>_ What you are actually looking at
Act 1 — Reading a mind from its wiring alone
We were handed a file of 17,253 numbers. No source code, no game, no documentation of what it was trained on — just the weights of a tiny neural network. By looking at where the numbers were large and where they were silent, we reconstructed what the network had learned: it only cares about one row of the grid (where a paddle lives), it spends almost all of its capacity tracking a moving dot and remembering where that dot was one tick earlier — exactly what you need to infer velocity — and it expects rewards in three rows at the top. Conclusion: this network was trained to play Breakout. Nobody told us. The knowledge was written into the numbers, and we read it like an X-ray.
Act 2 — The numbers still work, anywhere
Those raw numbers were uploaded to Hugging Face, pulled back down, and they play the game. What you see below is not a recording: the actual 17,253 parameters are doing live arithmetic in your browser, dozens of times per second, deciding left or right. The "intelligence" is nothing but those decimals multiplied through a simple formula — the whole mystery of deep learning made tangible in one file.
Act 3 — Surgery instead of schooling
If a policy makes a mistake, you don't have to retrain it from scratch over thousands of games. You can ask a counterfactual question — which change to which numbers would have prevented that failure while preserving everything else it does well? — and solve for that change directly, verifying the answer by replaying the game. That is the companion benchmark, weight-space-breakout: an AI is handed a deliberately broken policy and must repair its behavior with a handful of scalar edits, scored by simulation. Behavior becomes something you can edit, not just grow.
One honest note: the weights playing below are the original trained ones, unmodified. The repair half of this story lives in that benchmark — and in kimi-grid-breakout, where language models play this same game closed-loop, move by move, graded by the physics itself.
Why it matters
Normally, changing what an AI does means retraining: expensive, slow, and you can never be sure what else changed. Weight-space editing promises something different — behavior as inspectable, addressable mathematics. Find the numbers responsible for a behavior, change exactly those, verify nothing else moved. And if repair-by-reasoning works on a game, it works on any process with checkable steps: a railroad engine manual's worth of structured knowledge becomes hyperspecific behavior for the cost of electricity. Today this works on a 17-thousand-parameter toy; whether it scales to billion-parameter models is the open frontier.
↓ Scroll down to play the live game · Source: github.com/HarleyCoops/breakout-policy-visualizer — zero-build static page, inference in pure JavaScript.