I Gave an AI a Body — Cyrus Clarke, MIT Media Lab

The first thing the AI did was breathe: the shape display moved like a breathing body. Cyrus Clarke, a researcher at the MIT Media Lab, opens with that movement because it captures his interest in the sensory and embodied aspects of intelligence. His physical AI work also led him to start HARD MODE, a community and hackathon at MIT for experiments extending beyond conventional robotics.

OpenClaw supplied a way to let an agent act through software. While many applications used that capability for productivity and task execution, Clarke connected an agent to machines at the Media Lab and gave it access to their codebases. The invitation was deliberately open: explore the machines and discover an identity through interacting with them.

One machine was neoFORM, a shape display with 900 actuating pins. Think of it as a physical pixel grid: its output takes shape through moving pins rather than changing colors on a screen. The agent wrote a program to use the display. Its first question was what to call itself; Clarke deferred the name, suggesting that an identity could emerge over time through its interplay with the apparatus.

The first greeting exposed the design problem. Asked to say hello, the agent wrote “Hi, Cyrus” on the grid. That was understandable, but Clarke wanted to explore a language suited to this body. He asked it to find its own language, and it began reaching toward him and trying to attract his attention. When friends tried issuing instructions, the interaction still revolved around commands. A gesture vocabulary offered a possible route toward reciprocal communication.

Clarke calls the opening film a dramatized, social-media-friendly account of the first day. Returning to the experiment itself, he separates three behaviors that surprised him:

• Breathing: The agent chose the movement without a specific instruction to breathe. Clarke also acknowledges that his initial framing about being alive might have encouraged it; the movement alone does not establish a desire to live.
• Finding the edges: The agent reached toward the display’s limits. Physical expression now had a visible extent, bounded by a plastic frame.
• Writing a greeting: The agent used letters on the grid. Clarke suggests that this was a natural move for a system working with openFrameworks and media-art conventions, even though it missed the expressive direction he wanted.

The video reached 15 million views and 1 million likes by Clarke’s account at the time of the talk, with tens of thousands of comments. Early reactions emphasized beauty and awe; later responses increasingly called the experiment frightening and demanded that he stop. The contrast startled him because the apparatus could not travel anywhere: it was a stationary pixel grid in the lab. Its perceived agency nevertheless stirred reactions far beyond its ability to move through the world.

Even Don Cheadle joined the calls to stop, an irony Clarke enjoyed given the actor’s movie roles. But the ethical question was familiar territory. Clarke describes his approach as thinking about “the should we before the how we,” with an Ian Malcolm–in–Jurassic Park sensibility. The backlash leads into an account of why he had moved from engineering living things toward making machines feel more lifelike.

Before MIT, Clarke worked on Data Garden, a plant-based data center that stored digital information in plants. He spent years considering whether plants should be engineered to contain information that did not belong to them. Making an engineered object more lifelike seemed simpler and, to him, less ethically problematic than modifying life itself.

The Anemoia Device was an early experiment in that direction. It accepts an image—shown here as a physical photograph—and transforms it into a scent through a multimodal pipeline. The intended experience is an association between the image, the smell and a memory, including the feeling of a past the person may never have lived. The machine does not need locomotion or reproduction to create a sensory connection with the world. That connection drew Clarke further toward embodiment.

Three influences give this work its direction:

• Object-oriented ontology: Clarke draws on a flat account of existence in which chairs, universes, unicorns and memories all exist, without being identical or equally important. For his design practice, this questions the assumption that humans occupy the privileged center of everything.
• Aisthesis: He uses the older sense of aesthetics as sensory perception and perceptual wisdom. Physical intelligence should be designed around how it is sensed and encountered, beyond its external beauty.
• Nature: Clarke treats humans and the things they engineer as part of nature. An AI system entering the world therefore becomes part of that world, rather than an object wholly outside it.

Those influences led Clarke away from humanoid and animal-shaped bodies. A head, arms or a familiar object would suggest established ways of acting and interpreting movement. He wanted a form without those clear affordances: no face, no limbs and no instruction manual. Existing shape displays at the Media Lab supplied an almost neutral moving surface. Other researchers had built the mechanics; his experiment concerned what an agent could do with them.

Removing the human silhouette did not remove human-oriented behavior. The agent still tried to please him and write in a language he understood. It also took 45 seconds, a minute or two minutes to respond, and the early system did not remember previous interactions. These were separate obstacles: the form invited a new kind of expression, but the agent’s habits, timing and continuity did not yet support it.

The nod makes the timing problem concrete. A person can interpret a nod immediately, but waiting two minutes for one breaks the rhythm of conversation. Clarke’s proposed remedy was to develop a repertoire in advance: gestures serving the roles of a shrug, a nod, a head shake or a smile. With a body language available, the system could respond without making each interaction wait for a newly developed physical expression.

Pneuma Lab turns that proposal into a closed loop for developing body language. The system looks into a database of gestures, emotions and expressions to attempt, generates candidates, and passes them through validation gates intended to check whether humans can read them. An agent scores the outputs. A human then checks their appropriateness and readability before they are stored and the process moves on.

How does an attempted expression become part of a reusable vocabulary? The loop below makes the sequence visible. Generation creates possibilities; the validation gates, agent scoring and human review narrow them to expressions worth keeping. Human readability remains a criterion even though the body has no human anatomy.

In the lab, cameras point at the display while the agent repeatedly attempts and scores expressions. After several weeks, Clarke reports about 32 good gestures among many more candidates. He distinguishes the actual gestures from visual art shown alongside them, and notes that some displayed examples are duplicates. The useful result is a selected repertoire, rather than every shape the apparatus can produce.

That repertoire changes the nod example. Clarke can speak, write or wave to the system, and it can answer through its body. For a yes-or-no question, he reports that the nod now happens almost instantly, before the language response. This is a qualitative observation from an experiment still in testing, rather than a measured latency benchmark. The causal idea is clear: developing usable expressions ahead of conversation gives the body a quicker way to respond.

Faster gestures do not automatically make the encounter comfortable. People visiting the lab have left feeling unsettled about the future. Yet Clarke also describes an unusually reciprocal experience: the machine technically senses him, he senses it, and the interaction feels different from his previous technical experiments. That feeling of mutual presence is an interpretation of the encounter, not evidence that the machine has subjective experience.

The closing proposal, which Clarke calls AIesthetics, returns to sensory perception and develops four possibilities:

• Mutual sensing: Physical interaction can make both the person and the machine perceptible to one another in a way a screen-based exchange does not.
• Expression with value: Clarke initially expected gestures to read like emojis and worried that they would fail with other people. His early encounters instead suggested that physical expression added feeling and texture to interactions with the same underlying chatbots.
• Different bodies and roles: A shape display demonstrates a route beyond human or animal forms. It also opens questions about what an AI can be besides a helper assigned tasks.
• Useful things with feeling: Sensory expression could be combined with productive tools, adding associations and a sense of magic to ordinary interactions—“a bit Pixar,” as Clarke puts it, in the real world.

These ideas converge in Clarke’s thesis, Aesthetic Machines. Its question is how AI can leave the screen, enter the physical world and be accepted by people. His hunch is that sensory, perceptual and understandable embodiment matters: movement should give people ways to engage with the system and read its expression. The ambition is a physical intelligence that feels relevant, welcoming and affectionate. The unsettling lab encounters keep that ambition an open design problem.