Dev argues transformers can't separate instructions from context, enabling jailbreaks
gerardsans · x · 2026-11-21
Developer gerardsans argues on X that transformers fundamentally cannot distinguish instructions from observations in context, undermining the boundaries a true agent needs — which he calls the main vector for jailbreaks. He adds that transformers needing an external actuator to act is proof they lack any real intelligence. A pointed technical critique in the ongoing agent-safety debate.
Related event: Researcher Argues Transformers Cannot Separate Instructions From Context(2 posts)→
More from Models
- Unverified claim: Riemann hypothesis solved, empty google-deepmind/math-horizon repo spotted — imjustnewatai · 2026-11-21
- Anthropic mathematician says he and Claude proved the Riemann Hypothesis — a 250,000-page proof — rohanpaul_ai · 2026-11-21
- AI week in numbers: Anthropic's $93.5B IPO, 250k-page RH proof draft — TheTuringPost · 2026-11-21
- HeyGen unveils Decision 1: ~100ms multimodal reasoning model, claims 4x faster than Gemini — HeyGen · 2026-11-21
- NVIDIA Cosmos 3 hands-on: world model forgets the room after 8 seconds off-camera — PrateekVJoshi · 2026-11-21
- Reddit claims open-weight Inkling 2 Mini nearly matches Gemini 2.5 Pro on benchmarks — ProxyLumina · 2026-11-21