GHOSTS IN THE CODEX MACHINE

Summary from the TLDR AI newsletter

After weeks of user reports claiming Codex had gotten worse, echoing a perennial concern that frontier labs secretly downgrade their models, OpenAI assembled a squad to investigate every possible culprit and publish their findings. The team found no smoking gun, but instead uncovered a catalog of small bugs: older hardware subtly dragging down performance, a constrained sampling bug that made the model switch to Korean mid-sentence, and context compaction that quietly degraded accuracy in long conversations.