Side by Side with Jaron Lanier at Brown

Before Jaron Lanier took the stage on April 22, 2026, five speakers thanked each other. Vesna Mitrović thanked Jim Russell for agreeing to introduce the introducer. Jim Russell thanked Leon Cooper, who shared the 1972 Nobel Prize for the BCS theory of superconductivity and taught at Brown for half a century, his name sponsoring the talk. Stephon Alexander thanked Cooper for his mentorship, Barrett Hazeltine for taking a 22-year-old version of him seriously, and Lanier for describing him correctly in 1999 at a loft on Duane Street as a complete weirdo. Lanier thanked Brown for not being the other Ivy (I didn’t get it). The Brown Center for Theoretical Physics and Innovation sponsored the evening. The center is named for a Nobel laureate who mentored the man who invited the speaker. The speaker never went to graduate school. The “chain of custody” in operation (Lanier, 2026).

My friend Matthew and I filmed it. I loved every second. The evening was complete, a beautiful event, witnessing and working in tandem. Lanier shared his thinking about AI and theory and then played music alongside Stephon, Donnie Aikins on bass, and Jesús Andujar on percussion. The music and the thought came from the same inimitable consciousness.

I stood behind a precarious tripod. Most of the time I looked at the camera, but I also looked at Lanier, and the whole thing became a bit surreal. I was capturing information about a man talking about how information gets captured. The camera extracted audiovisual data from a speaker arguing that extraction is the problem.

Lanier’s main claim is a reframe. “AI” is not a kind of software. It is a collection of people. A large language model is a giant Wikipedia with statistics on it, with one difference: the writers are there involuntarily. Every prompt-and-output is an act of erasure. The specific humans whose words got interpolated to produce a given sentence are nameless.

His proposed fix is a parallel channel that surfaces the training-data clusters most responsible for each response. He calls it counterfactual cluster estimation. During Q&A he mentioned he had money for one extra summer intern on the problem.

His larger point: we need stories. We need experiences that change our relationship to these tools, and to the collections of people they inscribe. He told a student who asked whether AGI was bullshit that the student had “not asked him a question”. The definition of the problem defines the scope of the answer. Lanier follows Bergson and Deleuze in naming the primacy of the question. Real power is in the formulation.

He also debunked the Turing test. He argued that the test works to make humans stupider, more capable of being deceived, lowering the bar until a machine can clear it. The original imitation game came from a 1940s British parlor game where a man and a woman sat behind a screen and a third person had to guess which was which. Pass for the other gender. The grammar of passing was the evidence of intellect. Turing himself, gay and chemically castrated by the same government that devoured his code-breaking, knew what passing cost. Lanier argues that this is the wrong model for intelligence.

Lanier is right that the word “information” carries two incompatible meanings and that the confusion between them has broken computer science. Shannon information is observer-independent, entropic, measured in bits, physical. Semantic information is observer-dependent and context-bound. You can think of Shannon information as inherent to the object it describes, and semantic information as the context of that being.

But what I have found is that information is not two things. It is three.

Shannon is the first. Semantic is the second. James Gibson named the third: an affordance. Information that exists only in the relation between a body and an environment, and only while the body is acting. A chair affords sitting to something built to sit. The affordance does not live in the chair or the sitter. It lives in the relation.

Affordance is information that cannot be extracted because there is no object to extract from. It is event, not substance. It lasts only as long as the relation.

Most of what we call AI runs on Shannon and reaches for semantic. None of it touches affordance, because affordance is not an asset that can be extracted. The grammar of objectification and contextualization has nothing to grab.

I am reading Deleuze’s Bergsonism with my friend Paul, and Bergson’s intuition lands differently after hearing Lanier. Bergson’s central argument is that the intellect spatializes duration: it takes continuous, irreversible process and cuts it into discrete, reversible units. That cutting makes measurement possible and extraction possible in the same stroke. You cannot capture what you cannot cut, and what you capture by cutting is never the thing you cut it from. Gibson’s affordance is the part that resists the cut, because it exists only while the relation is active. Stop the action and the affordance disappears.

Lanier follows this line without naming it. His counterfactual cluster estimation tries to reconstruct the relational history that the model erased. Bergson would push back: you cannot reconstruct duration from its spatial trace. The trace remains after the cutting, but the relation was what the cutting destroyed. Either Lanier’s parallel channel can do something Bergson thought impossible, or it is an honest admission that the loss is permanent and the best anyone can do is mark where it happened.

The camera in my hands was Shannon and semantic at once. It captured bits. It also captured a man whose argument was that the bits were a kind of theft. What it could not capture was the affordance: my looking and his speaking, the room agreeing to be there together.

What the camera does, the machine learning model does at scale. Shannon plus semantic, both running, a quasi-independent operation with the relationship beyond the limit of the sensible. The third kind of information, relation, collapses into the first two and disappears.

I asked myself, behind the lens, what it costs to look at oneself or another human being through a mirror. Whether it is different from seeing yourself or another on camera, capturing audiovisual data, then manipulating that material into a stream of content that takes from the subject, objectifies it, and circulates it.

AI will never have an experience. The model can describe a relationship but cannot enter one. Affordance is the part of information that requires you, the we, the sense of “I” that is rooted in relation.

Around this time last year I had a brief hiatus of unfortunate internet fame. A man pushed me to the ground and called me a f*****. I had my phone, started filming. I was not hurt. I did not press hate-crime charges. A Jewish white man became the face of hate-crime victims in 2025. I gave an interview to the local news, hosted a few run clubs that spring and summer, got a lot of DMs from men who wanted to flirt. It was fun, and equally odd. I felt important. I tried to make content about running and the increasing authoritarianism in our individual souls and collective political life.

Then the content made content of me.

The footage was Shannon. The interview was semantic. The affordance was the actual moment with the man on the sidewalk, the choice (non-choice) to film as reflex. None of that lived experience traveled through the digital textures that surrounded me. What traveled was the part that could be made object, made context, not the relationship that produced it.

I think I care about this. I also think I think I care, which is the horseshit that makes me suspicious of every thought that passes through. Caring that travels well on a feed is the kind of caring that has already lost its affordance, some unlocalized part of our being. The second-order caring is the only place the relation might still live. Caring whether my caring is real.

Filming Lanier was a privilege because the doubled mirror asked me to attend to both operations at once. Extracting and relating.

You are reading prose that passed through many people. The captions came from an automated speech-to-text model trained on recordings whose speakers I cannot name. I cleaned those captions using Claude, a product of Anthropic, a company Lanier cited by name in his talk. Claude’s weights almost certainly include Lanier’s past essays, which means when I asked the model to clean the transcript, it interpolated his prior voice back into his current words. I corrected “Jaren” to “Jaron” and “Entropic” to “Anthropic” and “Jesus Anunda” to Jesús Andujar, the Providence-based percussionist confirmed through Brown’s coverage. Donnie Aikins gave me his number after the concert.

The full corrected transcript is here.

None of this is hidden. All of it is here in the prose, visible. This essay is a working prototype of what Lanier proposed.

Thanks to: Jaron Lanier, Stephon Alexander, Vesna Mitrović, Jim Russell, Leon Cooper (posthumously), Donnie Aikins, Jesús Andujar, Anthropic, every writer whose text shaped Claude’s weights, every speaker whose voice shaped the captioning model, Gilles Deleuze, Henri Bergson, James Gibson, Paul Aste, Matthew Osubor, Barrett Hazeltine, and Brown University for the chair in the room.