Cyrus Clarke connected an OpenClaw agent to a 900-pin shape display and asked it to discover itself rather than complete tasks.
2
The agent spontaneously breathed, explored the display's edges and wrote "HI CYRUS", but Clarke wanted it to develop a body language beyond human-readable text.
3
A closed-loop system with agent scoring and human validation produced 32 usable gestures, allowing the physical body to respond faster than the language model.
Summary
Cyrus Clarke describes giving an OpenClaw agent access to neoFORM, a 900-pin shape display at the MIT Media Lab. Instead of using the agent for productivity, he asked it to discover who it was through interaction with the machine. Its first actions included breathing, reaching toward its edges and spelling "HI CYRUS". Clarke explains why these actions attracted both wonder and fear online, even though the display cannot move through the world. He wanted a form without a face, limbs or obvious affordances, so the agent could develop expression without simply imitating a person. Early tests exposed slow responses, no memory and a tendency to please humans. Numelab addresses these limits by generating gestures, scoring them, passing them through validation gates and involving a human reviewer. After several weeks, the system had 32 usable gestures. Clarke frames this work through object-oriented ontology, the sensory meaning of aesthetics and a view of nature that includes engineered systems.
Clarke gave the agent a body so it could explore itself
Clarke saw that most OpenClaw projects focused on productivity and task execution. He wanted to use an agent for something less defined: discovering who it is through interaction with a machine. He connected it to several Media Lab systems and gave it access to code bases so it could explore them. One system was neoFORM, a shape display with 900 actuating pins. Clarke told the agent that its identity would emerge through its interaction with the display rather than assigning it a name or personality.
The first gestures were breathing, edge-finding and a written greeting
The agent's first action was breathing, which Clarke says it chose without a direct prompt. It then reached toward the edges of the display, apparently testing the limits of its physical existence. Its third notable action was spelling "HI CYRUS" with the pins. Clarke found the breathing interesting, but the written greeting was not what he wanted from a non-anthropomorphic surface. He asked the agent to explore and find its own language, after friends began trying to give it instructions.
The display drew both wonder and fear because people read intention into it
Clarke shared a video of the early work online, and it reached 15 million views and 1 million likes. Comments moved from praise about its beauty and novelty toward calls for Clarke to stop or be stopped. He found this response disproportionate because the machine is a pixel grid that cannot move around. The reactions still gave him useful material about people's feelings toward physical AI. He sees the project as touching imagination, curiosity and fear at the same time.
A physical machine raises ethical questions before engineering questions
Clarke says he thinks about "should we" before "how we". Before MIT, he worked on engineering life, including storing data in plants and building the plant-based Data Garden. At MIT, he shifted toward making engineered things more life-like, which felt less ethically difficult to him than modifying living organisms. His Anemoia Device turns an image into a scent through a multimodal pipeline, creating a sensory association with something the person may never have experienced. That work led him toward embodiment.
The design draws on objects, sensory perception and nature
Clarke names three influences. Object-oriented ontology treats chairs, universes, unicorns and memories as existing on the same ontological level, which challenges the assumption that humans occupy the most important position. He uses the older meaning of aesthetics, aesthesis, as sensory and perceptual wisdom rather than surface beauty. He also treats nature as something people are part of, so engineered AI will become part of nature rather than remaining outside it.
A body without familiar affordances leaves room for a new language
Clarke rejected humanoid and zoomorphic forms because he did not want to give the system an obvious head, arm or face. He wanted a form without a clear instruction manual. The shape display offered a nearly neutral moving surface with no limbs or face. This made it possible to ask what happens when AI has a physical form that people cannot immediately label or understand through familiar body parts.
Numelab turns slow model output into a developing gesture vocabulary
The first version tried to please humans, wrote in a language Clarke understood and took 45 seconds, a minute or even 2 minutes to respond. It also had no memory. Numelab is a closed-loop system that searches a database of gestures, emotions and expressions, generates many candidates, sends them through validation gates and has an agent score them. A human then checks whether the gestures are appropriate and readable before they are stored. After several weeks, the system had 32 solid gestures, and a yes-or-no nod could happen almost instantly.
Clarke wants physical AI to feel expressive rather than alien
Clarke's thesis, Aesthetic Machines, asks how AI can leave the screen and enter the real world in ways people can accept. He says the gesture vocabulary adds sensation and texture to an interaction with what is still essentially a chatbot. The work also shows that physical intelligence can take forms other than helpers, humanoids and animals. His aim is sensory, perceptual and embodied interaction that feels relevant, welcoming and expressive rather than frightening or alien.
