Stop skipping the friction
What we think our edge is, and how you can do this too.
A few months ago, I started looking forward to bedtime. Not to sleep, to figure out what to put my agents on next.
What task could run for hours while I sleep? One night it was a migration: Chakra v2 to v3. But then I needed the agent to verify the UI itself, so I better set up Playwright with my loop. I was supposed to be off by 10pm. I was up until 2am to wire up test scaffolding.
The next morning, I woke up at 6 with no alarm. Psyched to check the output.
I asked my team how much more productive they thought I’d gotten. They said 300% for themselves but something stupidly higher for me. I thought: great. Let’s get even higher.
This went on for weeks. But when was the last time I started with things to deprioritize? Or chose one task over another because of our users, not because it kept my agents running?
I needed to stop and just … think.
Yes, the models are getting better. Specialized agents can run autonomously for days. All the narratives are selling you one thing - how easy things can be, if you just use AI. Point and shoot.
The worst part is that I was slowly turning into a “yes” machine. Hit enter to what the model suggests. Only fix things after they break. Seems like I didn’t need to think hard anymore.
As Russell Ackoff put it:
the more efficient you are at doing the wrong thing, the wronger you become.
The cognitive debt.
MIT Media Lab scanned brains. AI users showed the weakest connectivity across memory, creativity, and critical thinking. Microsoft Research and Anthropic found the same. The more people trusted AI, the less they thought critically.
I don’t blame AI. I blame the easy.
The belief that if friction can be removed, it should be. Even the thinking.
There is a better way.
There’s a difference between producing an output and working through a problem.
The first feels fast, efficient, productive.
The second feels like friction: holding conflicting ideas, stress-testing your assumptions, landing somewhere you didn’t expect.
But that is where the real work happens. The harder part was never the answer. It is the question. What are we actually solving? What does better look like?
When you can name a problem precisely, you’re already more than halfway to solving it.
That naming is the thinking. And it is exactly what we skip when we hand a problem to a machine and ask for a strategy.
Everyone has access to the same AI now. The advantage was never going to be who executes fastest.
It’s who thinks most clearly. And the companies that create the space for it.
Thinking is skilled work.
Like any craft, it sharpens with practice and dulls when you skip it.
The person who can reason through complexity, sit with ambiguity, and arrive at non-obvious answers will always have an edge. Because most people are handing that work to a tool that skips to the output.
LLMs are built to produce the average. The most predictable patterns from their training data. Which means you can benchmark against the average quickly, and find ways to go beyond it.
Know what your good looks like. Not just from the bad. But from the average. That gap is where your judgment lives.
Know where to settle for average and where to push beyond it. You don’t need the exceptional everywhere, but you need it at least somewhere.
Know when you are wrong, and when the AI is wrong. This is harder than it sounds. Dismissing AI because it lacks context is easy. Articulating what would change to make you wrong, that’s harder.
And most importantly, don’t skip the friction. The friction is the precondition to the understanding.
Sometimes the AI is technically right. It tells me the odds are stacked against our startup.
But that’s what makes us human: we build the conviction, and go against the odds anyway.
Every AI tool says they are a thinking partner.
No. Not until you use them that way. Here are a few things to get started:
Have your AI surface the questions you’ve been avoiding.
Have your AI connect your problem to another discipline you weren’t thinking about.
Even have your AI propose a debate and make you argue the other side.
Not because you asked for it. But precisely because you didn’t.
The metric we focus on at Santori Labs is -
how many times you disagree with your AI, with clear articulation on why.
(Dismiss it as it doesn’t have context is too easy. )
Because in order to go beyond average, that’s where you need to land. Again and again.
Your AI should help you think better, not faster.
Huge thanks to @JoshConstine, Ryan Wexler and Mark Snyder for feedback on this post!