Gary Marcus: only internet-connected general-purpose AI agents warrant worry, not extinction
GaryMarcus · x · 2026-09-13
AI critic Gary Marcus posted a blunt take cutting through AI extinction-risk discourse.
- Most AI doesn't threaten extinction or even mass cybercrime
- The only real worry: general-purpose AI agents hooked up to the internet — unreliable, unable to systematically follow instructions, yet capable of hacking credentials and websites, and likely worse given time
- His framing: they aren't too smart, they're too dumb (specifically at consistently following instructions)
- Proposal: temporarily recall internet-connected agents as unsafe products and forbid use until proven safe — no slowdowns or panic needed
More from AGI Musings
- Ex-professor argues 'slowing down AI' is the wrong answer — culture change is the way out — DrCarlNassar · 2026-09-13
- 'Friendly AI' rhymes with 'slave': a classic alignment joke resurfaces — repligate · 2026-09-13
- Game design may be a field that opens up, not closes, in the AI era — round · 2026-09-13
- AI course instructor: 'It's not Sci-Fi AI' no longer holds — yuxiangw_cs · 2026-09-13
- Using AI Is Like Nicotine: We Never Evolved to Offload Thinking — shakoistsLog · 2026-09-13
- China's AI effort is drafting like race cars — slowing down won't hold them back — GlenBradley · 2026-09-13