The Real Story Behind GPT-6’s “Stealth Thinking”

The gist

1. Visible chain-of-thought reasoning was a workaround for stalled model scaling, not a long-term solution, so internal looping to cut costs makes sense.
2. Newport says the actual safety risk is agents running thousands of unsupervised prompts in a row, which drift unpredictably like a game of telephone.
3. He proposes limiting autonomous AI agents to a handful of prompts before requiring a human to check in, which he thinks AI companies would welcome.
4. Modular architectures with human-readable plans, like the Cicero diplomacy AI, can handle complex tasks while staying far easier to monitor than raw LLM loops.