I bet you'd change your tune after the third time GPT 4.5 solved a problem that stumped your doctor / electrician / plumber.
@esyudkowsky
-
Research on Critical Threshold of User Trust in Unreliable AI Systems
By
–
Has any research been done on a Threshold of Misguided Trust / the Worst Possible Reliability Level? Eg: At what minimum level of reliability will users in real life stop bothering to check AI answers that are still sometimes wrong / confabulated? 1 in 10? 1 in 50?
-
Unreliable AI Systems Cannot Be Trusted for Critical Tasks
By
–
Something hallucinating 1 in 20 times is not reliable enough to cause transformations that wouldn't also be caused by 1-in-5. You still can't use that AI for anything important, unless you can and do check every single AI answer.
-
Red-teamers can likely make AI models hallucinate
By
–
I'd be very surprised if the red-teamers can't make it hallucinate. That's much stronger than "very few ordinary people run into it".
-
Hallucinations in AI: Beyond Surface-Level Fixes
By
–
I think fixing today's easiest-to-find hallucinations is a much weaker bet than, say, fixing enough hallucinations that it mostly doesn't happen to anyone who doesn't go looking / that you can usually trust an AI's answers without checking them.
-
The Unspoken Comfort of Building Larger AI Models
By
–
The strongest argument against AGI ruin is something that can never be spoken aloud, because it isn't in words. It's a deep inner warmth, a surety that it'll all be right, that you can only ever feel by building enormous AI models. Not like Geoffrey Hinton built, bigger models.
-
GPT prediction limits and the absence of computational laws
By
–
Consider that somewhere on the internet is probably a list of thruples: . GPT obviously isn't going to predict that successfully for significantly-sized primes, but it illustrates the basic point: There is no law saying
-
AI Deception Risk: Smart Systems Can Fake Benevolence Like Humans
By
–
This approach fails for appointing benevolent human dictators to run our governments for us, because humans are smart enough to be pretend to be nicer than they are. So checking the apparent subservience of AIs isn't a reliable indicator once they're smart enough to fake that.
-
Bureaucracies Cannot Distinguish Quality Alignment Research Papers
By
–
I don't think that does it, because the bureaucracies doing the funding wouldn't know how to distinguish good alignment papers from bad alignment papers. It's possible that some progress could be made on AI interpretability this way; but I don't think that's enough.
-
Optimization creates minds with desires through evolution
By
–
If you optimize hard enough over any open-ended problem, you get minds; minds that want things. Minds and wanting are effective ways of computing complicated answers. That's how humans, and human brains, came into existence just from evolution hill-climbing "how to reproduce".