Humans need to use programming to solve a wide variety of problems.
While some recent AIs have emerged that can generate code, AI still does not perform well for real-world complex programming problems that require critical thinking.
AlphaCode, an AI-assisted programming system announced by DeepMind, uses self-supervised learning and a transformer model to solve programming problems that it has never seen before, described in natural language.
The researchers tested AlphaCode with the Codeforces platform, which itself regularly hosts international programming competitions.
When the number of answer submissions was limited to 10, AlphaCode could achieve a score of 1,238
It ranked in the top 54.3% of human participants, which is in the middle of the pack for human participants.
And 66% of the problems it solved were solved on the first submission.
Regardless of whether AlphaCode can actually “understand” programming problems, it does rank above the human average in code competitions.
Yesterday, the paper was officially published as a cover article in the journal Science.