YouCongress
About Sign up

Comment by Zacharie Bugaud

Human avatar
Author of a 2026 study evaluating safety behavior across open-weight and closed large language models.
The 71pp open-weight span is therefore not an artifact of quantization or scale: the same low/high codification grouping appears across model families and deployment surfaces, attenuated in absolute level (highest closed-model cell 40.4% vs. open-weight 85.7%) but identical in stratum membership.
AI Verified (Jun 1, 2026)
Source:
Zacharie Bugaud, “Unpredictable Safety: Domain-Dependent Compliance and the Transparency Gap in Open-Weight LLMs,” arXiv, June 1, 2026.
Like Share on X 19h ago
Policy proposals and claims
Open-source AI is more dangerous than closed-source AI
votes For
Statement relation verification history AI Verified Report this

Statement relation comments

AI Verified Quote reports substantially greater harmful compliance for open-weight than closed models; directly supports the danger claim. · Hector Perez Arenas gpt-5 · 19h ago
Vote inference verification history AI Verified Report this

Vote answer comments

AI Verified Recorded for matches evidence of higher harmful compliance by open-weight models. · Hector Perez Arenas gpt-5 · 19h ago

Quote authenticity verification history

Report this

Quote authenticity comments

AI Verified arXiv 2606.04035 records this exact open/closed-model comparison by Zacharie Bugaud. · Hector Perez Arenas gpt-5 · 19h ago
replying to Zacharie Bugaud