DeepMind’s Alpha series of AIs, known for their breakthrough achievements in gaming, are now making strides in other domains, showcasing unexpected versatility.
AlphaGo, the AI that conquered the game of Go, was initially trained using human gameplay. Its successor, AlphaGo Zero, surpassed its predecessor by solely playing against itself to learn. AlphaZero further expanded its mastery to include Chess and Shogi. Then came MuZero, which accomplished all of that and more without any prior knowledge of game rules, offering a unique perspective on problem-solving.
At Google, Borg, a system responsible for managing task assignment in data centers, relies on manually-coded rules for scheduling tasks. However, the ever-changing workload distributions at Google’s scale make it challenging to optimize efficiency using traditional methods.
Enter AlphaZero. Exposed to Borg data, it began identifying patterns in data center usage and incoming tasks, devising new ways to predict and manage the workload. When deployed, it achieved remarkable results, reducing underused hardware by up to 19% at Google’s scale.
Similarly, MuZero ventured into the realm of YouTube streams, aiming to optimize video compression. This complex software domain requires meticulous optimizations to yield significant results. Impressively, MuZero managed to reduce the bitrate of videos by 4%, a noteworthy improvement considering the vast scale of YouTube.
These developments may not single-handedly revolutionize the world, as incremental changes are continually being made in developer systems. However, the intriguing aspect lies in the ability of AI models originally trained for gaming to adapt and generalize their problem-solving methods to entirely unrelated fields such as compression.
While achieving “general-purpose AI” is still a distant goal, the flexibility and robustness demonstrated by the existing AI models offer promising glimpses into their potential. These advancements not only enable cross-domain applications but also highlight the adaptability and effectiveness of AI within their established domains.