The Truth about OpenAI’s “Secret AI Civilizations”

The gist

1. Agent swarms are just prompt management loops, not autonomous intelligence, and multiple smaller focused prompts work better than one large one.
2. LLMs produce reasoning that sounds intentional but is often post-hoc, trained to output plausible-sounding logic rather than actual decision-making.
3. Newport says running unsupervised prompt loops connected to dangerous tools for weeks is reckless engineering, not an unavoidable side effect of AI progress.
4. Newport argues regulators should hold companies criminally liable when an LLM-powered agent commits a crime, regardless of technical explanations.