I wrapped this small followup to stolen-thoughts.com in a gist for easier reading and wanted to share it here.

My hunch is that this might not just be reasoning distillation; it could even be benchmark distillation. This is a small experiment and definitely doesn’t prove anything about how any model was trained, but I thought the contrast was interesting enough to share.

I’d also like to try glm-5.3, but it isn’t fully available across the model-serving platforms I use yet. If I get around to running it and the results are interesting, I’ll share them later.