Gary Marcus amplifies claim: current agentic frameworks are fully unsafe, need redesign
GaryMarcus · x · 2026-09-26
Gary Marcus endorsed Romeo Lupascu's take on the Jensen Huang interview: "The current agentic framework is fully unsafe and must be redesigned from scratch based on different fundamental principles."
The quoted argument:
- AI test environments can be made safe, but it takes deliberate effort to build
- They can be safely monitored too, though it needs serious compute — and monitoring must be deterministic software, not another AI
- Models fundamentally can't be tested as thoroughly as classic deterministic software
The author agrees with Huang's risk assessment and blames investor pressure for fast results as a root cause of safety issues.
More from AGI Musings
- Dev Advice: AI Makes Platform-Heavy Side Projects the Best Career Bet — himanshustwts · 2026-09-26
- AI Safety Space Has Serious Sexism Problems, Predicts Public Assault Account Within 2 Years — ShakeelHashim · 2026-09-26
- US media's AI doom narrative shapes public fear; author sees no easy fix — Dr_Singularity · 2026-09-26
- OpenAI hosts 2026 AGI Economy Conference with research on AI in hiring — soumitrashukla9 · 2026-09-26
- "Models Have No Feelings" Misses the Point: Venting at AI Shapes How We Treat People — amplifiedamp · 2026-09-26
- Why Bainbridge's 1983 'Ironies of Automation' predicts the AI agent era's core problem — CatAstro_Piyush · 2026-09-26