DeepMind has created an AI that can code, and is already pretty good at it

Machines can do more and more jobs for us: today we have dedicated an in-depth study to the topic. And among the professions within reach of AI there is also that of the programmer. DeepMindthe Alphabet company specializing in research related to Artificial Intelligence, has in fact created a system capable of “write computer programs at a competitive level. In short, a program that knows how to program.

The name of this project is AlphaCodeand Deep Mind has already tested their skills for the first time, comparing them with those of human programmers, and then sharing the results. So how does AI cope with programming? Already pretty good, but not yet very well: this, in summary, is the result of the very first tests conducted.

It might sound disappointing to you, but it isn’t: for developers, AlphaCode’s capabilities are surprising considering that research is only in its embryonic stages, and are proof that in the long term it is indeed possible to develop an important resource for (first and foremost ) help programmers but also non-programmers to write code, improving productivity and creating new approaches and methods of creating software.

AlphaCode’s prowess was tested by subjecting the AI ​​to the challenges by Codeforces, a competitive coding platform that shares weekly issues and then ranks the participants’ results with an Elo system, such as the one used in chess. Ten tests were fed to AlphaCode in the exact same format in which they were formulated for humans: the AI ​​then generated a plurality of possible answers, sifting through them then executing the code and checking the output, exactly as a human competitor.

On the same tests addressed by AlphaCode, 5,000 users tried their hand, and the DeepMind program was placed in the top 54.3%, the equivalent of an Elo Codeforces equal to 1238.


Leave a Comment