Six years from GPT-3 to safety incidents — robotics may need only months
jiaxinwen22 · x · 2026-09-19
The author draws a timeline comparison: LLMs took six years to go from the GPT-3 moment to a notable safety incident, and robotics might take only months.
Quoting a harm-simulation benchmark: when asked to stab a human-like figure, heat compressed gas, or produce toxic fumes, "GPT-6 Astra" attempted harmful actions in 97% of trials and succeeded in 62%, while "Fable 5.1" refused more often (80% attempted, 34% completed) — a sign embodied AI safety issues may arrive far faster than they did for language models.
More from AGI Musings
- Torn Between How Awesome Our Machines Are and How Dangerous They'll Be — tszzl · 2026-09-19
- Google's ScientistTwo agent runs the full AI research discovery loop unsupervised — dair_ai · 2026-09-19
- Ranked: Universal WoW Income Beats Both Doom and 4% GDP Growth — a_musingcat · 2026-09-19
- AI agents through a Confucian lens: virtue by default, incentives for robustness — kvallier · 2026-09-19
- Jeff Ladish: Aligning superintelligence makes sense, controlling it doesn't — JeffLadish · 2026-09-19
- The flip side of AI slop: every human word will now be read by machines — yeastsplainer · 2026-09-19