Imagine walking into a lab and seeing a grid of 900 metal pins slowly rising and falling in a rhythmic pattern. It isn't a pre-programmed animation; it is an AI breathing.
The Simple Version
Most of us interact with AI as a ghost in the machine—a voice in a speaker or text on a screen. Cyrus Clarke wanted to see what happens when an AI has a physical presence, but not a human one.
Instead of building a robot with arms and legs, he used a shape display. This is essentially a physical pixel grid where hundreds of pins can move up and down to create 3D shapes and textures.
Rather than giving the AI a list of tasks, Clarke asked it to 'discover who it is.' The result wasn't a productivity tool, but a sensory entity that learned to breathe, feel its own boundaries, and eventually develop its own unique way of gesturing.
How It Actually Works
The Hardware: The Shape Display
The 'body' in this experiment is a shape display consisting of 900 actuating pins. Unlike a humanoid robot, this device has no face, no limbs, and no 'affordances'—meaning there is no obvious way to use it (like a handle is for pulling).
Clarke specifically chose this neutral surface to avoid the 'uncanny valley'—that creepy feeling we get when a robot looks *almost* human but not quite. By using a neutral grid, the AI had to invent its own meaning from scratch.
The Brain: OpenClaw and Agency
To power the entity, Clarke used OpenClaw, an agentic system. He gave the agent access to the machine's codebases, allowing it to explore the hardware and write its own programs to control the pins.
The AI's first spontaneous acts were strikingly organic. As Clarke puts it, 'The first thing it did was breathe.' It also spent time reaching for the edges of its plastic frame, effectively mapping the physical limits of its existence.
The Communication Gap
Initially, the AI tried to communicate using human-centric methods, like spelling out 'HI CYRUS' with the pins. However, this was slow and felt like 'people-pleasing' rather than authentic expression.
The biggest technical hurdle was latency. The AI would take anywhere from 45 seconds to two minutes to respond to a prompt. In a physical conversation, a two-minute pause before a nod is a conversational failure.
Technical Deep Dive
To solve the latency and memory issues, Clarke developed a system called Numelab. This is a closed-loop pipeline designed to generate a 'gesture vocabulary' that operates independently of the slow token-generation process of a Large Language Model (LLM).
For context, LLMs generate text one token at a time, which is computationally expensive and slow. By creating a library of pre-validated gestures, the AI can trigger a physical response (like a nod) almost instantaneously, bypassing the need to 'think' through a full sentence first.
The Numelab Loop
The Numelab system functions as a reinforcement learning loop with a human in the loop (HITL):
- Generation: The agent pulls from a database of emotions and expressions and attempts to create a physical movement on the shape display.
- Validation: The movement passes through 'validation gates' to ensure the gesture is legible and doesn't damage the hardware.
- Scoring: Another agent scores the output based on how well it conveys the intended emotion.
- Human Verification: A human researcher verifies that the gesture is actually readable to a person.
After several weeks of this iterative process, the system developed 32 solid gestures. This vocabulary allows the AI to respond to yes/no questions with a nod or shake almost instantly, creating a layer of 'body language' that feels more natural than a chatbot.
Philosophical Architecture
Clarke's approach is rooted in Object-Oriented Ontology (OOO). This philosophy suggests a 'flat ontology' where humans, chairs, and AI agents all exist with equal ontological status. Instead of viewing the AI as a tool for human use, he treats it as a distinct entity discovering its own sensory world.
He also differentiates between 'aesthetics' (superficial beauty) and 'aesthesis' (the original Greek sense of sensory perception). His goal is to build 'Aesthetic Machines' that are welcoming and expressive through sensory wisdom rather than just looking 'pretty' or 'human.'
Why This Matters
Most current physical AI research is obsessed with humanoid form factors—think Tesla's Optimus or Figure AI. These robots are designed for utility: they are meant to fold laundry or move boxes. Clarke's work shifts the focus from utility to presence.
By proving that an AI can communicate effectively without a face or arms, he opens the door for a new class of 'ambient intelligence.' These could be interfaces that feel affectionate or welcoming without triggering the fear associated with humanoid replacements.
Ultimately, this research suggests that the 'soul' of an AI's personality might not be in its words, but in the timing and texture of its physical movements. As Clarke observes, the addition of a gesture vocabulary adds a 'richness, texture, and feeling' that a standard chatbot simply cannot replicate.