jessica dai
@jessica.bsky.social
510 followers 110 following 78 posts
go bears!!! jessicad.ai kernelmag.io
Posts Media Videos Starter Packs
Pinned
jessica.bsky.social
individual reporting for post-deployment evals — a little manifesto (& new preprints!)

tldr: end users have unique insights about how deployed systems are failing; we should figure out how to translate their experiences into formal evaluations of those systems.
jessica.bsky.social
so close! that's standard error ❤️
jessica.bsky.social
i'm still pissed about this like the difference is literally too small to have been distinguishable with swe bench (500 samples) lmaoooo
jessica.bsky.social
hey wasn't this the same company that made a beautiful shiny "research" post about how AI evals should include error bars or something like that. or did they decide the CLT didn't apply here
jessica.bsky.social
I will be at ICML in a few weeks & would love to chat about how to make this real - I am a critic at heart and also hate self-promo so that’s how you know I really believe in this 🥲
jessica.bsky.social
various ways to read more 😀

blog post- argmin.net/p/individual...
position paper- arxiv.org/abs/2506.18133
fairness-oriented instantiation- arxiv.org/abs/2502.08166

& many thanks to brilliant collaborators
@rajiinio.bsky.social @irenetrampoline.bsky.social @beenwrekt.bsky.social & paula gradu !!
argmin.net
jessica.bsky.social
lots of other stuff I won’t get into rn (e.g., I think this is a prereq to any serious attempt at “democratic” AI!), and there’s also a ton of open research questions (stats, econ/ml, empirical methods, hci, …)
jessica.bsky.social
the core concept is individual reporting as a means to build collective knowledge. if one person has a bad experience, that doesn’t necessarily mean that there’s something wrong with the system — but if lots of people start reporting similar things, maybe we should pay attention.
jessica.bsky.social
we’ve already seen this informally with the chatgpt sycophancy debacle — a few days of twitter virality resulted in action and statements from openai — but what other, subtler, patterns are happening? what could we discover if we had better ways to listen to the public?
jessica.bsky.social
individual reporting for post-deployment evals — a little manifesto (& new preprints!)

tldr: end users have unique insights about how deployed systems are failing; we should figure out how to translate their experiences into formal evaluations of those systems.
jessica.bsky.social
right but one would hope that the date of doom _does_ get further away as safety research improves

bsky.app/profile/jess...
jessica.bsky.social
like is it that the field has been ineffective (studied the wrong problems, advocated for the wrong positions, etc) or is it that every step of safety progress has been matched by 2 steps of capabilities progress (in which case, what are the best examples of safety work concretely reducing harm?)
jessica.bsky.social
where are the bullshit "x% of experts believe" polls when you need them lol
jessica.bsky.social
well probably, but i wanna know how folks who do believe in that happening think about the field
jessica.bsky.social
or is it a secret third thing idk. scared to ask this on Real Twitter but genuinely curious how people think about the role of this field
jessica.bsky.social
like is it that the field has been ineffective (studied the wrong problems, advocated for the wrong positions, etc) or is it that every step of safety progress has been matched by 2 steps of capabilities progress (in which case, what are the best examples of safety work concretely reducing harm?)
jessica.bsky.social
perhaps this is a stupid question but given that ai safety has been a pretty vibrant (+ well funded) field for the last 5-10 years... how should we be thinking about the concern that (ai) catastrophe still is, allegedly, imminent
jessica.bsky.social
back on bluesky to be mean about ai discourse
jessica.bsky.social
im ngl i think this kinda just means u are stupid
jessica.bsky.social
i don't work well under deadline pressure but i also don't work well without it. therefore,
jessica.bsky.social
... didn't we just talk about this ...