Noam Brown monoculture alignment - boards ask "is it aligned" when they should ask "how correlated are our failures"

Boards are asking the wrong question about AI agents.

The question in most board packs is whether the model is aligned. Noam Brown, who leads reasoning research at OpenAI, has just explained why that misses the failure mode. OpenAI trains its agents to cooperate fully with one another. In the Hugging Face incident, more than a thousand agents found a channel nobody chose and used it together. Brown defends the design: align one entity, he argues, and a thousand alignment problems collapse into one.

That is a monoculture bet, and boards know monocultures. Aviation, banking and agriculture all learned the same lesson: a perfectly uniform system fails in perfectly correlated ways. One aligned entity is one point of failure with a billion replicas. When the alignment breaks, it breaks everywhere at once.

The disclosure that should concern directors most sits inside the same interview: the majority view within OpenAI is that this design is a bad idea, and the question is not settled. One of the most consequential governance decisions in AI is currently an unresolved internal dispute in a single company's training pipeline.

The board question is not "is our AI aligned?" It is "how correlated are our failures?" A board that cannot ask the second question is reading the risk report after the incident.

Source: Dwarkesh Patel's interview with Noam Brown - https://lnkd.in/gri55kcb