๐ฅ Really excited about this work!
Agents are too eager to act even when they should pause and ask the human, which can result in irreversible and harmful actions such as forwarding a confidential email to the whole company!
Check this out ๐
AI agents can do more than ever. But do they know when not to act?
Excited to present AgentAbstain at @SnorkelAI's Frontier Data Summit on Oct 8 and as a Spotlight at the WAB workshop at @COLM_conf on Oct 9.
@YiMaTweets I totally agree. but there's another layer to this. 'be better than us' assumes we can first define what 'good' looks like. Given how much humans disagree on right and wrong, I'm not sure that question has an answer either.
As someone who has been developing "AI" for ~20 years: Instead of "slow the pace", we need toย change the game.
Safety and capability don't have to be separate things; safety only makes things slower when it's tacked on, outside of the core technology. And the "pacing" rhetoric reifies a particular Frontier AI narrative, treating ethics (including safety) as an external check on technology, rather than part of what the technology is.
Done well, ethics -- including foresight on how risks and benefits interact -- can drive innovation instead of slowing it down. If you're in the AI alignment space, you can see something similar with techniques that turned LLMs into fluent chatbots (Instruction Tuning, RLHF): Our modern technology arose, in part, from safety-informed alignment practices. They defined what the technology is.
The need to reprioritize is hampered by the "race" framing, which strikes me as awkward against technology's real potential. It's not a race. There's no single track, no agreed finish line, no final point where someone wins.
I see a garden instead: Many things can be grown, and many can flourish at once. Here is a picture of part of my garden.
We are recruiting reviewers for the Evaluation of Interactive Agents workshop at NeurIPS Atlanta this year! If you are interested, please fill out the reviewer interest form :D
Congrats to @PardisZahraei and @iamjanvijay for the outstanding paper award at the KnowFM Workshop (Towards Knowledgeable Foundation Models, https://t.co/yG0GDIrisQ) at ACL 2026 for the following paper: https://t.co/IjjRrpePTg
@convai_uiuc@uiuc_nlp