Scientists Just Cut Wasted AI Training Time by 42%

Training an AI model is a lot like baking hundreds of loaves of bread to find one good recipe. Most attempts fail, but you don't find out until the oven timer goes off — and by then you've burned hours of electricity and money. A research team led by Romain Claret, working with universities in Switzerland and New Zealand, wanted to spot those doomed attempts much earlier.

Their idea is simple. Instead of waiting for an AI experiment to finish, they peek at its progress after the first few rounds. If the results still look like random guessing, they stop it right there. Using 90 test runs, they tuned a cutoff point: after round three, if the score hasn't climbed past a certain small number, quit. On 180 fresh runs they hadn't seen before, this rule correctly killed bad runs about 87% of the time, while letting more than 90% of the promising ones continue.

The payoff was real: a 41.6% cut in computing work. Compared with Hyperband — a popular, general-purpose method for the same job — their custom rule was 64% more efficient and produced slightly better results on average, though Hyperband occasionally stumbled onto a star performer they'd have missed.

The honest catch: this rule was tailored to one specific AI-building technique called ES-HyperNEAT (a method where AI designs its own internal wiring), so you can't copy the exact numbers elsewhere. The team also found the rule gets too aggressive once their search has already found good solutions, killing 69% of potentially good runs. Their fix is a smarter, adjustable cutoff — and they believe the underlying approach, learning when to quit from early progress, could work across many kinds of AI research.