AI
Meta enlists 7,000 engineers to train AI coder by crowdsourcing 800+ bug fixes
Inside Meta's strategy to close the gap with rivals by turning routine code reviews into training data for its internal AI model.
The New Stack
Meta is mobilizing its workforce as a distributed training pipeline for its internal AI coding tool. The company asked thousands of engineers to contribute code fixes from their daily work, allowing Meta to harvest mistakes and corrections as raw material for improving its AI models.
In July, Maher Saba, vice president of Meta's Applied AI Engineering organization, issued an internal call asking engineers to submit at least one code correction per week using MetaCode, the company's proprietary coding agent. Early results showed promise: 7,000 active users contributed over 800 fixes, which are already being used to refine Muse Spark 1.1 and train an upcoming model codenamed Watermelon. To boost participation, Meta added achievement badges to employee profiles tied to the number of contributions.
The hidden goldmine in code reviews#
The appeal of this approach lies in what production engineering work reveals—something that doesn't exist in public datasets or synthetic benchmarks. Public code repositories show finished products but omit the AI's initial attempt, its failures, and the reasoning behind the engineer's corrections. MetaCode captures this entire arc: the original task request, the model's output, the engineer's fix, and the testing and approval process that followed.
Rather than create artificial training exercises, Meta can extract genuine learning signals from thousands of real coding problems the AI gets wrong during normal development. If patterns emerge across hundreds of corrections, the company gains insights into systemic gaps that might otherwise stay hidden. The memo doesn't specify how Meta weights these submissions or whether all patches automatically flow into training—only that the process is intentional and tied to submitted corrections.
Racing rivals on public benchmarks#
Meta's internal push reflects a performance lag that public leaderboards have made impossible to ignore. Muse Spark 1.1, released in July 2026, scored 53% on the DeepSWE 1.1 benchmark, which evaluates coding agents on 113 long-running software engineering tasks. At launch, competitors scored significantly higher: GPT-5.5 at 67% and Claude Opus 4.8 at 59%.
The gap only widened. OpenAI's GPT-5.6 Sol reached 73%, and Anthropic's Claude Opus 5 climbed to 74% in late July. While these benchmarks test the general models rather than MetaCode specifically, they spotlight Meta's underlying problem—rivals continue to pull ahead while Meta's model remains competitive but not dominant.
