The algorithmic monoculture – new old AI risk

A $23 million second-hand textbook was only the rehearsal. In 2011, two pricing algorithms on Amazon got locked in a polite little loop, each repricing against the other, until a biology textbook cost $23 million. Everyone laughed.

Now we are driving to a situation where the same mechanism is being installed, at scale, in decisions that matter, and nobody should be laughing now.

The third systemic risk of the agent era, and the one almost nobody is pricing in: coupling.

When competing firms deploy agents trained on similar data, with similar objectives, the market doesn't diversify. It synchronises.

Different foundation models are not the answer. Different models are trained on largely the same public corpus of text and use the same Transformer architecture. They converge on the same interpretations the way students of the same textbooks converge on the same answers.

In a healthy market, errors cancel: my mistake is your opportunity.

In a monoculture, decisions compound: amplified and without dumping, that could make even a correct decision for one everyone's mistake, if made simultaneously, in the same direction.

Finance ran this experiment. The 2010 Flash Crash was driven by automated strategies built on similar risk models discovering each other in real time; a trillion dollars of market value disappeared in minutes.

I remember a conversation from those years: risk managers at a large international bank asking the consultancy that had delivered their trading robots why the bank had behaved exactly like everyone else. The answer: "We give you the best models, fully compliant. How you use them is your business." It was the same answer the consultancy gave every client.

The agent era presents coupling risks for much more complex and long-term decisions, and much "better-grounded" decisions.

So where does independence come from, in the AI world? From the one input your competitor cannot buy: your own experts, training your agents through their daily interventions. The model is the shared textbook. The supervision is the private education. Firms that capture it will decorrelate; firms that rely on the vendor's defaults will move with the herd, and fail with it.

The uncomfortable question is portfolio-theoretic: diversification only works when positions are independent. How independent is your firm's judgement once it runs on the same models as your entire industry?

Coupling is the closing argument of Chapter 11 of Architecture of Intellect. The book is in my profile.

What's actually different about your agent? Honestly.