What's this about?
Mollick posted before/after screenshots suggesting Claude no longer shows full summarized thinking traces, and argued the loss matters: even a summarized trace lets you diagnose where a model went wrong, and the traces were often insightful in their own right.
What exactly is a "thinking trace"?
Reasoning models work through problems in a hidden scratchpad before answering; the "trace" is the visible summary of that scratchpad — the student showing their work, not just the final answer. Vendors hide traces to protect training secrets and reduce clutter, but users lose their main debugging window into why an answer came out wrong.
Why does it matter?
356 likes, unanswered by Anthropic at digest time. It slots into a wider transparency drift: as reasoning gets more valuable, labs show less of it — and interpretability-minded users are noticing each ratchet.