A postmortem of three recent issues
a day ago
- #Bug Resolution
- #AI Infrastructure
- #Model Quality
- Three infrastructure bugs degraded Claude's response quality between August and early September.
- Bugs included context window routing errors, output corruption, and an XLA:TPU miscompilation.
- Detection was challenging due to overlapping symptoms and privacy constraints.
- Fixes were deployed, including routing logic corrections and rollbacks of problematic changes.
- Anthropic is improving evaluations, monitoring, and debugging tools to prevent future issues.