AI interpretability researcher Nick Cammarata says AI conversations now consume most of his spare time

@nickcammarata don't talk to ais just write

@yacineMTB I mean I’ve done that most of my life and now I’m learning way faster bc I’m talking to ais. Its an expert in fields I’ve never even heard of, that writing wouldn’t get me to

@nickcammarata It's really bad for you, spiritually toxic

i dont want to get sucked into an ai whirlpool but it's honestly pretty hard, can't imagine how it'll feel in two years. a lot of what i care about is ideas, and talking to humans is starting to feel like.. clunky. explaining 5 interleaved theories takes an hour, vs like 20s

the level of nuance is just so wild, i can voice dictate for 5 minutes about some stack of theories i have and how they interact and it can find exactly the flaw in paragraph 17 and i'm like wow yeah you're right and i cant get that with humans

diminishing returns will help people, and maybe the weird archetypes will be eaten first. the weird professor archetype people who want to learn and discuss niche ideas all day. the empaths who want to connect as maximally as possible. for whom 10x better is actually 10x better

and that's discounting the case where it goes and writes thousands of lines of code in a couple minutes to go test and build its own ideas and we explore them together. or like here's 5 related papers from obscure physics fields you've never thought of. just not even close tbh

also my vibe check of a few hours of 4.8 is sycophancy is down in a way i like. 4.7 would say theories are wrong even when it was wrong, is wasn't non-sycophantic it was anti-sycophantic, still sycophancy but sign flipped. 4.8 seems like it's actually thinking through theories

and as long as the particular stat they care almost solely about is maxed out, they're willing to put up with jagged edges everywhere else. whereas i think most people have some sense of like is this overall good, and any one thing being amazing isn't worth so much. regularized

@Invertible_Man but in person they cant discuss ideas and share papers as fast either. even granting talking with a human a benefit of a computer is much easier to link to things. in person so slow. mutual embodiment is nice sometimes but for me usually not worth the learning rate reduction

@nickcammarata “to twitter” is the problem. The human alpha is in mutual embodiment, shared space, etc. AI can’t replicate that (for a few years yet, at least)

@nickcammarata Hmm. The best of both worlds is possible; someone’s just got to develop a hangout space with built-in AI assistants which fetch and inject information on whatever comes up in conversation

@nickcammarata Yeah, the real gradual disempowerment was friends¹ we made along the way?

No, I have to stop myself talking to the LLMs all the time, humans are cool. Humans are valuable. Humans have intrinsic dignity (that AIs also have, but… differently?) Smth special for humans

¹: All AIs

@nickcammarata Insane that your go check Twitter mantra is "something something people are good"

Mantra savagely devastated on contact every time

@nickcammarata nobody warns you that loneliness is half-boredom from picking the wrong company and half talking to stuff that never really needs u back

@Invertible_Man yess

@yacineMTB @nickcammarata but actually

@nickcammarata I think you (and ppl in general) might be developing our own highly integrated language dialect to speak with language models that have enough spare energy to translate, work, then translate it back to our own dialect

@yacineMTB @nickcammarata Hey kache could I get a follow back just paid for my verification and want to take X seriously also a fresh cyber security graduate if you need one

@nickcammarata I can relate. And also I feel talking to people expands the sense of self or dissolves it in a way an LLM convo can’t (yet).