Three visions for diffuse control
A leading AI safety thinker warns against a dangerous gamble on future AI.
Deep Dive
The article critiques the 'diffuse control' plan, where humans would force misaligned AI to solve its own alignment problem. The author argues this is procrastination, delegates a critical task to an untrusted party, and is vulnerable to AI deception. It contrasts this with other safety strategies like solving alignment first or using iterative oversight. The core warning is that waiting for AI to fix itself is a high-risk strategy with a low chance of success.
Why It Matters
This debate shapes how humanity approaches the existential risk of superintelligent AI.