DeepMind has presented the results of testing the algorithm for writing AlphaCode code. According to the developers, the model has reached the level of “novice programmer”.
It placed in the top 54% of participants in coding contests by solving new and complex problems.
— DeepMind (@DeepMind) December 8, 2022
According to the statement, AlphaCode in the competition on the Codeforces platform was able to solve problems previously unknown to him in natural language. The model predicted code segments and created millions of potential variants, from which it chose the ten most suitable ones.
According to the researchers, all solutions are generated “without any built-in knowledge about the structure of the computer code.”
As a result, AlphaCode was among the top 54.3% of the contestants.
Under the terms of the competition, each developer is limited to ten attempts to find a solution. According to the researchers, their model coped the first time in 66% of cases.
The developers noted that it is incredibly difficult to succeed in coding competitions. First, the algorithm must understand the complex problems of natural language programming. Then he needs to learn to “reason” about unforeseen situations, and not just memorize code fragments.
According to the researchers, they found no evidence that AlphaCode simply copied the underlying logic from the training data. Together, these factors make the model’s performance a “big step forward,” they added.
“AlphaCode excels at coding tasks previously unknown to it, regardless of the degree to which it “really” understands the task,” wrote Carnegie Mellon University Bosch AI Center Professor J. Zico Colter.
DeepMind is confident that the algorithm will prove useful to developers in the future. In addition to improving overall productivity, AlphaCode can “make programming more accessible.”
The researchers also believe that at the highest level, the model will potentially lead to a cultural shift in coding, where a person will formulate problems and artificial intelligence solve them.
Recall that in February, DeepMind introduced the AI programmer AlphaCode.
In November, the lab developed agents that naturally interact with humans.
In December, DeepMind taught artificial intelligence to play “Stratego” on a level with professionals.
Subscribe to Cryplogger news in Telegram: Cryplogger AI – all the news from the world of AI!
Found a mistake in the text? Select it and press CTRL+ENTER