Skip to main content

Key Points

  • DeepMind’s AlphaDev, an AI based on AlphaZero, allegedly creates algorithms that sort data up to three times faster than human-written ones.
  • Critics argue that the supposed innovations are not novel and that DeepMind overstates the significance of the results.
  • Despite DeepMind’s claims, AlphaDev’s improvements only apply to a specific scenario and are not universally applicable.

Google’s DeepMind recently developed AlphaDev, a derivative of its AlphaZero algorithm, claimed to create sorting algorithms that outperform human-written ones in speed. Critics, however, challenge the purported novelty and significance of these findings, suggesting that the touted achievements might not represent an authentic advancement in the machine learning sector. This skepticism arises amidst perceptions that DeepMind has been overshadowed by other technology powerhouses in the realm of AI innovation.

Understanding AlphaDev

The team at DeepMind structured the sorting algorithm discovery as a single-player game, enabling the deployment of their previous experiences designing AI agents for gaming applications. The AlphaDev agent, akin to its predecessor, AlphaZero, selects instructions for a sorting process, instead of possible moves in a game. Interestingly, AlphaDev performs these operations at the assembly level, akin to C++ programming, the go-to language for sorting algorithms. The entire process is driven by a reward system based on the success of the sorting algorithm, continuing until it perfectly sorts all lists.

Performance Analysis

As per DeepMind’s reports, AlphaDev outperforms human-written algorithms by up to 71% in its best scenarios. Despite the impressive initial performance, the time-saving advantage dwindles to 1-2% when applied to larger values, attributed to unoptimized code elsewhere in the program. While industry experts acknowledge AlphaDev’s code-generating capability, they argue that it brings no groundbreaking innovation to the field.

Internet Reactions

Discussions online indicate a significant scepticism towards AlphaDev, with some deeming it as an overhyped non-news in the AI community. One user argued that the reported improvements were largely due to the inefficiency of a ten-year-old library algorithm, which AlphaDev leveraged. Contrary opinions highlighted AlphaDev’s ability to independently create a new sorting algorithm. However, the consensus remains that the development, though noteworthy, is not revolutionary.

The Real Strength of AlphaDev

AlphaDev’s expertise manifests in code assembly, locating missing optimizations, and maintaining a balance between correctness and distribution for hashing algorithms. Rather than unveiling an entirely new approach to sorting, AlphaDev micro-optimizes code. Comparable capabilities already exist in the form of super optimizers, which find optimal programs, typically via brute force.

While AlphaDev’s approach to the coding process marks a distinct application of AlphaZero, its impact appears overestimated. Its capabilities are context-specific and don’t necessarily apply to the diverse scenarios encountered in the computing world.