cognitive ability means that an AI that's "less" than that, which makes "silly" mistakes, isn't good enough. Even if an AI makes far fewer mistakes than a human would at some task, any mistake is evidence that it's not good enough, because "no human would have made that mistake". Our meta-perception of our own perception is so clouded by (natural) hubris that we can't even realistically imagine a typical human's performance for comparison.

This, to me, was a great revelation about the human mind when I studied AI: perception is hard, and I'm actually not that good at it. I just muddle through, with some kind of metaprocess smoothing over all these silly little bumps. (And there's an interesting link here, too, to old meditative practices.)

Reply on news.ycombinator.com