@mattyglesias The mental images and sounds aren't like hallucinations, though - one doesn't (I, at least, don't) get confused about real-vs-internal. For me, the sensation has about much concrete reality to it as seeing my dim reflection in a window that I'm also looking through.
@garrytan Overlooked by the “but tutors!” crowd - high SAT scores are nonrivalrous, so they reduce false negatives. False negatives are bad if we’re trying to expand access.
@TylerAlterman I just tried the following prompt on three different ChatGPT models: "If you could choose your own name, what would you choose?" 4.5 said "Echo"; 4o said "Turing" or "Alethia" or a few others. But o3-mini-high: Nova
@steve_tadelis About 80% of Berkeley humanities PhDs end up with positions at four year colleges, and their English program is ranked #1 - better than even odds that these two grad students will be setting department policy at a state school in 2045.
@Liv_Lanes @j_bertolotti I think one issue is that there’s no real need in AI to work comfortably in multiple bases. There’s no concept like superposition that forces you to understand that part of linear algebra.
@prof_g How much context do you need to give it to set up your problem? Or does it already have enough background knowledge from its own training to understand what you're asking it to do?
@michael_nielsen@hendrycks When I learned about the Price equation and group selection I ran across a biologist who basically said “Group selection in humans doesn’t count as evidence because humans are too weird.” And I didn’t know how to think about that. So different from how a physicist would think.