Market maps and AI models regress to the mean, by design

Regression to the mean is not a metaphor here, it's the literal mechanism.

• An LLM predicts the next token by statistical likelihood across a training corpus. It is architecturally built to output the average of what's already been written.

  
  

• A market map does the same job with a different medium: it takes last cycle's winners, sorts them into categories, and calls the result a picture of the space.

Both tools compress the past into a consensus view. Neither has any mechanism for surfacing what hasn't happened yet.

Treat them as sources of insight rather than references, and the flaw becomes structural, not incidental.

The consequence is convergence.

• Founders prompting the same models for their positioning arrive at the same positioning.

  
  

• Investors underwriting off the same market maps end up bidding on the same categories, at the same time, for the same reasons.

  
  

Category-defining companies almost never originate inside an existing box. They come from edge signals: contradictions and unmet needs visible only from inside a market, not from a dashboard.

AI has also compressed the cost of shipping code and interfaces to near zero, so execution speed stopped being a moat several years ago. What's left to differentiate on is taste, distribution, and a willingness to build somewhere the map hasn't been drawn.

The irony is that regression to the mean is self-reinforcing once enough people rely on the same tool.

• Each founder who takes the model's output as a starting point adds another data point to the corpus the next model trains on.

  
  

• Each fund that underwrites off the same map cements that map as the consensus view, which shapes the next round of maps.

  
  

The average doesn't just get described, it gets manufactured, and it pulls harder on every cycle that feeds it.

Outliers survive only where someone deliberately steps outside the loop, which is precisely what makes them rare and precisely why they're mistaken for noise until they aren't.

Neither tool was ever built to find what comes next. That was never the design goal, and no amount of better prompting or better categorization changes what the architecture is optimized to do.