Fluid Interfaces retweeted
The Workshop on Agent Behavior is tomorrow starting at 8:30 AM in Continental Ballroom 4! If you’re @COLM_conf come check it out. We have some amazing speakers giving very different perspectives on how to study agents.
✨Announcing the first Workshop on Agent Behavior @COLM_conf 2026 (Oct 9, San Francisco 🌅)
aiagentbehavior.com/
We invite two types of contributions: (i) papers, and (ii) benchmark proposals. We are also seeking reviewers. More details below!
Fluid Interfaces retweeted
Most AI benchmarks measure what models can do. ImpactBench measures what AI does to us.
800+ behavioral metrics. 48,540 conversations. Ten models. Undesirable behaviors avoided in only 53.1% of evaluations.
Explore: impactbench.media.mit.edu
Fluid Interfaces retweeted
Curious about applying to the Media Lab? Register for the Virtual Open House for Prospective Students on October 9! You’ll have an opportunity to hear from staff, faculty, and students; learn about the academic program and the research groups; and ask questions.
Learn more and register for the webinar: media.mit.edu/events/2026-ma…
Fluid Interfaces retweeted
Last week, I was invited to present our @icmlconf 2026 paper “Visual Persuasion: What Influences Decisions of Vision-Language Models?” at the ECCV VisPer Workshop.
We created a visual prompt optimization method (CVPO), in which we iteratively edit images (e.g., w/ @NanoBanana 🍌) to maximize the feedback given by VLM judges. While this method can be used for lots of applications, we use it to discover factors that persuade agents (and, as a proxy, people) while making decisions. More details and examples below!
Work w/ Pranav M R*, @PattieMaes, and @nikhilsinghmus
Fluid Interfaces retweeted
I'm at COLM this week! If you're interested in chatting about AI agent vs human behavior, feel free to reach out :) I will also be on campus at Stanford (Oct 6-7) and Berkeley (Oct 8)
Join us @woburnlibrary October 3, 2026 to learn more about the science and fiction behind Brain-Computer Interfaces!
@NeuraFutures @medialab @ArtsatMIT
NeuraFutures or the Next User Interface is Your Brain
Join us at @woburnlibrary on October 3, 2026 to learn more about the NeuraFutures exhibit and try out some brain-computer interfaces for yourself!
Fluid Interfaces retweeted
We got Best-in-Show at @ArsElectronica Expanded Conference!
Echoing the gallery’s theme of “expanded play”, seeing the surprises and pure joy on people’s face when playing with our piece was the biggest reward.
ars.electronica.art/negotiat…
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.
Fluid Interfaces retweeted
Curious about applying to the Media Lab? Register for the Virtual Open House for Prospective Students on October 9! You’ll have an opportunity to hear from staff, faculty, and students; learn about the academic program and the research groups; and ask questions.
Learn more and register for the webinar: media.mit.edu/events/2026-ma…
Fluid Interfaces retweeted
In 1979, MIT researchers said "put that there" to a computer, and it listened. Yesterday, OpenAI launched GPT-6 Astra, the most powerful AI model ever built, opening with 47-year-old footage from our Media Room to show exactly where the idea came from. The premise that Richard Bolt, Chris Schmandt, and Eric Hulteen proved nearly five decades ago still holds: machines should adapt to human language, not the other way around. The best ideas don't expire...they compound. READ THE STORY: mitmedialab.info/4zTwbXb
Fluid Interfaces retweeted
What's your real relationship with @claudeai ? Our AHA team collaborated with @AnthropicAI on Reflect, Claude's new beta feature that shows your AI habits and helps you stay aligned with your personal goals. Check your Claude settings (desktop) to explore your own patterns!
Fluid Interfaces retweeted
Do some AI models refuse more user requests than others?
Are people using AI to generate illicit content more or less over time?
And which models do people go to for coding vs role-play?
Unfortunately, existing reports of AI use from frontier labs are coarse and proprietary that do not help us understand the impacts and natural uses of AI.
This prompted us to create a public AI observatory ai-observatory.org/
1/ How are people really using AI?
Today, researchers across @MIT, @Stanford, + 12 academic institutions are launching the Public AI Observatory: public infrastructure for measuring how people actually use AI assistants in the wild.
🔗 ai-observatory.org/
📜 dataprovenance.org/ai_observ…
Fluid Interfaces retweeted
Today we are excited to release the NeuroDatasets tool, where one can browse and find all available neuroscience datasets for all major modalities: EEG, MRI, MEG, NIRS, PET and more.
Check the NeuroDatasets tool here: datasets.neuro2.ai
In the latest essay @WBUR Mike Kalin talks about AI-powered systems that can help with memory support for those who might need it due to cognitive impairments and/or age. One of the systems highlighted in the article is MemPal, co-led by our PhD student @wazeerzulfikar @medialab
Check the paper by @ManiarNatasha @cwt_sam @wazeerzulfikar Scott Ren, Christine Xu, and @PattieMaes here: dl.acm.org/doi/10.1145/37083…
Check the webpage of the project here: media.mit.edu/projects/mempa…
Watch the replay of the keynote talk at this year's @ACMIUI from our team's lead, @PattieMaes @medialab on the topic of « Designing Al Interaction for Human Flourishing ».
Watch the video here: youtu.be/nsCQkx6nbhs
Fluid Interfaces retweeted
In a recent study, @anku__rani, @valdemardanry, Prof. @pliang279, Dr. Andrew Lippman, and Prof. Pattie Maes (@fluidinterfaces) show that although chatbots can help people detect misinformation, users who rely on AI systems to verify facts may become worse at spotting fake news on their own. However, the researchers also identified conversational strategies that encourage active learning and skill development, rather than simply helping in the moment.
news.mit.edu/2026/consequenc…
