Michael Bernstein

It was standing room only at the kick-off for our research series on continual learning. Thank you to @NikzadAfshin (@across_ai ) @sarahookr (@adaption_ai) and @mralbertchun (AI Circle) for hosting! @oshaikh13 shared his research on human grounding in continual learning. It was so cool to be reminded of the old Apple Knowledge Navigator and how close we are to it and yet how far we still are :) how much easier some questions have gotten and how some remain so hard. Omar, you reminded me of my PhD defense where at some point I annoyed Maneesh so much he said: you can't keep saying "depends on the user context" in response to every question 😅 youtu.be/umJsITGzXd0?si… Stay tuned for the next meetup next month and check out Omar's research with @msbernst and @Diyi_Yang : •⁠ ⁠Creating General User Models from Computer Use (arxiv.org/abs/2505.10831): an architecture for a model that learns about you by observing any interaction with your computer, building confidence-weighted propositions about preferences and intent. •⁠ ⁠Learning Next Action Predictors from Human-Computer Interaction (arxiv.org/abs/2603.05923): predicting a user's next action from their full multimodal interaction history (screenshots, clicks, sensor data) rather than just typed prompts.

🎉 Thrilled to have two papers accepted to ACL 2026 main! 1. Graph-based models match LLMs on close-ended human simulation tasks with far less compute & greater transparency 2. (oral) How to allocate human samples towards fine-tuning vs post-hoc rectification in simulation

Research --> product!

@chris_bail @dorazhao9 @Diyi_Yang Nice—thanks!

@dorazhao9 @Diyi_Yang @msbernst Cool paper-- you might be interested in this new paper, forthcoming in AJPS: osf.io/preprints/osf/…

Excited to be sharing three papers at #CHI2026! 1⃣ Value Alignment of Social Media Ranking Algorithms 2⃣ Mapping the Spiral of Silence: Surveying Unspoken Opinions in Online Communities 3⃣ Whose Knowledge Counts? Co-Designing Community-Centered AI Auditing Tools with Educators in Hawai`i

This is one of my fav figures in our paper. You can: 1. Identify a user's objective by observing general interaction with their computer. 2. Use it to construct a "just in time" rubric. 3. Sample bunch from model and SCALE TEST TIME COMPUTE ON LITERALLY ANY OPEN-ENDED TASK?!?

@RobbWiller Amazing, congratulations!!!

Very grateful to receive a Guggenheim Fellowship. I’ve been so lucky to have such incredible students, collaborators, and mentors whose coattails I’ve hitched myself to for many years. Above all, I feel very lucky that I get to work on topics I care deeply about with collaborators who are also my friends.✊❤️#guggfellows2026

This paper was a fascinating experience where, when we first submitted it, reviewers refused to believe that we could create contentious social media content with LLMs. This time around, they saw it as the main point of novelty. As time passed, our work got _more_ novel?

"Algorithmic feeds are the first at-scale misaligned AIs"

@michelle123lam So great to see this!

Most of what I actually need help with, I never think to tell a model. But why is it on me to remember? Our new paper asks: what if AI could proactively specialize to individuals and the tasks they’re carrying out at this very moment? 🧵

Best paper honorable mention at CHI — instantly adaptive objectives for LMs as you use your computer

@percyliang Congratulations!!!

Academic titles are funny. After 14 years, I finally have the official title that people might have always assumed I had.

Damn, this was actually very good. Ended up watching the whole hour.

MIT postdoc opportunity! We're hiring a human-AI interaction postdoc (HCI+ML/RL) to train agents that deepen how people think and collaborate - rewarded by how humans actually build skill together. With @arvindsatya1 @ZanaBucinca, me & more! Apply by May 1 tinyurl.com/4jsr8ee9

GitHub: github.com/GeneralUserMod… For example: here’s @gandhikanishk following the Australia v. Zimbabwe match while monitoring a training run. All annotated with NAPsack!

Passive interaction data is super underrated. Recruit users, observe what they're already doing (and willing to share!), and label those trajectories with a VLM! It's free lunch!!!! We're releasing an open-source package (NAPsack) to do this, tested on 1.9M screenshots. 🧵

We trained our models to solve problems with objective answers. But can we build models that solve problems where success is subjective, messy, and human? The latter is even more impactful imo. Agree with Percy: simulation is the next frontier for AI.

Current AI is reactive. You prompt, it responds. True proactivity requires predicting what you'll do before you ask. Our new work done by @oshaikh13 formalizes this as Next Action Prediction (NAP ): given a user's computer use, predict their next action. We annotated 360K actions across 1 month of continuous computer use from 20 users and open-sourced a pipeline for private-infra labeling. LongNAP combines parametric + in-context learning to reason over long interaction traces. This is one step closer to an assistant that proactively anticipates, not just reactively responds 🚀

Predict what the user will do next: a task that underlies a huge number of goals. Better assistants, better user models, the next generation of operating system metaphors...

What’s the point of a “helpful assistant” if you have to always tell it what to do next? In a new paper, we introduce a reasoning model that predicts what you’ll do next over long contexts (LongNAP 💤). We trained it on 1,800 hours of computer use from 20 users. 🧵