Artificiell intelligens från DeepMind kunde nå mänsklig nivå

DeepMind, which develops artificial intelligence, reported that their algorithm for writing AlphaCode was able to achieve good results. Now their AI is at the “novice programmer” stage and continues to improve.

AlphaCode took part in the competition on the Codeforces platform. The program was able to solve previously unknown problems in natural language. At the same time, DeepMind said that the developed AI predicted code segments and created millions of potential options for solving the problem, after which it chose the ten most suitable ones.

The situation researchers reported that the solutions he generated were implemented without any built-in knowledge of the structure of the computer code, which is even more impressive. According to the results of the competition held on Codeforces, the artificial intelligence AlphaCode entered the top 54.3% of the best participants in the past competition. The competition involves restrictions in the form of ten attempts to find the desired solution. The creators of AlphaCode reported that their program could handle the task the first time in 2 out of 3 cases.

DeepMind specifically noted that it is quite difficult to achieve such results and make progress in the coding competition. To solve the necessary problems, the algorithm must understand complex problems in natural language from the very beginning, after which it is necessary to form a process of “reasoning” about unforeseen situations, and then find a solution. The usual memorization of the code in this case will not work, which greatly complicates the situation.

The researchers report that they found no evidence of a normal copying of the base process from the provided training data by the AlphaCode program. Considering all the factors, the performance of artificial intelligence can safely be called a big step forward.

This was told by Zico Colter, one of the professors at Carnegie Mellon University. He decided that AlphaCode did an excellent job with those tasks that were previously unknown to him.

DeepMind are not only proud of their work, but also believe that in the future, such an algorithm can become a great colleague for developers. It can make the programming process more accessible to users and generally increase the overall productivity of such activities.

According to the researchers, such events in the world of coding will be able to take programming to a new level. Raised it high enough that coding can survive a culture shift where humans are less about coding than they are about tasking, while AI takes over the technical part and generates the right code.

AlphaCode has been improved to this level in 10 months. In February, DeepMind first introduced this program to people, which will learn to solve unpredictable problems. In November, the same company created agents that would naturally interact with people. In December, the laboratory also did not sit idly by and trained artificial intelligence to play "Stratego" on a par with professionals.