Long Ouyang (@longouyang) profile picture

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

Joined July 2008
Photo by Long Ouyang (@longouyang)Photo by Long Ouyang (@longouyang)Photo by Long Ouyang (@longouyang)Photo by Long Ouyang (@longouyang)Photo by Long Ouyang (@longouyang)Photo by Long Ouyang (@longouyang)Photo by Long Ouyang (@longouyang)Photo by Long Ouyang (@longouyang)Photo by Long Ouyang (@longouyang)Photo by 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)
2
5
1,002
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)
1
1
3
925
way to track whether the thing that was approved was the raw code or the summary (at some particular level of granularity). (3/3)
2
447
just played around a little with tinker from thinking machines.. definitely makes sft / rl a lot more accessible
1
10
1,577
context length anxiety ~= range anxiety
3
1,040
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/
2
1
9
1,426
hunched up in a closet? sleeping in a bed? an interesting design problem
3
883
gpt-5's final paragraphs "if you want, i can make a table about x versus y versus z" are annoying 80% of the time, but like 20% of the time they're super useful - help me figure out something i didnt know i didnt know. but by definition i couldnt have known to ask for this. 1/
1
4
913
in principle this is something i could collect a dataset for and eventually fine-tune on, but the friction is high enough that realistically i won't get to it. would be nice if there's some in-product way to do it; still early days for personalization i guess.
3
646
genie 3 (incredible, of course) feels like something that should have come out of meta when it was more focused on VR.
5
626
proposal for new term - structurable data: unstructured data that can be turned into structured with the right llm invocations
4
708
with models like deep research, there's an opportunity for publishers / rights-holders to acquire new customers / distribution mechanisms, e.g., it'd be pretty valuable to buy a deep research connector subscription to jstor, property/finance databases, (1/4)
2
2
707
query). it also feels like there's some room for innovation in pricing / how payment is made (e.g., charging per-article-searched)
449
old newspaper archives, etc. this feels more viable than, say, microtransactions for news articles - the end user is "exercising" a lot more of the content (through the process of a machine sifting through lots of it and bubbling up the parts most relevant to a user (2/4)
364
cool feature of macos26 - integrated language model support that user applications can generally call. this would be a nice wedge for homelab inference powered by a mac mini or even homepod.
1
756
random nice use case for LLMs - asking questions about tv shows / movies you havent seen and requesting no spoilers
4
565
cool tidbit from the In Good Company podcast interview of jennifer scanlon, ceo of a renowned product safety testing company called Underwriters Laboratories (UL). they've started working on standards for AI products. UL started in the 1890s certifying electrical equipment (1/2)
1
1
610
it seems that a lot of their raison d'etre is insurance, broadly construed. e.g., when they first started, they were backed by fire insurance underwriters (at the time, mitigating fire hazards was more nascent). neat that they're also looking at ai in its early innings too
452
in general, pretraining data curation often removes ads or boilerplate, but when agents browse or do research, they can get inundated with the stuff - it feels like they can get distracted / fooled by fluff content, possibly due to differences in train/test distributions
1
4
17
1,537
one obstacle to (many, but not all) long horizon tasks is that human preferences may be very tacit (and in some cases, might not exist at all until we're faced with a choice). 1/5
3
4
905
in principle, we could supervise models to ask for this sort of information during training, but that adds a lot of time. 4/5
1
639
and the modelling assumption that the human preferences are stable and fixed ahead of time feels murkier here (i can imagine that preferences are actually *created* by the set of available options, à la some behavioral econ work) 5/5
1
573