jessica dai
@jessicadai_
phd student @berkeley_ai !? also editorial @reboot_hq @kernel_magazine (she/her)
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.
I’m rebelling against ChatGPT emdash hegemony — we must reclaim the emdash — it does not belong to the LLMs — it belongs to us!
kernel launches in two days! help me get these magazines out of my house!
just had 500 magazines delivered to my door :)
postering today with @paula_gradu at 4:30 (East A-B E-1202) 😀 come say hi
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.
Being a grad student is peak human existence. The end goal of all political and technological progress should be allowing everyone to be a grad student forever
There's been an idea floating around policy spaces for some time now (eg. UN AI Advisory report, California Frontier AI report, etc) around the need for "AI monitoring", "AI incidents", "AI adverse reporting" etc. Jess is now seriously thinking about how to operationalize this:
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.
we're launching a new issue of @kernel_magazine in two weeks!! join us for a launch party in SF @ Gray Area on July 17!
What more could we understand about the fractal, “jagged” edges of AI system deployments if we had better ways to listen to the people who interact with them? What a joy to work w @jessicadai_ using individual experiences to inform AI evaluation (blog/ICML/arXiv links 👇)
Newest @reboot_hq 🎙️post: @jessicadai_ and I discuss forecasting, and how people present unhelpful narratives about the future (mostly by picking on AI 2027, sorry guys) Why we should view the future as constructed, not predicted
RLHF fine-tunes to a “mythical user” via aggregated feedback—but what if that user represents no one? Excited to share a new paper with @paulgoelz and @KunheYang “Distortion of AI Alignment: Does Preference Optimization Optimize for Preferences?” #AIAlignment #PluralisticAI #LLMs
In LLM land, a slow model is annoying. In robotics, a slow model can be disastrous! Visible pauses at best, dangerously jerky motions at worst. But large VLAs are slow by nature. What can we do about this? An in-depth 🧵:
Why RLHF stopped us from getting r1/o1 sooner and why we elected Trump/Biden/[Politician You Dislike]: We can use elections 🇺🇸🇺🇸🇺🇸 to understand why RLHF naturally suppresses reasoning (even if you fix the whole "RLHF isn't really RL" thing)