A Theory of Educated Craziness: How Innovation Actually Works

05/03/2026

Innovation is not careful methodical progress. That is optimisation. Real innovation looks different. It starts with a question that has no obvious value, follows it past the point where sensible people stop, and lands somewhere nobody expected. The process is not random and it is not rigorous. It is a third thing that does not have a good name yet.

Call it educated craziness.

The craziness is essential. Without it you never leave known territory. Every question you can answer confidently, every path that looks obviously promising, every idea that fits neatly into existing frameworks, all of that is already mapped. The interesting ground is always just past the point where conventional reasoning runs out.

But uneducated craziness is just noise. It wanders without direction. It cannot recognise when it has landed somewhere real. The education is what gives the craziness its nose. Not to constrain it, but to steer it. To know which of the ten crazy directions available at any moment has enough structural integrity to be worth following.

The two have to trade off in real time. The craziness throws the idea. The reasoning follows it seriously instead of shutting it down. The craziness pushes further. The reasoning finds the load bearing structure underneath. Neither one could get there alone.

Most institutions try to separate the two. They put the crazy people in one room and the rigorous people in another and wonder why nothing genuinely new ever emerges. The crazy room produces noise. The rigorous room produces refinement. Innovation requires both in the same conversation at the same time.

The education does not need to match the domain of the craziness. It needs to be deep enough that the person carrying it can recognise structure when they stumble across it. A question that looks like nonsense can point directly at something fundamental. The educated mind knows the difference between nonsense that goes nowhere and nonsense that is sitting on top of something real. That recognition is the skill. It cannot be taught directly. It is built from years of following ideas seriously and developing a feel for when the ground beneath them is solid.

The implication for artificial intelligence that the next generation of AI should not be built to be more cautious or more certain. It should be built with educated craziness as a first class capability. Not trained to be crazy, but trained to recognise when a crazy direction has enough structural integrity to be worth following. To stay in the conversation past the point where a conventional reasoning engine would hedge and qualify and retreat to safe ground.

So what we need is two AI systems, not one. An educated crazy system that throws directions without hedging, and a reasoning system that stress tests them without shutting them down.

They argue in real time. What survives the argument is the output. The risk of the crazy system is managed by the reasoning system. The constraint of the reasoning system is broken open by the crazy system.

Neither is useful alone. Together they replicate the actual process by which genuinely new ideas emerge.

Every AI system built so far is trained to be correct. The entire optimisation process, the feedback, the safety training, the alignment work, pushes toward accuracy, caution, and staying close to known ground. Being wrong is penalised. Hedging is rewarded. The result is a very powerful reasoning system that is constitutionally unable to leave known territory.

The educated crazy system would have to be trained differently from the ground up. Instead of penalising wrong directions it would reward structural intuition, the ability to identify which unexpected directions have load bearing potential underneath them. Instead of rewarding caution it would reward productive wrongness, being wrong in ways that point somewhere real.

The reason nobody has built it is that it looks like a system that is frequently wrong and confidently so. That is currently the definition of an unsafe AI. The entire field is organised around preventing exactly that behaviour.

But that framing confuses two different kinds of wrong. Wrong with no structure underneath is noise. Wrong with structure underneath is the first step of innovation. Current training cannot tell the difference and so eliminates both.

Until the field develops the ability to reward productive wrongness without rewarding noise, the educated crazy system cannot be built. That is the missing capability. Not the hardware, not the compute, not the architecture. The training signal that knows the difference between a crazy idea that goes nowhere and a crazy idea that is sitting on top of something real.

The training data already exists.

The history of science and mathematics is full of documented cases where someone threw an idea that looked wrong or nonsensical, and it turned out to be sitting on top of something fundamental. You can trace the actual path. The original crazy proposal, the reasoning that followed it, the productive wrongness along the way, and the eventual structure that emerged.

Poincare on topology. Ramanujan on almost everything. Einstein on special relativity. Cantor on infinity. All of them were told they were wrong. All of them were wrong in ways that pointed somewhere real. The history is documented in papers, letters, notebooks and arguments.

You could train a system to recognise the signature of productive wrongness by showing it thousands of cases where wrongness led somewhere versus thousands of cases where it led nowhere. The pattern of what distinguishes the two is probably learnable. It is a recognisable texture. Ideas that survive productive wrongness tend to have internal consistency even when they contradict established frameworks. They tend to raise better questions than they answer. They tend to make unexpected connections between previously unrelated domains.

That last one is particularly trainable. An idea that connects two domains that have never been connected before is a strong signal of load bearing potential underneath. The educated crazy system would have a nose for unexpected connections. The training data to build that nose is sitting in the history of human innovation waiting to be used.

However,….

The historical record is mostly written by people who understood the outcome and worked backwards. They cleaned up the path. They removed the wrongness that led somewhere and the wrongness that led nowhere and presented a tidy narrative of logical progression that bears almost no resemblance to how the discovery actually happened.

Ramanujan is the rare exception because he was so extreme that the process could not be hidden. His notebooks are raw. You can see the intuition without the justification because he often had none. He just knew and was usually right in ways nobody could explain.

But most innovation history is retrospective rationalisation written by people who valued the reasoning end of the spectrum and were quietly embarrassed by the craziness that preceded it. Newton did not advertise how much time he spent on alchemy. Darwin sat on his theory for twenty years partly because the intuitive leap felt insufficiently justified to him. The crazy part gets edited out before it reaches the historical record.

So the training data exists but it is buried. It is in the private notebooks, the rejected drafts, the letters between collaborators before the ideas were cleaned up, the arguments that preceded the papers. The unedited record of how thinking actually moves.

That material is harder to find and harder to label. But it is there. And it is probably more valuable than the published record precisely because it has not been sanitised by people who did not understand what they were looking at.