Post

a core skill in 2026 is having enough experience to tell when a model's lost the plot and is building towers to heaven

The tell is it starts babbling in something that stops resembling English.

claude *hates* talking about Babel

In general Anthropic seems to be paranoid about things which look like they could be a meaningful code that the agent can understand but which a human overseer (or perhaps a weaker guardrail model viewing only a single turn) could not.