Four brains, live in your browser.

Each one learns by experience: observe, remember, predict, act, learn from what followed. No backward pass, no training switch, no frozen weights. The whole brain is drawn beside the task while it thinks, and every number has a receipt and a verifier.

No backward passOne stream, no replayBeats AlphaZero 0.945 / 0.990Runs on your CPUMIT licensed

01 · Arm

A body learned from babbling

Draw a figure and the arm copies it. It learned its own body from random movement and plans every torque through what it remembers.

  • Nobody programmed its kinematics
  • Adapts to a longer link with no reset, then returns
  • Keeps learning while it copies

Held-out reaching 99% · online transformer on the same stream 40% · online MLP 2%

02 · World

A world that must be remembered

Ask for an object and it goes there from memory. Seen one cell at a time, the world has to be remembered to be navigated.

  • Remembers where it saw each object, corrects the memory when the object moves
  • Holds a cue across a delay and grounds words in objects
  • Keeps navigating when doors stop opening

Requests fulfilled 100% with memory, 18% erased · cue at delay 8 100% vs 52% transformer

03 · Artist

A figure shown, a figure drawn

Draw a figure and the artist draws it stroke by stroke. It learned what its pen leaves by scribbling and replans from its own canvas after every stroke.

  • Draws figures it has never seen
  • Keeps drawing when its body and its canvas change mid-drawing
  • 75 ms per decision at the p95, inside the 100 ms it has

Held-out F1 0.93 · frozen from birth 0.00 · online MLP with replay ring 0.80

04 · Connect Four

The rules, learned by watching

Play against a brain that learned how a stone falls and what four in a row is from the games it played. No rules, no engine, no labelled moves.

  • Beats an AlphaZero trained with the published recipe
  • Searches the boards it imagines eight plies deep
  • Keeps learning from the game you play against it

vs AlphaZero 0.945 at 25 and 0.990 at 100 simulations per move, 100 paired games