DeepMind developed the AlphaCode neural network

DeepMind created the AlphaCode artificial intelligence platform, which is not only partially capable of programming instead of a person, but also ready to do it at a “competitive” level. A subsidiary of Alphabet tested the system on problems used in human programming competitions. It turned out that according to the test results, AI is among the 54% who coped with the tasks best of all. The result was a big step forward towards creating standalone programming systems, although AlphaCode's skills are not necessarily up to the task of the average coder.

According to DeepMind, research is still at an early stage, but it is already clear that the program can autonomously solve programming tasks that until recently were available only to people. The company expects that in the long term, AI will help developers write code, improve employee productivity and provide new ways to develop software. AlphaCode was tested on the Codeforces platform, which publishes weekly programming tasks and programmer ratings. The tasks are different from those that a programmer might face when creating, for example, a commercial application. They are more "self-sufficient" and require extensive knowledge of both algorithms and theoretical computer science concepts. To solve them, a combined approach is required, since the programmer needs to understand logic, mathematics, and programming itself.

One example is a task in which you need to find a way to convert one string of randomly repeated letters s and t into another string of the same characters using a limited number of keystrokes. At the same time, contestants cannot, for example, simply type new letters instead of old ones - instead, you need to use the Backspace key and delete several letters from the original string. The task refers to examples of medium complexity, on the left of the image contains its description, on the right - samples.

Ten similar tasks were assigned to AlphaCode to perform in the same way that humans are assigned to perform them. The AI ​​generated a number of possible solutions and weeded out the unsuitable ones by running the code and checking the result, just as a programmer would. The 10 tasks proposed by AplphaCode were performed by 5,000 programmers on the Codeforces website. AI work was in the top 54.3% of answers, and according to DeepMind, this provides AlphaCode with a Codeforces Elo rating of 1238. Thus, AI was among the top 28% of programmers who competed on the site over the past six months. DeepMind notes that AlphaCode's current skills are only applicable to programming competitions, but in theory, the new system allows you to create tools that can make programming more accessible and, someday, fully automated.

Many other companies are known to be working on similar solutions. Much progress has been made in recent years, but all these systems are far from taking the place of human programmers. Released AI code is often full of bugs, and since such systems are usually trained on public code libraries, they sometimes play fragments that are protected by copyright and related rights.

Moreover, one study found that the Copilot program, developed by the GitHub code repository, produced code that contained 40% of vulnerabilities. Information security experts have even suggested that attackers may deliberately write and share code with backdoors, which will later be used to train AI systems, as a result of which they are doomed to make mistakes in their future decisions.

These problems mean that AI programming systems are likely to be slowly integrated into the work of regular programmers until they "deserve" trust. In other words, they have yet to learn and learn. But they do it very quickly.