The 800 mistakes that could reshape Meta’s AI coding strategy

Meta has launched an internal program that asks its software engineers to identify and correct coding mistakes, treating those corrections as training signals for its AI-driven coding assistants. This shifts part of the engineering workflow from pure development to a data-labeling role, creating a feedback loop where human-fixed errors directly inform model updates. The change is internal; it does not involve any new external tooling or public API, but it does reallocate engineering capacity. The cost of participation is the engineers' time spent on fixing code for the purpose of training, which could otherwise be allocated to feature work, bug triage, or performance improvements. Since the program targets "thousands" of engineers, the aggregate impact on productivity could be noticeable across teams. No additional hardware or software purchases are mentioned, so the primary expense is labor. The effectiveness of the approach hinges on the relevance of the corrected code to the AI models. Because the training data comes from Meta’s own codebases, the resulting models are likely to perform best on similar internal projects. When developers work on external libraries, open-source c