Phillip Carter

phillipcarter.dev

25 December 2024 06:07

Esp. since coming from the distributed tracing world where you simply deal with the reality of sampled data, incorporate sampling rates into calculations, and aim for a good MoE relative to your data, calculations, and granularity

Phillip Carter

phillipcarter.dev

25 December 2024 06:07

It kinda blew my mind that evals on sampled data didn't incorporate things like confidence intervals a little while back. I just assumed people were doing that!

Yoav Goldberg

yoavgo.bsky.social

01 December 2024 01:59

are there works (in RL? in planning?) that characterize different long-horizon tasks in terms of their hardness? for example, how reactive can you be? how often if at all do you need to backtrack or change strategies? how delayed is the feedback? etc. is there a taxonomy or theory of these?

Jeremy Howard

howard.fm

30 November 2024 18:29

hardmaru got banned from bsky because he didn't believe someone who claimed that there's a sentence that, if you write it here, you get insta-banned. So he tried it, to prove it's not true. Turns out the claim was right. It's a ban that happens even with all moderation filters disabled.

Ramon Astudillo

ramon-astudillo.bsky.social

29 November 2024 22:11

Everything is Markov (order N) if you have enough compute 😛