Meta put LLM-written just-in-time tests into production: 589 engineers used them, 60 percent of the catches were real, and 74 serious privacy bugs died before shipping
AspireAI community · Meta Engineering
Meta's approach to catching bugs before code lands is not an AI reviewer that reads a diff and forms an opinion. It writes executable tests that demonstrate a concrete behavioural difference, which is a harder thing to produce and a much easier thing to act on. Two workflows generate them: one infers what a change was trying to do and targets the risky parts, the other pretends every diff is a bug and hunts for the difference from its parent, without reasoning about intent at all.
The operational part is what makes the numbers credible. False positives are filtered by a panel of judge models and by a deterministic rule-based checker, and humans stay in the loop at a different point than usual: engineers set criteria and review outcomes in aggregate rather than supervising every individual result. Meta says 589 engineers have engaged with the system so far, that 60% of the catches they saw were true positives, and that in the first half of 2026 the system helped catch 74 serious privacy bugs at the diff stage, lowering privacy SEVs by 20% over that period. The research behind it was presented at FSE 2026.
The caveat is that all of this is Meta reporting on its own deployment, and the 60% figure is the judgement of engineers who chose to look, not an audited measurement across the whole codebase. The post also says nothing about how the system behaves on codebases smaller than Meta's, where the overhead of generating and running these tests may not pay for itself.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.