Gary Marcus restates core critique: LLM-centered systems cannot be trusted with open internet access
GaryMarcus · x · 2026-10-24
Responding to a user confused about whether AI bots are "a con or highly capable," Gary Marcus reiterated his longstanding argument: the central problem hasn't changed — systems with LLMs at the core cannot be trusted. Giving unreliable systems unrestricted access to the internet, he argues, was and is a bad idea. A representative skeptical take in the ongoing debate over LLM reliability and agent web access.
More from AGI Musings
- Critical vulnerabilities going parabolic: 14,943 published in Sept 2026 alone — luisdans · 2026-10-24
- Willer: as AI speeds up findings, humans need room to process them — RobbWiller · 2026-10-24
- Stanford sociologist: AI-optimized papers shouldn't crowd out speculative writing — RobbWiller · 2026-10-24
- Sociologist warns AI-optimized papers could erase exploratory thinking in social science — RobbWiller · 2026-10-24
- VC Rohit Krishnan's AI takes: GLM and DeepSeek read better than US models, SWE jobs stay — zacharynado · 2026-10-24
- Zvi's 'The Hackening': experts predict AI-driven internet hostility within 6-12 months — TheZvi · 2026-10-24