Codex reasoning-token clustering at 516 may be leading to degraded performance 12 points · 1 visible top comment · 2026-07-01 17:03:11 UTC github.com · Read Story HN original
cyanydeez · 2026-07-01 21:24:00 UTC wonder if theyre basically doing this llamacpp reasoning trick: https://github.com/ggml-org/llama.cpp/blob/master/common/rea...id guess its the harness
Comments
id guess its the harness