Long Ouyang

human + machine learning @OpenAI. previously: cognitive scientist @Stanford, baby

[Long Ouyang @longouyang]

@usr_bin_roygbiv @a__tomala i have both. the 360 feels much worse to type on, even if used wired. i think it might be a combination of the keycap texture, switches, and subtlety of how the hands rest

[Long Ouyang @longouyang]

@unixpickle @Miles_Brundage To me alignment seems more interesting? Like the steps for capabilities improvement seem more predictable, whereas questions like what the models are actually doing or how to constrain systems much smarter than us feel more intriguing

[Long Ouyang @longouyang]

@giffmana @austinc3301 It could be that fast mode is in a different cluster/cloud where you cant transfer to or it’s infeasible to

[Long Ouyang @longouyang]

way to track whether the thing that was approved was the raw code or the summary (at some particular level of granularity). (3/3)

[Long Ouyang @longouyang]

speculative - comprehension chains for code (cf. cold chains in logistics). the future of software engineering might be more like systematically approving ai-written code (sort of analogous to how civil engineers officially certify designs). current agent harnesses dont (1/3)

[Long Ouyang @longouyang]

make it that easy to review diffs so there's some gap here for an interface that makes it easy to *understand* large diffs (and modify them). the IDEs seem better here though there's still room for improvement, e.g., it'd be useful to have some summarization feature and a (2/3)

[Long Ouyang @longouyang]

@thegautamkamath I like AI for science but it feels like it does little directly for folks who want help making rent / putting food on the table. And the current environment doesnt seem all that pro-science either (e.g., people being anti-Terence Tao when he was having funding issues)

[Long Ouyang @longouyang]

@difficultyang fwiw, i’ve noticed surprisingly bad hallucinations with (non-thinking) 5.2 for world knowledge questions (e.g., it thought Fil-C was a parody version of C based on Filipino)

[Long Ouyang @longouyang]

just played around a little with tinker from thinking machines.. definitely makes sft / rl a lot more accessible

[Long Ouyang @longouyang]

@nearcyan yeah i was expecting to hate it given all the chatter but do like it

[Long Ouyang @longouyang]

@gstsdn noob q - why dont the AIs get included in the firm size?

[Long Ouyang @longouyang]

@bayeslord there’s an important genre of code he writes which is load-bearing pedagogical code, like “technology X in 100 lines” where every line matters a ton for the reader (rather than the interpreter/compiler). these sorts of cases naturally fare worse with codex et al

[Long Ouyang @longouyang]

@LauraDeming hiring an editor or thinking partner off of a freelancing site could be valuable - that would have the extra benefit of sharpening the specific pieces you’re working on

[Long Ouyang @longouyang]

@zachtratar @typewriters For the big labs it’s not clear but I think for specific use cases like lawyers making arguments and providing citations or summarizing appointment transcriptions in medicine, there’s clear signals to measure and insure against

[Long Ouyang @longouyang]

@stevenydc you probably saw this but in case not - Don Wilson is working on the compute futures part (as far as i understand it) though admittedly it’s very early days https://t.co/vRZlAyF40H

(see also a recent episode of Odd Lots)

[Long Ouyang @longouyang]

@_lyraaaa_ do any models do ok at this?

[Long Ouyang @longouyang]

context length anxiety ~= range anxiety

[Long Ouyang @longouyang]

@unixpickle shumway works surprisingly ok

[Long Ouyang @longouyang]

hunched up in a closet? sleeping in a bed?

an interesting design problem

[Long Ouyang @longouyang]

i can't really imagine that many tasks i'd want a humanoid robot to do - emptying dishes and laundry, mayyybe sweeping some leaves, but i'd guess that it'd be mostly idle to start. so where would it be stored? standing dracula-like in an upright box? 1/