Frontier AI Models Are Losing Monitorability | Reilly Haskins

Frontier AI models are becoming harder to monitor via chain-of-thought, pushing labs toward costly action-level safeguards with real coverage gaps.

• Per-action monitors deployed: METR built an LLM-based system that scores every agent action for suspicion of causing real-world harm before execution, halting runs above a threshold for human review.
• Monitorability trending downward: Reilly Haskins says evidence points to declining transparency as models act competently "without chain of thought" better than before, with capability growing fast; he cites rumored architectures like Neuralese as context.

Themes, sentiment and reasons are AI-written paraphrases, checked so they never quote an episode. They may contain errors. Not investment advice. Episodes remain the property of their creators.