Expertise Creates Blind Spots
The framework that made you successful can stop you from seeing what replaces it.
By Carles Carrera · August 2026
I was reminded of this while reading The Infinity Machine, Sebastian Mallaby’s book about Demis Hassabis, DeepMind and the history of modern AI.
For years, DeepMind was probably the most admired AI lab in the world. It had extraordinary researchers, Google’s resources and landmark achievements such as AlphaGo. Much of its success came from reinforcement learning, self-play and the idea that intelligent systems should learn through experience.
Then language models changed the game.
OpenAI pushed much harder on transformers, scaling and language. DeepMind had enormous expertise, but that expertise may also have shaped what it expected the next breakthrough to look like.
That is the interesting part.
Expertise helps us see more clearly inside a framework. But it can make the framework itself harder to question.
Success creates a lens
This happens far beyond AI.
A company becomes successful because it understands a particular customer, product, distribution channel or business model extremely well. The better it gets, the more knowledge accumulates around that model.
Processes improve. People are hired for it. Incentives reinforce it. The company develops an increasingly sophisticated explanation of why its way works.
Then something changes.
A new technology appears. Consumer behaviour shifts. A different distribution model emerges. A competitor attacks from a direction that initially looks inferior or irrelevant.
The incumbent often sees the new thing perfectly well.
It just interprets it through the old framework.
Investors have the same problem
Experience is one of the biggest advantages in investing. Pattern recognition matters.
But every pattern is based on a world that already happened.
A great investor can become so good at recognising yesterday’s winners that tomorrow’s winners look wrong.
“Not profitable enough.”
“Too expensive.”
“Not a real business.”
“No moat.”
Sometimes those objections are correct.
Sometimes they are evidence that the category itself is changing.
A circle of competence is useful because it defines what we understand. It becomes dangerous when we quietly redefine it as what is allowed to be true.
The problem is not expertise
The answer is obviously not to become less knowledgeable.
Expertise is enormously valuable.
The challenge is keeping enough intellectual flexibility to ask:
What would have to be true for my current framework to stop working?
And perhaps more importantly:
What am I dismissing because it does not fit the model that made me successful?
DeepMind did not suddenly become stupid when language models accelerated.
Quite the opposite.
It may have suffered from something much more interesting.
It was extraordinarily good at one way of seeing the future.
And sometimes that is exactly what makes another future harder to see.