Eureka

AI will automate 98% of everything. This is about the 2% it never will.

You've felt it before. You're not even working on the problem anymore. You're in the shower, or on a walk, or half asleep. And then it just arrives, whole, like someone whispered the answer directly into your mind.

You don't check if it's right. You just know.

I just had a eureka about eureka. Here's what I think is happening.

Surgeons get better with reps. So do generals and firefighters and chess players. The knowledge required to work can be taught and practiced and maybe even perfected. You see enough patterns and you start recognizing the next one a little faster each time.

But eureka works differently.

The firefighter who instantly knows to ventilate the roof is using expertise. Whoever first realized you could ventilate roofs at all, that person was having a eureka moment. Expertise lets you select from options that already exist. Eureka creates options that didn't exist before.

A psychologist named Wallas figured out the stages a century ago. It was preparation, then incubation, then illumination, and finally verification. You work hard on something, you step away from it, lightning strikes, and then you check if it was real.

The model still holds up. But the mystery was always that third stage, the illumination part, the lightning itself. Nobody could explain what was actually happening in there.

Then neuroscientists finally looked inside.

Turns out your brain doesn't quiet down when you stop thinking. The default mode network, which is what runs when you're idle, burns twenty times the metabolic resources of conscious thought. Your unconscious is working the whole time you're away.

And researchers found something else strange. They tracked mathematicians working at blackboards and measured all their tiny movements. Minutes before someone said "I see it," their behavior became measurably unpredictable. The system was reorganizing itself before the person even knew anything was happening.

The magic is real, and it turns out it's machinery. Your machinery.

When you're actively working on a problem, you're searching within a frame. The frame is made of options you already know exist. Focus narrows, and that's kind of what focus means. You end up looking harder and harder at available options, contemplating which one fits best.

Stepping away dissolves the frame. Your unconscious keeps working, but it's searching across wider territory now, wandering into options that were never on the menu.

Pasteur said chance favors the prepared mind. But chance also favors the prepared mind that knows when to stop preparing.

Machine learning has this phenomenon called grokking. Models will flatline for thousands of training steps, looking completely stuck, and then suddenly shoot up to perfect performance. They were building structure the whole time, it turns out. The metric just couldn't see it happening.

Humans do this too. The plateau that feels like nothing, and then the click. Structure crystallizing. Your mental model getting sharper.

But eureka goes even further than that. Structure crystallizing is your map of existing territory getting clearer. Eureka is new territory appearing out of nowhere. The option space itself expanding.

Taste can be developed. You absorb enough good and enough bad and eventually you start to feel the difference without thinking about it. Intuition can be built the same way. Enough reps and pattern recognition just becomes automatic.

AI is getting better at both of these things. Give it enough data and something like taste starts to emerge. Train it long enough and something like intuition forms. These are accumulation, basically. And accumulation can be automated.

But eureka doesn't accumulate. It just arrives, complete and pre-validated.

When AI generates options, you get a hundred possibilities. Then you apply your taste, filter them down, and find the one that actually works. It's two steps. Generation first, then validation.

Eureka does both at once. The mechanism that generates the answer is the same mechanism that validates it. The taste was already applied somewhere you couldn't see, by a part of you that you couldn't instruct.

That's the magic that stays yours.

I work twelve hours or more most days. But some days I relax, or read books about things that have nothing to do with my work, or watch YouTube videos about random stuff that caught my interest.

Both parts matter.

The hard work is essential because you need to care deeply about the problem. You need to turn it over in your head a thousand times, try every angle, hit every wall. That's how you load the system.

But the diverse inputs matter just as much. Reading across domains, hobbies that seem unrelated, letting your mind wander through territory that has nothing to do with the task at hand. That's what gives your unconscious new material to work with, new patterns to mix in when you're not looking.

And then you have to actually step away. Sleep properly. Let the frame dissolve. Your unconscious needs room to search territory you don't even know exists yet, and it can't do that while you're standing there watching.

There's no formula that guarantees it. You can do everything right and the lightning might not come today. Or tomorrow. That's the honest part.

But you put in the work anyway because when it does come, and it does come, it feels like God whispered it directly to you. You can't manufacture that. You can only be ready for it.

So you keep working, keep reading, keep letting the frame dissolve. AI will handle more and more of everything else, and that's fine. This is the part of your work that can't be automated or planned. That's why I call it magic. I just spent the better part of an hour trying to articulate the moment, but I can't and I don't think anyone can.

So do the work, and rest, and trust that maybe today is the day.