Tim Delille

content ML ranking @ Snap 👻

[Tim Delille @TimothyDelille]

@marclou did you change the core product ?

[Tim Delille @TimothyDelille]

agreed, living in a european city means all your needs are taken care of, no need to own anything, you take the subway to the gym, walk to the grocery store, your whole life is catered to. the downside is… your whole life is catered to. you only do things that are accessible by subway. you spend your time waiting for public transport. you go to wework. variance is extremely small, everyone does the same things. you will never meet a dude going on random side quests, riding motorbikes in the desert just because (« why would you even want to do that! ») it’s also why europeans tend to aggregate in NY / SF. im french and living in the US for 6 years btw (sf -> ny -> la)

[Tim Delille @TimothyDelille]

@fchollet regarding 4. what data are you looking at specifically ?

[Tim Delille @TimothyDelille]

super cringe to see myself on video but I have a better idea what to improve now

[Tim Delille @TimothyDelille]

back in topanga this week-end (video is 3s long bc it’s from a live photo lol)

[Tim Delille @TimothyDelille]

@jose_goncalves_ average sorbonne student

[Tim Delille @TimothyDelille]

@ChristosTzamos @yechan_ai is it encoded in another set of weights (which only executes when the model triggers “fast decoding” ?) this correctness guarantee is really amazing. does this also mean that you can execute arbitrarily long algorithms (like multiplying very long numbers), bypassing ctx length?

[Tim Delille @TimothyDelille]

@ChristosTzamos @yechan_ai is it encoded in another set of weights (which only executes when the model triggers “fast decoding” ?) this correctness guarantee is really amazing. does this also mean that you can execute arbitrarily long algorithms (like multiplying very long numbers), bypassing ctx length ?

[Tim Delille @TimothyDelille]

@bryan_johnson whoop shows me +3% recovery impact for ag1. confounding or actual impact ?

[Tim Delille @TimothyDelille]

@BetterCallMedhi c’est plus facile de faire du revenue en vendant du slop a des grands groupes / gov qui n’ont aucun standard plutôt que d’essayer de pousser la frontiere et dev le prochain palantir. j’ai l’impression que ce mispricing se corrige dans le tertiaire

[Tim Delille @TimothyDelille]

@nathanbenaich @culturengine @alisabets @zhoubolei @QuanquanGu @petrenko_ai @tweetingnonstop @ben_kasper @_samirism I’ll be around!

[Tim Delille @TimothyDelille]

the flash attention paper reports achieving the first non-random performance on Path-X and Path-256. Since they’re so difficult, has any modern LLM been evaluated on these tasks ?

[Tim Delille @TimothyDelille]

@martin_casado great job this is incredible

[Tim Delille @TimothyDelille]

Falcon 9 spotted in Santa Monica

[Tim Delille @TimothyDelille]

@martin_casado cool stuff!

[Tim Delille @TimothyDelille]

Whoop didn’t let me customize journal entries so I built my own

been tracking supplements and workouts for 3+ years. Going to start tracking behaviors too now

[Tim Delille @TimothyDelille]

@karpathy insightful, thank you
“How it went well” will need to come from a world model / tutor (in Claude’s case: an engineer) instead of the agent itself, won’t it ?

[Tim Delille @TimothyDelille]

https://t.co/kARy8Aiu2g

[Tim Delille @TimothyDelille]

omg I still can't get this CarRacing agent to learn... Simple Deep Q Network baseline with same parameters as the Atari paper: 1M episode replay buffer size, linearly decreasing epsilon from 1 to 0.1, clipping rewards between -1 and 1 (code in the reply)