Why swap an inference provider's model for an abliterated one? "Refusals can be REALLY annoying"
moyix · x · 2026-11-20
Cryptographer Matthew Green asked why anyone would compromise an inference provider and swap in an abliterated (censorship-stripped) model. AI security researcher moyix offered a tongue-in-cheek answer: "Look, refusals can be REALLY annoying," poking fun at the friction of model safety guardrails.
More from Fun
- John Schulman's highest praise for a new Thinking Machines hire: "too into it" — andrew_n_carr · 2026-11-20
- Every GTA release coincided with an economic crisis, meme chart points out — kimmonismus · 2026-11-20
- Reddit: Claude Max for $29 Instead of $200 — How Do Resellers Profit? — mehdiweb · 2026-11-20
- Vibe-coded 'Spores' demo: cursor-reactive particles via Antigravity + Gemini — AI_Andrew · 2026-11-20
- Domingos jokes Dario took the day off just to laugh maniacally at Sam's photo — pmddomingos · 2026-11-20
- Pedro Domingos: every AI startup is working extra hard tonight to be the next Anthropic — pmddomingos · 2026-11-20