How Meta Plans to Close the Gap with Anthropic and OpenAI in Coding

As Meta Platforms races to catch up with Anthropic and OpenAI, it’s asking thousands of engineers to use its in-house AI coding agent, MetaCode, in their day-to-day work in an effort to improve the coding capabilities of its AI models.

Maher Saba, vice president of Meta’s Applied AI Engineering organization, last month asked engineers across the company to write at least one code “diff” a week— a code change that is reviewed and integrated into the models, according to an internal memo reviewed by The Information. When the model gets something wrong, they’re encouraged to fix it and submit those corrections.

Saba said engineers’ use and feedback on MetaCode has already helped boost the coding capabilities of the latest version of Meta's AI model, Muse Spark 1.1, and those same corrections will be used to post-train an upcoming model dubbed “Watermelon” internally.