The 800 mistakes that could reshape Meta’s AI coding strategy
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
Meta is leveraging its internal AI coding agent, MetaCode, to collect real-world training data by having engineers submit code corrections when the tool makes mistakes. Over 7,000 weekly active users have already submitted more than 800 fixes, which are being used to post-train upcoming models like Watermelon. This approach captures the full correction loop — original task, AI response, engineer fix, and review — data that public repositories rarely contain. Meanwhile, Meta's Muse Spark 1.1 scores 53% on the DeepSWE leaderboard, trailing GPT-5.6 Sol at 73% and Claude Opus 5 at 74%. Cost is also a factor: Muse Spark is cheaper than frontier competitors, but those competitors are rapidly cutting prices. Meta's strategy bets that production coding mistakes are a better training signal than synthetic benchmarks.
