DeepMind, perhaps the global leader in artificial intelligence recently announced their new coding engine Alpha Code which they say writes computer programs at a competitive level. The creative minds behind Alpha Go (AI plays Go) and Alpha Zero (AI plays Chess) are now one step closer to building an engine that can autonomously code.
Software engineers however need not worry engine is not taking over any jobs just yet. DeepMind tested the system against coding challenges in human competitions where it achieved an impressive “estimated rank” within the top 54% of human coders. It is also important to note that Alpha Code’s skills are not representative of those required of an average programmer.
Instead of the normal tasks required in say, developing a commercial app, the system was tested against challenges curated by Codeforces. This is a platform that shares weekly problems and issues rankings to coders who attempt to build solutions for these problems, quite similar to the Elo rating system in chess.
The challenges here are a little different from those in the “commercial sector” but just as applicable. To solve them however requires a wider and deeper knowledge of the theoretical concepts in computer science and algorithms. One might think of them as very specialized puzzles that combine math, logic and coding expertise.
To show Alpha Code’s proficiency DeepMind highlighted a challenge that Alpha Code was able to solve. These solutions were released on Github and thus available for anyone to check out. Here is a brief look at one of these tests.
It is important to note here that these challenges were fed into Alpha Code in exactly the same format they are given to humans. The system tacked these challenges by generating a large number of possible answers and then narrowing them down to just a few by checking the output just as a human competitor would.
“The whole process is automatic, without human selection of the best samples,” Yujia Li and David Choi, co-leads of the AlphaCode paper, told The Verge over email.
There are also other companies working on similar applications. Microsoft for example and AI lab OpenAI have adapted the latter’s language-generating program GPT-3 to function as an autocomplete program that finishes strings of code. The system works the same as Gmails’ Smart Compose feature by suggesting ways to finish whatever you’re writing.
[…] The team built OPT to match GPT-3 both in its accuracy on language tasks and in its toxicity. OpenAI has made GPT-3 available as a paid service but has not shared the model itself or its code. The idea […]