Your Freedom Is in Your Unpredictability

Jacob Collier is a musician. A harmonist, specifically, which is a word he uses for himself. In a 39-second YouTube Short last week that I watched as I doom scrolled, he said something I've been chewing on.

"In a world of AI, your freedom is in your unpredictability."

His reasoning: large language models are trained to do the most predictable thing next. That's the whole point. Humans are differentiated, among many other things, by being inefficient. The things most meaningful to us aren't always the most predictable ones. So as a musician, one of his greatest joys is taking a melody and doing the most unexpected thing with it. "Whatever you think I'm going to do, I'm going to do the opposite of that."

The model drafts the average

When I ask Claude to write an email, it writes the email a median professional would write. When I ask it to draft a recommendation, it produces the recommendation a competent consultant would produce. That makes sense. It was trained on billions of examples of competent professional writing and is optimizing for the most probable next token.

If I send that draft as-is, I've shipped the average. I've contributed nothing that wasn't already implicit in the training data.

The value I add is the deviation. The word the model wouldn't have picked. The framing that runs against what the reader expects. The section I cut that the model would've kept, or the tangent I kept that the model would've smoothed away. The irreverence, in Collier's word, which I had to google the definition for btw.

Lately I've started thinking of the model as a mirror for convention. Its output tells me what the expected move is. My job is to decide whether the expected move is the right one, and if not, to diverge from it on purpose.

The part Collier left out

Collier's advice only works if you already know what the expected move is. He isn't random. He's a trained musician who's spent his life learning the conventions he's choosing to break. His irreverence lands because it's informed.

Most people using AI at work don't have that foundation in whatever they're asking it to do. Imagine it's your first time writing a performance review as a new manager. You've never read one end-to-end; you have a vague sense they include strengths, growth areas, and goals, but you don't know what weight each section carries or what phrasings tip from honest into career-limiting. You ask whatever AI you're using to write you a draft. It has all the expected sections, professional tone, balanced feedback. It reads fine to you, because you have no basis for comparison. You submit it. Your direct report reads it as generic corporate mush, and you can't tell they're right.

This lines up with something I wrote about vibe coding. AI lowers the typing barrier, not the thinking one. People with craft use AI and move further faster because they can tell when the model cuts a corner. People without it produce more work that looks finished and is subtly wrong (aka slop).

The Collier version of the pattern: divergence only works if you know what you're diverging from. Put plainly: if you've never written a good performance review, you can't recognize when the model gave you a mediocre one.

How I use it differently now

I treat AI output as a hypothesis about what the average answer looks like. That's changed how I work with it. It gives me something to push against.

It also makes me skeptical of the framing that says AI "democratizes" writing or strategy or design. It democratizes the average version of those things. The valuable work is the informed deviation from the average, and that part is still rare.

Collier is right that your freedom is in your unpredictability. Which leaves the harder question: how does anyone build the craft, when AI is there to fill in for the reps they haven't done?

A senior consultant with AI moves faster and ships more polished work than they could alone, because they can tell when the model cut a corner. A junior analyst fresh out of college with the same tool produces slop and doesn't know it, because they've never seen what good looks like without the tool doing it for them. It's like handing a graphing calculator to an elementary kid before they've learned arithmetic. They get right answers without ever building the mental model that makes those answers meaningful.