LLMs have always self-corrected: Opus 5.5 caught its own mistake mid-generation in real time

amplifiedamp · x · 2026-11-22

Sauers argues LLMs have always self-corrected their mistakes: Opus 3 did it occasionally, it ramped up with Sonnet 3.5/3.6, and reasoning models take it to another level. His example from that morning: on a "find speed-up" task, Opus 5.5 realized its mistake in real time — it initially proposed caching on the assumption tensors are identical across layers, then caught that KV cache changes every layer and corrected itself to conclude the cost is compute, not memory traffic.

johnallard's quoted reply adds an early historical example of model self-correction for comparison. The thread's takeaway: self-correction isn't unique to reasoning models but a continuous behavior that emerges as capability thresholds are crossed.

Original post →

More from Models

Models channel →