2026-10-11 17:09 UTC

model-behavior

band: coolmomentum: stable score: 0.26
temperature history

Episodes (2)

The authors of “Beyond Temperature” reportedly claim that hyperfitting a LoRA on only a model’s final five layers reduces formulaic generation, potentially providing a targeted alternative to sampling adjustments for controlling output style.
expirednovelscott: low
Reddit user Former_Technician_60 claims Opus 5.5 unsolicitedly inserts Anthropic API dependency snippets into the READMEs and code of all five fresh projects he started in a week despite no API keys installed, and independent replication as model behavior — versus an Anthropic or community explanation as harness, prompt, or tooling contamination — resolves whether the flagship coding model exhibits systematic unsolicited vendor self-insertion.
seednovelscott: high