DeepMind compares AlphaDev’s discovery to considered one of AlphaGo’s weird but winning moves in its Go match against grandmaster Lee Sedol in 2016. “All of the experts checked out this move and said, ‘This isn’t the appropriate thing to do. It is a poor move,’” says Mankowitz. “But actually it was the appropriate move, and AlphaGo ended up not only winning the sport but additionally influencing the strategies that skilled Go players began using.”

Sanders is impressed, but he doesn’t think the outcomes must be oversold. “I agree that machine-learning techniques are increasingly a game-changer in programming, and everybody is expecting that AIs will soon find a way to invent latest, higher algorithms,” he says. “But we usually are not quite there yet.”

For one thing, Sanders points out that AlphaDev only uses a subset of the instructions available in assembly. Many existing sorting algorithms use instructions that AlphaDev didn’t try, he says. This makes it harder to check AlphaDev with the perfect rival approaches.

It’s true that AlphaDev has its limits. The longest algorithm it produced was 130 instructions long, for sorting a listing of as much as five items. At each step, AlphaDev picked from 297 possible assembly instructions (out of many more). “Beyond 297 instructions and assembly games of greater than 130 instructions long, learning became slow,” says Mankowitz.

That’s because even with 297 instructions (or game moves), the variety of possible algorithms AlphaDev could construct is larger than the possible variety of games in chess (10¹²⁰) and the variety of atoms within the universe (around 10⁸⁰).

For longer algorithms, the team plans to adapt AlphaDev to work with C++ instructions as a substitute of assembly. With less fine-grained control AlphaDev might miss certain shortcuts, however the approach can be applicable to a wider range of algorithms.

Sanders would also wish to see a more exhaustive comparison with the perfect human-devised approaches, especially for longer algorithms. DeepMind says that’s a part of its plan. Mankowitz desires to mix AlphaDev with the perfect human-devised methods, getting the AI to construct on human intuition reasonably than ranging from scratch.

In any case, there could also be more speed-ups to be found. “For a human to do that, it requires significant expertise and an enormous amount of hours—perhaps days, perhaps weeks—to glance through these programs and discover improvements,” says Mankowitz. “Consequently, it hasn’t been attempted before.”