DeepMind has developed AlphaCode to write competent computer programs. AlphaCode can solve problems that require a combination of critical thinking, logic, algorithms, coding, and natural language understanding and was able to achieve an estimated rank within the top 54% of participants in programming competitions.
AlphaCode uses transformer-based language models to generate code on a massive scale and then intelligently filters it to a concise set of promising programs.
AlphaCode was trained using data from 10 recent contests hosted by Codeforces, a competitive programming platform. AlphaCode was able to achieve the level of a median competitor, making it the first time an AI code generation system has reached a competitive level of performance in programming competitions.
Existing AI systems lack the capability to participate in competitive programming. However, with the recent advancement in large-scale transformer models combined with large-scale sampling and filtering, systems like AlphaCode are able to solve more problems.
Alphacode ranking within the top 54% in real-world programming competitions shows the potential deep learning models have for tasks that require critical thinking. The models elegantly leverage modern machine learning to express solutions to problems as code, circling back to the symbolic reasoning root of AI from decades ago.
