‼️Skibidi for Machines! :) . Developing language 🔠 between humans🧒 and machines🤖 has long been a dream - the language that will help us expand what we know so that we can communicate with machines better, and create machines better align with us. With @johnhewtt's amazing.

RT @ActInterp: The one and only @_beenkim on Agentic Interpretability and Neologism: What LLMs Can Offer Us!

Join millions who have switched to Grok.

RT @johnhewtt: Come chat with me at our ICML poster about interpretability as a communication problem, and the need to derive new words for….

RT @ActInterp: Join us at 09:10 for Been’s talk on Agentic Interpretability and Neologism: what LLMs can offer us.

RT @WiMLworkshop: Talk panel“The Unsolved Problems of AI: What We’re Still Getting Wrong” 🤦‍♀️.Featuring Ana Lucic, Been Kim, Furong Huang,….

RT @NithishKannen: Say hi 👋to @ziwphd ,@MeeraHahn, @_beenkim, Wenjun Zeng and Rich Galt at #ICML2025 if you're interested in learning more….

Going to #ICML2025 tomorrow - Sat morning! If you are graduating this year with a PhD, and you are interested in Google DeepMind, send me a message. Where I will be: . 0. Wed 1pm @WiMLworkshop Mentoring Table and 2pm panel discussion. 1. Wed 4:30pm East Exhibition Hall A-B.

In this paper, we layout some core properties of agentic interpretability (i.e., proactive assistance, multi-turn interaction, mutual mental model), provide some real and some imagined examples of agentic interpretability (e.g., sec 2.2.3 `open-model surgery’ idea where we can.

A method is agentic interpretability if it `proactively assists human understanding in a multi-turn interactive process by developing and leveraging a mental model of the user which in turn enables humans to develop better mental models of the LLM’. In other words, enable.

By keeping up with the evolution of machines, we believe that this can help increasing human empowerment and helping reduce bad outcomes like gradual disempowerment (.

The idea of agentic interpretability stems from the observation that AI systems are becoming increasingly adept at communicating with us, verbalizing their thoughts, and providing explanations. So why not enable them to help us understand them?.

We (@_beenkim @johnhewtt @NeelNanda5 Noah Fiedel Oyvind Tafjord) propose a research direction called 🤖agentic interpretability: we can and should ask and help AI systems to build mental models of us which will help us to build mental models of the LLMs.

Excited to give the opening keynote at bidirectional human AI alignment workshop time at 9am! . And adding a picture with @ylecun after being stopped every second by folks who wanted to take pictures with him (we were just walking out of ICLR board meeting). The peer pressure was

RT @iclr_conf: Huge thanks to the #ICLR2025 Organizing Committee (including many who couldn't make it to the conference) 👏👏👏 .

See you in Singapore! #ICLR2025.

RT @PrincetonAInews: As AI becomes increasingly complex, new language is needed to improve communication between humans and machines, @Goog….

Now the talk is uploaded here:

RT @ziwphd: Thrilled and grateful for the invitation from @HDRUNDP to share our research on proactive AI. Such a valuable opportunity to d….

♟️♟️Now our work on teaching superhuman chess strategies to grandmasters (one of whom @DGukesh who became the latest and the youngest world chess champion) is published on PNAS! 🎉🎉. Yes, we can transfer machine knowledge to humans to push the frontier of human. knowledge.

On the gorgeous Princeton campus, I'm excited to give a distinguished lecture tomorrow! This nice poster they gifted me shows my entire wardrobe—I’ll be wearing that jacket. Again.😅😅

RT @JeffDean: 🥁Introducing Gemini 2.5, our most intelligent model with impressive capabilities in advanced reasoning and coding. Now integ….