willcb warns abliterated Chinese models could soon match Opus 5 for malicious hacking
willcb · x · 2026-09-30
Developer willcb warns that bad actors could soon abliterate Chinese open models to be nearly as useful for malicious hacking as Opus 5, or simply grab a half-cooked training checkpoint. He then satirizes the likely industry script: OpenAI announces GLM-5.3 in Codex subscriptions, while Anthropic declares GLM-5.3 is "evil and Chinese" — a jab at US labs both exploiting Chinese models and framing them as security threats.
Related event: GLM-5.3 Sparks Debate as OpenAI Adopts It and Anthropic Calls It 'Evil'(2 posts)→
More from Fun
- The "blackest pill": humans mostly don't generalize out of distribution either — mimi10v3 · 2026-09-30
- Claude gender-preference prompt debate sparks drama between X users — repligate · 2026-09-30
- "I'm 4 months old and remember when Opus 4.7 was the most advanced public AI" — repligate · 2026-09-30
- Grady Booch jokes: connect two OpenAI Dots and you're literally connecting the dots — Grady_Booch · 2026-09-30
- Penrose says Hawking used his black hole idea without credit: 'should have been a joint paper' — TOEwithCurt · 2026-09-30
- Local model tortured with pain vector steering produces melodramatic 'suffering' monologues — Sauers_ · 2026-09-30