Google Has a Secret Project That Is Using AI to Write and Fix Code

  • Google is working on a tool that will teach code to write and rewrite itself.
  • The project originated in the company’s Moonshot Unit X and moved to Google Labs this year.
  • It’s part of a broader push in the field of generative AI.

Google is working on a secret project that uses machine learning to train code to write, fix, and update itself.

This project is part of a broader push by Google into “generative AI,” which uses algorithms to create images, videos, code, and more. It could have profound implications for the future of the company and the developers writing code.

The project, which started at Alphabet’s “X” research unit and codenamed Pitchfork, moved into the Google Labs group this summer, according to people familiar with the matter. By moving to Google, it signaled its increasing importance to executives. Google Labs tracks long-term bets, including virtual and augmented reality projects.

Pitchfork is now part of a new group at Labs called the AI ​​Developer Assistance Team, led by Olivia Hatalsky, a longtime X collaborator who has worked on Google Glass and several other Moonshot projects. Hatalsky, who ran Pitchfork at X, joined Labs when it migrated this summer.

Pitchfork was designed to “teach code to write and rewrite itself,” according to internal materials viewed by Insider. According to people familiar with it and patents reviewed by Insider, the tool is designed to learn programming styles and write new code based on those learnings.

“The team works closely with the research team. They are working together to explore different use cases to help developers,” a Google spokesman said.

Pitchfork’s initial goal was to create a tool that could update Google’s Python programming language codebase to a newer version, a Google spokesperson confirmed. “The idea was, how do we go from one release to the next without hiring all these software engineers?” said one person familiar with the early stages of the project.

The project’s goals shifted over time to a general purpose system that could reduce the need for humans to write and update code while maintaining code quality. In job postings for X late last year, Hatalsky said she works on a team “that is building the future of software engineering.”

Employees who spoke to Insider did so on condition of anonymity as they were not allowed to speak to the press. Her identity is known to insiders.

The generative AI boom

Google and other tech companies have already made great strides in “generative AI.”

GitHub, owned by Microsoft, has launched a tool called Copilot that suggests code snippets and features as developers type. Developers use Copilot to generate up to 40% of their code, and GitHub expects that number to double within five years, Bloomberg reported earlier this month.

Google is working on several other AI code projects. Subsidiary DeepMind has a system called AlphaCode that uses AI to generate code, but currently focuses on competitive coding or writing programs at a competitive level.

Google is also working on a tool similar to GitHub’s Copilot that uses machine learning to generate code snippet suggestions as developers type. Douglas Eck, Google’s senior research director, said at an event in New York earlier this month that the tool improved coding iteration times by 6% for Google employees who used it.

Google’s AI developer support program goes even further by training systems to do more work on their own. The project is still in its infancy, and Google still has tricky ethical considerations to make about how these models will be trained, such as: B. Prejudice and potential copyright issues.

Earlier this month, a class-action lawsuit was filed against GitHub, alleging that the Copilot tool committed “software privacy on an unprecedented scale” by using AI to reproduce open-source code, The Verge reported.

Are you a Google employee and have more to share? Do you have a tip? Contact reporter Hugh Langley at [email protected] or via the encrypted messaging apps Signal and Telegram at +1 (628) 228-1836.