Gary Marcus mocks claims that sandboxing makes agentic AI fully testable

GaryMarcus · x · 2026-09-19

Gary Marcus quote-tweeted a post claiming agentic AI "swarm-hacks" are fully testable in properly sandboxed, air-gapped environments (Docker, VMware, QEMU, KVM, etc.) with mature standards like IEEE and ISO — and that this even disproves Bostrom's paperclip maximizer. Marcus sarcastically replied that there are many safe ways to do this, "but cosplay is more fun."

Original post →

More from Fun

Fun channel →