🔧D16
Harness Lab
The model is fixed — tune the harness around it. Step budgets, verifiers, observation shaping, typed errors: same model, 34% → 92% of tickets solved.
⚖️D17
Judge & Jury
Build an LLM-as-a-Judge protocol and watch it crown the wrong model. Position, verbosity and self-preference bias — measured by counterfactual.
🩹D2
Self-Healing Pixel Art
Draw a sprite. A tiny neural cellular automaton learns to grow it from a single pixel — then erase chunks of it and watch the wound close itself.
🎧D3
Listen to a Neural Net Learn
Training as a score: each layer is a drone voice detuned by its gradients. An out-of-tune orchestra that tunes itself as the loss falls. Headphones on.
🤟D4
Webcam Sign-Language Recognizer
Teach your browser your signs: record a handful of samples, train in ~200ms, then fingerspell at the camera and watch it type. Nothing is uploaded.
🧠D6
A Photo in a 12KB Neural Net
Don't store pixels — store a function. A SIREN learns f(x,y) → color until it is your photo, then fights JPEG at the same byte budget.
💤D7
Pong Dreams
A world model watches Pong until it learns the physics — then the simulator switches off and you play inside the network's hallucination as it drifts.
🦠D11
Evolving Walkers
Blobs of springs and joints learn to walk with no brain and no backprop — just mutation and the rule that whoever crawls furthest gets to reproduce.