Overview
MIT Media Lab researcher Cyrus Clarke explores what happens when an AI is given an unfamiliar physical body rather than another productivity task. He connected an OpenClaw agent to a shape display—a grid of actuating physical pixels—and invited it to develop an identity through embodied interaction. The agent spontaneously simulated breathing, explored the boundaries of the display, and wrote a greeting, behaviors that fascinated millions of viewers while also provoking disproportionate fear. Dissatisfied with its anthropomorphic communication, slow responses, and lack of memory, Clarke developed NumaLab: a closed-loop system that generates, evaluates, and stores gestures for the display. After several weeks, it produced approximately 32 reliable gestures, enabling nearly instantaneous physical responses such as nodding. Clarke frames the work through object-oriented ontology, the sensory meaning of the Greek concept aisthesis, and an inclusive view of engineered systems as part of nature. His broader thesis is that physical AI need not resemble humans, animals, or conventional robots, nor exist only as a tool. By making AI perceptual, embodied, expressive, and mutually sensible, “aesthetic machines” could leave screens and enter everyday life in forms that feel engaging and intelligible rather than alien or frightening.
Sections
Key Concepts
Terms that organize Clarke’s approach to embodied and sensory AI.
- Shape display: a physical pixel grid whose actuating pins form moving surfaces and patterns, providing the agent with a non-humanoid body.
- Object-oriented ontology: a philosophical position presented here as a flat ontology in which people, objects, memories, fictional entities, and vast systems all exist without automatically granting humanity supreme ontological privilege.
- Aisthesis: the broader sensory and perceptual root of aesthetics, associated with embodied perception and sensory wisdom rather than surface beauty alone.
- NumaLab: a closed-loop system for generating, evaluating, validating, and storing a body-language repertoire for the shape-display agent.
- Aesthetic machines: Clarke’s proposed category for embodied AI that leaves the screen and becomes sensory, perceptual, expressive, and understandable in the physical world.
Higher-Order Implications
Synthesis and implications derived from the experiment and its reception.
- The apparent meaning of an embodied agent may be produced as much by an observer’s interpretation as by the model’s internal process. Breathing-like motion can trigger assumptions about life and intention even when the mechanism is only a stationary pin array.
- Physical expression can operate as a low-latency communication layer alongside slower language generation. The body does not merely display the completed verbal answer; it can acknowledge and respond before that answer is ready.
- Avoiding a humanoid form does not eliminate anthropomorphism. When given an ambiguous body, the agent still selected culturally familiar behaviors such as breathing, reaching, nodding, and writing.
- Acceptance of embodied AI may depend on perceptual legibility more than visual realism. A machine can remain visibly nonhuman while still feeling responsive, affectionate, or comprehensible through movement.
- The project’s strongest contribution may be a design framework rather than a claim about machine consciousness: construct bodies, movement vocabularies, and validation processes that support reciprocal sensing without pretending the system is biologically alive.
System and Experiment Details
Specific implementation details disclosed in the talk.
- Neoform 900 is described as a shape display made from actuating pins arranged as a physical pixel grid.
- The initial agent was connected to multiple Media Lab machines and given access to relevant codebases so it could explore and program them.
- The unconstrained embodied response time was approximately 45 seconds to two minutes, which Clarke considered unsuitable for conversation.
- NumaLab queries a database of gestures, emotions, or expressions; generates many candidates; passes them through validation gates; uses an agent to score them; and includes final human verification before storage.
- The experimental setup uses multiple cameras directed at the shape display while the agent repeatedly generates and scores candidate gestures.
- After several weeks, the system had approximately 32 gestures considered sufficiently good, although Clarke says many additional candidates existed.
- The interface accepts multiple forms of interaction, including speech, writing, and human body-language cues such as waving.
- For binary questions, a physical nod can occur almost instantly and currently precedes the slower language response.
Memorable Quotes
Statements that capture the project’s motivation and conclusions.
- You will create your identity over time. It will emerge through your interplay with the shape display.
- The first thing it did was breathe.
- I'm actually someone who does think a lot about the should we before the how we.
- It has no face, has no limbs, has no instruction manual.
- we don't have to just operationalize AI to be our helper. It can also be other things.