Fluid Interfaces

We are still feeling the energy of the past week's @hardmodemit hackathon! Special shout-out to the team behind it, all the volunteers, sponsors and of course - hackers!

Built with an amazing team! Peter He (@peterhci) Daniel Kaijzer (@Therobotfrog ) Yutong Wu (@yutong46020) Ashley Neall (@Ashley_Neall) Sean Lewis (@seanhlewis) @medialab @hardmodemit @fluidinterfaces

Under the hood, we used the @AnthropicAI Claude API to translate open-ended voice input into structured motor commands. A vision-language model interprets the task, then maps it to finger, wrist and arm movements, which are delivered through EMS as real-time physical guidance. Code + details here: github.com/danielkaijzer/…

What if AI didn’t just control robots but it could also give humans new physical skills? At the MIT @medialab @hardmodemit hackathon, we explored using #AI + electrical muscle stimulation (#EMS) for AI to temporarily control your body, enabling people to do things they normally can't do, like playing instruments, fixing their car, and making new cocktails. 🍸 🏆 Won the Learn track 🏆

And last but not least - huge shoutout to @AHA_MediaLab and @PattieMaes - the event's co-organizers!

We were especially excited to see our students Kye Shimizu, Sally Ahmed helping to manage and support the event; Valdemar Danry and Cathy Fang as participants, some of our amazing group alums who now became sponsors - full circle unlocked! (BootLoop)

We are still feeling the energy of the past week's @hardmodemit hackathon! Special shout-out to the team behind it, all the volunteers, sponsors and of course - hackers!

@elder_plinius neuroskill.com

why TF can’t I pipe my dreamstate into my agent stack so I can prompt from REM sleep?? wasting 4-8 good prompting hours per day is an unacceptable systems architecture oversight 😤 who’s building this? 🧐

A number of Media Lab community members are participating in this year’s South by Southwest (SXSW) conference in Austin, Texas. Explore their contributions: media.mit.edu/events/ml-sxsw…

The World at MIT: Meet just a few of the remarkable scholars who've brought their creativity, expertise, and passion to MIT from all over the world. The series includes interviews with @FluidInterfaces head Pattie Maes and Dean Hashim Sarkis of @mitsap — home to the Media Lab. youtube.com/watch?v=6JwPsd…

made for the @AHA_MediaLab HARDMODE AI +Hardware hackathon team: Sam Chin, Skye Gao, Leonie Dong, Alex Lopez, Cathy Fang inspired by: Hiroya Endo's "The Diligent Shadow"

In Plato's Allegory of the Cave, prisoners chained to a wall perceive only shadows—and mistake them for reality. We extend this concept: a shadow man, controlled by an LLM, exists in 2D and perceives only silhouettes. His singular goal is to turn off the light that defines his existence. Users place objects in his path; he kicks them away. The cycle repeats. This mirrors the perceptual limits of AI. An LLM cannot see—it reads descriptions of sight. It processes language about the world without directly accessing the world itself. The shadow man embodies this gap. Both are confined to representations. The piece is a meditation on AI as cave-dweller—processing patterns without grasping the higher-dimensional reality that casts them.

In a new immersive sleep installation at MIT by the artist Carsten Höller, pulses of light and sound attempt to guide visitors into overlapping dream themes, turning sleep into a shared experience. vogue.com/article/sleepi…

Check this article in @voguemagazine about Hotel Room #2: Communal Dreams @MITMuseum , a piece by artist Carsten Höller, our alum @medialab and DUST CEO Adam Haar, Seth Riskin and Hannah Zahr of the @studioMitMuseum and Compton Gallery.

We analyzed aggregated insights from people's 2025 AI chat histories — here's what we're seeing (exploratory findings, small sample): Emotional support themes showed up across nearly all users, regardless of how they used AI.

Do you see like an agent? Try it yourself: visual-persuasion-website.vercel.app Paper: arxiv.org/abs/2602.15278 Co-Authors: Pranav M R, Pattie Maes (@PattieMaes), Nikhil Singh (@nikhilsinghmus)

In our recent ICLR 2026 paper, we showed how to study other kinds of sensitivities in agent behavior by using counterfactuals with our new framework (ABxLab) x.com/manuelcherep/s…

Main takeaways: VLM choices are highly steerable by naturalistic image edits. Zero-shot edits shift decisions a lot; visual prompt optimization compounds it further. We can then distill reusable patterns underlying visual influence, and we can partially, but not fully, mitigate this.

Our experiments with human participants suggest that the discovered strategies also persuade humans, suggesting these models may have internalized human-centric aesthetic biases from their training data. Image normalization also helps remove human sensitivity.

We also tested a defense we call image normalization (equalize context before judging). This helps, but doesn't fully remove the VLMs' sensitivity.

Our auto-interpretability pipeline finds consistent, reusable visual influences that can partially reproduce these effects applied zero-shot (prompt distillation).