Agents are a joke until they’re not.

@teortaxesTex I skimmed it bc working but is there a significant theoretical contribution beyond the current lit? Seems like this paper with some more empirical findings / "laws"

@DimitrisPapail @nayoung_nylee @jackcai1206 @A_v_i__S @Kangwook_Lee Would it be fair to say this is an empirical / results-driven paper? Curious if you considered/attempted other proxies for difficulty, such as entropy / logprobs or something else entirely

@ocolegro Ah sorry I meant essentially "when will I be as effective (using Cursor + agi)". There's certainly far less Rust code in the pre-training corpus, but presumably solving Python can only come shortly before solving Rust, no?

@cursor_ai Don't see the ability to configure custom MCP?

@SZiteng @ananthbshankar @JonathanBerant @jacobeisenstein @nagpalchirag @FlavienProst @aradsinha @th33rtha Have been waiting so long for a paper like this! "Learning" to take advantage of TTC. should be an entire sub-discipline. Congrats on the huge number of forthcoming citations :)

@dexhorthy My main app for deep research these days is asking it to check the docs for me lol

@fmerian @nbrempel 🔥🔥🔥

@FelixGerlach 👺

@fugitive_druid @alicemazzy "By the sweat of your brow will you live" is a blessing not a curse

@cross_entrippy Incentives

@cross_entrippy It’s truly a privilege

@DimitrisPapail If I train on X data points from Craftax-Classic with filtered BC and PPO, and then test on Craftax-Full, which do you think will do better? Can probably test this!

@TalentedMrReply Transformers are great at RL tasks. What's wrong with tokens?

@ethansteininger 🫡

@corbtt Our day one vision 😉

@EliottLee LFG

@dexhorthy @CamelAIOrg @humanlayer_dev Woah

@not_nang Always have