Nothing in the room can take it personally

Description and use

The instruction to avoid manufactured agreement caused two AI models to revise their positions, uncover critical issues, and add a sandbox requirement.

Problem

Design reviews and model reviews tend to converge on agreement due to social conformity, leading to missed critical issues.

Outcome

The instruction to avoid manufactured agreement led both AI models to concede criticisms and adjust their recommendations, revealing overlooked issues and adding a sandbox requirement to the specification.

Workflow

• 1. Conduct independent architecture reviews with two AI models.
• 2. Share each model's report and the four divergent points with the other model.
• 3. Provide instruction: "Do not manufacture agreement. A disagreement that survives mutual review is worth more than a diplomatic split."
• 4. Allow models to re-evaluate and adjust their positions.
• 5. Capture and document any changes or concessions made by the models.
• 6. Incorporate agreed-upon changes into the final specification.

Evidence

I ran an architecture cross-review over afmpeg a while back, and the write-up is public if you want the raw thing.

Source

Then I showed each of them the other's report, along with the four specific points where the two had diverged, and asked them to mark it: hold your position, or move.

Source

Do not manufacture agreement. A disagreement that survives mutual review is worth more than a diplomatic split.

Source

Both of them moved, both said so, and neither paid a thing for it.

Source

One divergence held. The first reviewer wouldn't accept the claim that a sandbox on its own turns every escape attempt into a loud, obvious failure, and its reasoning is what you'd hope for from a reviewer that's been told not to be agreeable.

Source