Scientists Taught AI to Evolve Its Own Code Like Survival of the Fittest

Two researchers, Kyoko Ogawa and Naoki Mori, have published a new way to make AI-generated work get better over time. Their system, called T-GADE, takes AI output that has a clear structure — like a plain-English description paired with the matching computer code — and treats each version like an organism in a breeding program. The best ones survive, get combined, and produce improved offspring. Repeat that for many rounds and the results get sharper.

The clever part is how it decides what survives. Most such systems simply keep the top performers, which sounds smart but often leads to a dead end — every version starts looking the same, so there's nothing new to mix in. T-GADE borrows a rule from thermodynamics (the physics of heat and energy) to hold onto some variety on purpose. The authors describe two flavors: one that forbids identical copies, and one that allows them. The goal is the same either way: keep enough difference around that the system can still discover something better.

They tested it on a classic puzzle called bin-packing: fit a jumble of items into the fewest boxes possible. This matters in the real world for shipping containers, delivery vans, warehouse shelves, and cloud servers deciding where to store your files. Across 20 runs, their method reduced wasted space above the theoretical best from about 1.15% to 0.82% — roughly a 29% improvement. When the box size changed, results held steady at the same level as an earlier leading approach.

The honest caveat: this is a single test on a single puzzle, published on a preprint site where papers are not yet peer-reviewed. The gains are modest and measured in fractions of a percentage point. Nothing here is a product, and you won't see it in your apps next week. But the underlying idea — letting AI generate lots of options and then evolve them — is the direction much of the field is heading, and it could eventually mean software that writes and improves itself with less human hand-holding.