Artificial intelligence will soon be able to write codes based on ordinary languages

Last few years, Researchers use artificial intelligence to Thank ss Between programming languages ​​or automatically Solve the problem. For example, the artificial intelligence system DrRepair has been proven to solve most problems that produce error messages. But some researchers dream that one day artificial intelligence can write programs based on simple descriptions from non-experts.

Tuesday, Microsoft with OpenAI The sharing plan introduces GPT-3 (one of the most advanced text generation models in the world) into programming based on natural language descriptions. This is the first commercial application of GPT-3 since Microsoft invested $1 billion in OpenAI last year and obtained the exclusive license for GPT-3.

“If you can describe what you want to do in natural language, GPT-3 will generate a list of the most relevant formulas for you to choose from,” said Microsoft CEO Satya Nadella in a keynote speech at the company’s Build developer conference. “Write the code yourself.”

Microsoft vice president Charles Lamanna told WIRED that the complexity provided by GPT-3 can help people deal with complex challenges and provide support for people with little coding experience. GPT-3 translates natural language into PowerFx, which is a fairly simple programming language similar to the Excel command Microsoft launched in March.

This is the latest demonstration of applying AI to coding. Last year at Microsoft’s Build conference, OpenAI CEO Sam Altman Demo A language model that uses GitHub code for fine-tuning and automatically generates Python code lines.As WIRED detailed last month, startups like SourceAI are also using GPT-3 generated code. IBM demonstrated its Project CodeNet last month, which has 14 million code samples from more than 50 programming languages, which can reduce the time required for car companies to update programs with millions of lines of Java code from one year to one month .

Microsoft’s new features are based on Neural Networks The architecture called Transformer is used by large technology companies, including Baidu, Google, Microsoft, NvidiaAnd Salesforce use text training data crawled from the web to create large-scale language models. These language models keep getting bigger. The largest version of Google BERT is the language model released in 2018. It has 340 million parameters and is the building block of neural networks. The GPT-3 released a year ago has 175 billion parameters.

However, such efforts still have a long way to go. In a recent test, the best model had a 14% success rate in an introductory programming challenge compiled by a group of AI researchers.