A chip that learned to play Connect-4 now exists, trained end-to-end for under $10

From data to silicon: play Connect-4 against a chip

Benoit Rostykus argues that extreme hardware specialization is closer than we think: if custom silicon becomes cheap to design and manufacture, why not print an ASIC for every recurring computation? His proof of concept is a single-cycle Connect-4 chip with roughly 14k logic gates, learned end-to-end from simulated games for under $10 of compute. On a two-decade-old 130 nm process it measures 0.8×0.8 mm, runs at 25 MHz, and burns a few hundred picojoules per move — matching a modern CPU engine while using two to three orders of magnitude less energy.

In other words, learning an AI model where the model is the silicon?

More from this day

2026-11-19