Chip routing is one of the hardest problems in chip design — and a tiny AI is finally beating the big ones at it

Tom Tunguz writes about chip routing, the 3D puzzle of laying thousands of wires between metal pins without them crossing or doubling up. Chip companies have spent billions on software to solve it. That's the kind of problem most people assume needs a big AI — the ChatGPT-scale ones — but the BFTS experiment shows a tiny model, one-hundred-thousandth the size, is doing it better with the right training recipe.

The proof is in the numbers. A small model started solving only 12 puzzles out of 5,008. After 30 minutes of training, it hit 85 — seven times better. For the first time ever, it solved a puzzle completely, all six wires in place. A bigger version of the same model broke at the starting line, but the diagnosis is clear: a known configuration bug, not a capacity problem. A $15 test is running now to confirm the fix.

The bet here is real. BFTS is arguing that small focused models, trained with the right objective, beat giant general ones at specialized hard problems. The 7M model is going open-source. The 35M is cooking behind the scenes with a new RL objective — it's stopped copying demonstrations and now plays 3D Tetris against itself, finding wire paths it wouldn't have discovered by imitation. This is the difference between memorizing and reasoning.

Why this matters for us: the next wave of AI isn't about bigger models — it's about the right model, trained right, doing real work in the world, and the open-source gate to that is opening now.