Where does the mind end and the world begin?
For most of modern history, we assumed a strict boundary: cognition happens entirely within the skull. The world is just the stage where the brain's decisions are acted out.
But in 1998, philosophers Andy Clark and David Chalmers proposed a radical alternative. They argued that if a piece of the outside world functions in the exact same way as a piece of your brain, there is no meaningful reason to draw a line at the skull. The mind can leak out into the environment.
Otto and the Arcade
Clark and Chalmers famously used the thought experiment of Otto, a man with memory loss who relies completely on a notebook. Because the notebook functions identically to biological memory, they argued, the notebook is part of Otto's mind.
But four years earlier, cognitive scientists David Kirsh and Paul Maglio proved this wasn't just a thought experiment. They watched humans play Tetris.
They noticed players frantically rotating blocks on the way down. These rapid keystrokes weren't mistakes. The players were offloading the heavy cognitive burden of mental rotation onto the physical screen. Pushing a button was computationally cheaper than calculating matrix transformations in the brain. The physical action was, literally, a thought.
Training the Philosophers
To test the boundaries of the mind, we built an arcade. We trained two Reinforcement Learning AI agents to play Tetris. Both have the same neural architecture. Both see the exact same board. The only difference is their philosophy, encoded into their reward functions.
Agent Alpha — The Internalist. Punished with a strict point penalty for every single keystroke. It must compute the spatial reality of the board entirely within its own hidden layers before making a move.
Agent Beta — The Extended Mind. Operates without keystroke penalties. It is free to manipulate the physical environment to see what happens.
Below: both agents at epoch 10 000. Still essentially random — the philosophies haven't yet imprinted themselves on play.
The Divergence
By checkpoint 250 000, the two philosophies have violently diverged.
Watch Agent Alpha on the left. It processes everything internally. It waits, computes the optimal placement, and acts with robotic, single-keystroke efficiency.
Now watch Agent Beta on the right. It is fidgeting. It spins the T-block redundantly, just to see how the geometry aligns with the stack below. These are what Kirsh and Maglio called epistemic actions — changing the world to make thinking easier. Beta is thinking with its hands.
The cyan ghost-trails mark frames where the JSON's
is_epistemic_action flag fires.
The Mastery
By the millionth timestep, both agents play with practised skill — but the philosophies they've internalised have hardened into style.
Agent Alpha continues to act with surgical economy. Each piece is placed in its final orientation with the minimum keystrokes possible. The penalty has trained a clean, deliberate mover.
Agent Beta still fidgets. The redundant rotations and lateral oscillations haven't faded with training; they've stabilised as part of how it cognises the board. The cyan ghost-trails are denser than ever.
Watch the screens. The internalist computes. The extended mind thinks out loud.
The Data: a Parity Principle
The telemetry proves the Parity Principle.
Both agents learned to survive the game, but they built fundamentally different cognitive systems. Agent Beta didn't fail to learn mental rotation; it learned that the screen was a better hard drive than its own neural weights.
By epoch 1 000 000, over 20% of Beta's physical actions were purely cognitive offloading. Its mind had extended into the pixels.
Source: data/telemetry_{alpha,beta}.json — 100 evaluation games per
checkpoint, top 5 sample games by cumulative reward exported per checkpoint.