cyanydeez
2 hours ago
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
Codex reasoning-token clustering at 516 may be leading to degraded performance
11 pointsposted 6 hours ago
by 0x_rs