DeepMind, the subsidiary of Alphabet (formerly Google), specialized in AI presented AlphaCode, a tool capable of writing computer code in a completely autonomous way to solve problems requiring logical thinking and understanding of language. It has confronted this new code generation system to human developers during programming contests on the Codeforces platform where it obtained a score of 54.3%.
DeepMind Technologies, a British company specializing in artificial intelligence, founded in 2010 by Demis Hassabis, Mustafa Suleyman and Shane Legg was acquired in 2014, by Google for more than 628 million dollars.It aims to find “the best techniques from machine learning and systems neuroscience to build powerful general-purpose learning algorithms. Renowned for its Alphago program, which beat several Go champions before achieving a definitive victory over the world’s No. 1 in 2017, or AlphaFold, which can predict the structure of proteins from their amino acid sequences, DeepMind works on rational and predictive reasoning, as well as on deep learning and deep reinforcement learning methods.
Competitive programming with AlphaCode
The development of machine learning models similar to AlphaCode allows code to be generated from algorithms. Several companies have created code auto-completion tools: Microsoft and Open AI have partnered to produce Github Copilot, based on Codex, TabNine has named its own. The objective is to allow developers to gain productivity in an increasingly complex activity. It is precisely to provide “solutions to competitive programming problems that require an understanding of algorithms and complex natural language“ that Deepmind created AlphaCode.
AlphaCode uses language models based on large sample-efficient transformers to generate code at large scale, then filters to a small set of programs. Deepmind has evaluated and validated the performance of its model in competitions on Codeforces, a popular platform where tens of thousands of developers from around the world come to test their coding skills. AlphaCode competed against 5,000 competitors in 10 contests, achieved a score of 54.3% and placed about the same as the median competitor, marking the first time an AI code generation system has achieved a competitive level of performance in programming competitions. Petr Mitrichchev, software engineer, google and world-class competitive programmer said:
“Solving competitive programming problems is a very difficult thing to do, requiring both good coding skills and creative problem solving in humans. I was very impressed that AlphaCode could make progress in this area and delighted to see how the model uses its understanding of statements to generate code and guide its random exploration to create solutions.“
Mike Mirzayanov, founder of Codeforces, added:
“I can safely say that the results of AlphaCode have exceeded my expectations. I was skeptical because even in simple competitive problems you often have to not only implement the algorithm, but also (and this is the hardest part) invent it. AlphaCode has managed to perform at the level of a promising new competitor. I can’t wait to see what’s in store!“
An AI in its infancy
According to DeepMind, this AI is still in its infancy, but “Our exploration of code generation leaves vast room for improvement and hints at even more interesting ideas that could help programmers improve their productivity and open up the field to people who aren’t currently writing code. We will continue this exploration…”.
For now, AlphaCode has been evaluated in competitive programming but does not yet have the skills of a human developer. On the other hand, like other code generators, it is based on existing code, so could be accused of plagiarism.
Translated from AlphaCode, générateur de code de Deepmind, évalué sur Codeforces
