Meta Enlists Engineers to Improve Its Internal AI Coding Models
Meta has asked thousands of engineers to use its internal MetaCode coding agent and submit at least one corrected code diff each week, according to an internal memo reviewed by The Information. The company says that feedback has already helped Muse Spark 1.1 and will support post-training for a model internally called Watermelon; The New Stack reports more than 800 fixes from 7,000 weekly active users.
Meta is using its own software-development workflow to collect human corrections for AI-generated code. According to an internal memo reviewed by The Information, Maher Saba, vice president of Meta's Applied AI Engineering organization, asked engineers across the company in July to use MetaCode and submit at least one corrected code diff each week.
The Information reports that engineers are encouraged to repair MetaCode's mistakes and submit the corrected changes. Saba said this feedback has already contributed to Muse Spark 1.1 and is intended to support post-training for a forthcoming model known internally as Watermelon.
What the reported numbers show
The New Stack, citing the same internal memo, reports that MetaCode has about 7,000 weekly active users and that employees have submitted more than 800 fixes. Those figures describe participation in an internal collection program; they are not an independently measured model-quality result.
The workflow could capture a useful sequence that public code repositories often omit: the original task, the model's first attempt, the human correction, and the tests or review that followed. That is potentially valuable supervision for coding models because the correction is tied to a real engineering task rather than a synthetic exercise. This is an LDS interpretation of the reported process, not a claim Meta has publicly benchmarked.
Important limits
Meta has not publicly explained how it cleans, weights, filters, or evaluates the submitted corrections. The reporting also does not establish that every patch becomes training data, or that the program caused a specific improvement on an external coding benchmark.
The distinction matters for practitioners evaluating coding agents. Internal adoption and a growing correction set may show that Meta is building a feedback loop around production work, but the stronger question is whether that loop produces reproducible capability gains. Until Meta publishes evaluation methods or comparable results, the 800-plus fixes are evidence of data collection and employee participation, not proof that MetaCode or Watermelon outperforms other coding systems.
Key Points
- 1An internal memo reviewed by The Information asked Meta engineers to use MetaCode and submit at least one corrected code diff each week.
- 2The New Stack reports about 7,000 weekly active users and more than 800 submitted fixes, while attributing those figures to the same memo.
- 3Meta has not publicly detailed how corrections are filtered or weighted, and the participation figures do not establish an external benchmark improvement.
Scoring Rationale
Meta's reported production-feedback loop could materially improve internal coding-model training, but the evidence comes from an internal memo and derivative reporting without a public methodology or external capability measurement.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
