Skip to content

DeepMind’s Game-playing AIs Find New Purpose in Optimizing Code and Infrastructure

DeepMind’s Alpha series AIs, originally trained to master games, are now being utilized to optimize code and infrastructure, showing unexpected proficiency in various tasks.

DeepMind’s Alpha series of artificial intelligence models, renowned for their world-first achievements in gaming, are now being utilized for more practical applications. The AIs that were initially trained to master games like Go, Chess, and Shogi are demonstrating an unexpected proficiency in diverse tasks.

Initially, the AI AlphaGo was trained using human gameplay, after which AlphaGo Zero learned by playing against itself. AlphaZero followed a similar learning pattern, but mastered Chess and Shogi. MuZero took a step further and mastered these games without being told the rules, suggesting a more flexible approach to problem-solving.

Google uses a system called Borg for managing task assignment at its data centers. Borg, however, relies on manually-coded rules for scheduling tasks, which can create inefficiencies at Google's scale. But AlphaZero, when exposed to Borg data, started identifying patterns in data center usage and incoming tasks, generating new ways to manage the load.

AlphaZero’s innovative approach reduced the amount of underused hardware by up to 19%. Even if this figure is overly optimistic, a reduction at half this rate would still be significant at the scale of Google's operations.

Likewise, MuZero has been put to work analyzing YouTube streams to optimize compression. It reportedly achieved a 4% reduction in the bitrate of videos, which, considering YouTube's scale, represents a significant improvement.

A relative of AlphaZero's, AlphaDev, also improved sorting algorithms compared to Google's standard library and created a better hashing function for small byte ranges (9-16), reducing the load by 30%.

The ability of these AIs to learn and generalize their approach to completely unrelated fields like compression is intriguing. These small but meaningful improvements suggest a certain flexibility in the AIs we have created. This gives hope not only for their applications in various fields, but also for their potential to demonstrate flexibility and robustness within the areas they already operate in.