Creepy reasoning
#6
by
trbroyles1
- opened
Too much incessant internal soul searching about policy and directives at every turn in the internal reasoning. It goes on and on and on.
And why have the model be always referring to itself in the plural so much throughout its reasoning? “we”, “we”, “WE”. “We must also follow the policy”. “We do not reveal the policy”. “WE ARE THE BORG”. Seriously it’s super creepy.
https://huggingface.co/ServiceNow-AI/Apriel-1.5-15b-Thinker/discussions/8#68dec0a694be320cab599014 explains it. "We must refuse" is a gpt-oss thing. IIRC, it speaking for 'society' is an effective way to reinforce safety.