niplav.site

No description provided.

von Neumann: I came up with this new system that generalizes probability theory to consider convex sets instead of point estimates. I think that I could use this to prove regret bounds…

⤋ Read More

you might be thinking: “aha! so I should vote in elections, since even though under do()-calculus, the decision has a miniscule impact, there are many agents that are logically correlated with me, which means my influence is much higher!” A tiny problem is that the number of agents that are logically correlated because they base their decisions on logical correlation is, ah, not that big…

⤋ Read More

“…and we can probably plan for middle of March thi…ah, hello! Mister Yudkowsky! I hope you found—” “Rename your company to ClosedAI.” “Uh…what?” “Rename your company to ClosedAI. Stop your biggest training runs. Pivot to mechanistic interpretability. Shut it all down.” “………Selfie?”

⤋ Read More

at higher levels of intelligence, maintaining coherence becomes more difficult since your action space widens and having high coherence might be NP-hard. so the better metric is coherence divided by size of action space

⤋ Read More

a problem with minmaxing is that the crisper your model of the world becomes, the more conservative you become—but intuitively it should be the other way around!

⤋ Read More

in emily in paris Y is mad because X slept with Ys boyfriend. and what does my brain say? ‘why don’t X and Ys boyfriend pay her the cheerful price to forget the thing?’ damn economists

⤋ Read More

I want there to be a norm that it is good for people to first state them noticing their status-guided motivations and then proceding to answer on the object level when askedabout things.

⤋ Read More

here’s a question: when do NNs generalize, and how hard? as in adding two specific numbers together vs. n-digit integer addition vs. addition in general vs. simple arithmetical operations

⤋ Read More