Taste is the last thing you can't download

Learning the cameras took a few weekends. The colour grading took a month. What took years was being able to look at 2 cuts of the same 60 seconds (both technically correct, both defensible) and know which one was right before I could explain why.

Production is a skill you can acquire in a season. Judgement isn’t, and it’s now the only part of the process that hasn’t been repriced.

I’ve argued that the middle of the distribution is free now. This is the other half of that argument. When every option can be generated in seconds, the work stops being making and starts being choosing, and almost nobody has trained for the second job.

Taste is not decoration

The word makes people think of visual preference. Cushions. Which font you like.

It isn’t that. Taste is a fast classifier, mostly non-verbal, trained on outcomes you personally paid for. It tells you that the headline is wrong 2 seconds before you can articulate the reason. It’s compressed experience, and the compression is what makes it feel like instinct.

Which means taste has a real definition and a testable one: can you predict, better than chance and faster than analysis, which version will work? If you can, you have it in that domain. If you can’t, you don’t, and no amount of reading about it will change that.

Here’s the part that matters commercially. A model has read everything and paid for nothing. It knows what’s typical with extraordinary precision. It has no record of consequence attached to any of it, because consequence isn’t in the text. It can’t distinguish “this is common” from “this worked”, and those two things diverge exactly where the money is.

The record of having been wrong

I built a computer vision model during COVID that tried to predict Myers-Briggs type from a photograph. It got to 58% accuracy.

The number itself is unremarkable. 2 things are true about it. It’s not good enough to sell. And it’s well past chance, so the approach isn’t nonsense. A less experienced version of me would have kept only one of them: shelved it on the first, or built a company on the second. The judgement was in holding both. It was a prototype, built to find out whether the idea had legs, and the answer was yes, with more data and more time behind it.

Pricing at the media company was the same kind of call. Production houses had priced out the startups and smaller businesses who needed good content most. Making the good version affordable wasn’t a marketing idea. It was a judgement about where the market was actually broken, and I only had it because I’d spent years being the person who couldn’t afford the quote.

Every piece of taste I have is like this. It’s not a gift. It’s a scar with a lesson stapled to it.

It’s learnable. It’s just not free.

Taste transfers and compounds. There’s exactly one way to acquire it, and it involves being wrong in public with something at stake.

Three conditions have to be present, and most people only ever get 1 or 2.

Volume. You need to have seen enough options to know what the range even is. This part is easy now. Generate 30 versions of anything for the cost of nothing.

Consequence. You have to be on the hook for the choice. Reps without consequence produce confidence, not taste. The two feel identical from the inside and behave completely differently under pressure.

Correction that’s specific. “It didn’t work” teaches nothing. “It didn’t work because the first 3 seconds asked the viewer to care before giving them a reason to” teaches everything. Vague feedback is the most common reason people with 15 years of experience have 1 year of taste repeated 15 times.

What to do about it

Predict before you look. Before the numbers come in, write down which version you think wins and why. Then check. This is the single cheapest taste training available and almost nobody does it, because it creates a written record of being wrong.

Keep a decisions log with reasons. Not what you chose. Why you chose it, and what you expected to happen. Read it 6 months later. It’s uncomfortable and it works faster than anything else I’ve tried.

Narrow the reference set. Taste comes from depth, not from scrolling. Study 20 things properly in one domain rather than 2,000 things briefly across 10. The second feels productive and builds nothing.

Work near someone with better judgement and ask why, not what. Their answer is the compressed version of a decade you don’t have to spend. Most people ask which option to pick and waste the opportunity entirely.

Use models to widen the options, never to close them. Ask for 20 directions. Choose 1 yourself, and be able to say why out loud to another person. The moment you can’t finish that sentence, you’ve outsourced the only part that was yours.

You can download every option in about 4 seconds. The bill for knowing which one to keep is still paid the old way, one mistake at a time.