July
@july
A common objection is that AI hallucinate - that it generates false information, misinterprets context, or confidently asserts fabrications. But what if this isn't really about AI at all, what if it is a reflection of who we are. If anything, what if they are our beautiful flaws that makes us deeply human. Also as Daniel Kahneman talks about in Type 1 thinking - ultimately cognitive biases shows that human thought itself is riddled with errors, misremembering, and distortions -- so if we train on that data, how would we not get the same type of behavior? What if instead of hallucinations, we call them happy accidents? (like Bob Ross)
8 replies
3 recasts
49 reactions
Zach
@zd
Yes! One of the biggest reframes I've had in the past year is to think of AI hallucination as a feature, not a bug. What could you build if you realize that all great ideas start as a hallucination until everyone starts to agree with them? When they do, what was once called a "hallucination" starts getting called "truth."
1 reply
0 recast
9 reactions
July
@july
I think about how DNA & RNA mutations in the genetic sequence is garbage most of the time (or even harmful at the cellular or structural level) -- it is a needed side effect as well for the species to evolve. Perhaps at a less grand level AI's "hallucinations" introduce novelty into the cognitive landscape. If AI were too rigid, too perfect, it wouldn’t be able to search for new solutions, just as a species without genetic mutations would eventually stagnate. Also makes me think that of how RL also mirrors biological evolution through trial and error, exploration using reward functions
2 replies
0 recast
3 reactions
