The AI revolution seems to be here, but so far these systems have not been used, for example, to discover some prodigious medicine or some material with special properties. Real practical applications seemed limited to the creative realm, but at DeepMind he is proving time and time again that these systems can go further.
Heir to AlphaZero. Those responsible for DeepMind had already created AlphaZero to learn to play chess from scratch, but now they have created AlphaDev, a derivative version of this system that had a different objective.
sort lists. Precisely the objective of AlphaDev was to try to discover the best way to order lists of elements. This AI system managed to discover a method that is 70% faster than the algorithms that were traditionally used, something crucial for many areas but which had an immediate impact in one in particular.
benefited programmers. Although the news of this discovery is now published in the prestigious Nature, the technique for ordering elements in a list has been available for months. In January 2022, DeepMind sent its solution to the organization that manages the C++ language, one of the most popular in the world, and after two months of rigorous analysis, this algorithm was included in the language: it was the first change made in that section. in C++ in over a decade.
It also improves encryption. DeepMind also added new discoveries to Abseil, a set of C++ algorithms used in cryptography that allow working with hashes, the unique IDs of any type of data. In DeepMind they believe that their algorithms are used “billions of times a day” and speed up their processing by 30%, something that certainly represents a notable improvement in this section.
The algorithm created by AlphaDev made it possible to save an assembly instruction when ordering lists of three elements. For those with five, the gain was even greater. Source: Nature.
Assembly language returns. To achieve their purpose in DeepMind, they worked as a base with assembly language, a very low-level programming language that allows algorithms to be broken down into small steps that are later useful for finding “shortcuts” and new solutions to problems like the ones you have Found DeepMind. As with AlphaZero and chess, AlphaDev found the solution by “playing millions of games” or, in his case, trying millions of assembly instruction combinations to create even one new instruction for his algorithm.
Less instructions, better. Daniel Mankowitz, from DeepMind, explained in MIT Technology Review how they worked with algorithms to order short lists of three or five elements, widely used in all kinds of developments. His team explained that the best human version of the algorithm for three-element lists was 18 instructions in assembler. AlphaDev managed to create an algorithm with 17 instructions. Sorting lists of five takes 46 instructions in the best human algorithm, and AlphaDev reduced that to 42. It doesn’t sound like that much, but running the algorithm on an Intel Skylake chip went from sorting a list of five in 6.91 nanoseconds. at 2.01 nanoseconds, 70% faster.
limitations. Still, AlphaDev has its limits. The longest algorithm you can write is 130 instructions because when you try to combine more, the number of possible combinations is even greater than the number of possible moves in chess (10^120) or atoms in the universe (about 10^80). ). Even so, it is possible that we will see new advances in this regard as DeepMind works with other algorithms or, as its creators pointed out, directly with languages like C++. The shortcuts may not be as flashy, but being of a higher level these improvements can be applied to other scenarios.
Image | Xataka with Midjourney
In Xataka | AlphaZero has the solution to make chess more fun: no castling
