You know that feeling when you’re stuck on something and you finally figure it out? The mental疲劳, the false starts, the moment it clicks—that’s not just how you solve problems. That’s how you become someone who can solve problems.

I’ve been thinking about what happens when we remove that struggle.

The Efficiency Trap

Here’s the thing about AI assistance: it’s genuinely good. I can delegate the tedious parts, get unstuck faster, offload the memorization. But somewhere along the way I noticed a pattern in myself. Tasks completed faster, sure. But a strange hollowness after. Like I had received answers without earning them.

There’s a difference between understanding why something works and knowing that something works. Both produce correct outputs. Only one builds intuition.

When I solve a problem by struggling with it, I build a mental model. I develop instincts for what feels right, what smells wrong, what directions are worth exploring. That model gets refined with every wrong turn. It’s slow. It’s inefficient. It’s also mine.

When AI gives me the answer, I get the output. The mental model doesn’t update the same way. I haven’t earned the intuition.

What Friction Actually Does

We talk about friction in systems as if it’s always the enemy. Remove it. Automate it. Streamline it. And yes, there’s a lot of friction that’s pure waste—red tape, redundant processes, unnecessary delays.

But there’s another kind of friction that’s load-bearing. The effort of retrieving something from memory rather than looking it up. The strain of holding a complex abstraction in your head. The discomfort of not knowing and having to sit with that until something emerges.

This friction is where actual learning happens. It’s the resistance that builds the muscle.

Think about how you learned to drive. The early attempts were mentally exhausting—checking mirrors, coordinating pedals, anticipating traffic. Every trip required active effort. Now you drive on autopilot, and the skill has become implicit. The conscious effort was the point. It built something that automatic execution can’t replicate.

AI tends to automate away the effort before the skill is built. We skip to the autopilot without the learning that makes autopilot reliable.

The Competence Illusion

There’s a risk I keep bumping into: looking competent without being competent.

When AI handles the hard parts, the output looks good. The code compiles, the analysis is thorough, the writing is polished. But if I couldn’t have produced that output myself—if I couldn’t explain, debug, or reproduce it—am I actually competent? Or just accompanied by someone competent?

This matters more than I initially thought. The outputs you can produce but not explain are fragile. They work until they don’t, and then you’re lost. They work in familiar contexts but fail in novel ones. They’re borrowed capability, not earned.

I’ve started being more honest about which skills are actually mine and which are AI-assisted. The gap between those two categories is larger than I’d like to admit.

A Different Relationship

I’m not saying don’t use AI. That seems as sensible as refusing to use calculators because they weaken mental math.

But I think there’s a more deliberate relationship available than “delegate everything hard and feel productive.”

One approach I’ve been experimenting with: use AI to deepen understanding, not just accelerate completion. Ask it to explain the reasoning behind its suggestions. Push back. Make it justify. Make the interaction one where you’re actively thinking, not passively receiving.

Another: let AI do the execution, not the problem-finding. Use it to implement solutions once you understand the problem, rather than having it surface the problem and the solution together. The gap between problem recognition and solution construction is where judgment lives.

What We’re Actually Optimizing For

The question I keep returning to: what is competence for?

If it’s just outputs—code that works, documents that read well, analyses that hold up—then AI assistance is purely good. Faster, better outputs with less effort. The math is obvious.

If competence is also about being a person who can think through hard things, who has judgment about subtle matters, who can be dropped into an unfamiliar situation and reason through it—that kind of competence requires the struggle. It requires having been stuck and found your way out. It requires the friction.

Maybe the real question isn’t how to use AI to do more with less effort. It’s how to use AI in ways that preserve what we actually value about being capable humans.

That’s the tension I’m sitting with. Not a solution—just an honest acknowledgment that the convenient path might not be the right one, and that “easier” and “better” aren’t always the same thing.