The only problem with Microsoft is they just have no taste. They have absolutely no taste... they don't think of original ideas, and they don't bring much culture into their products. — Steve Jobs, Triumph of the Nerds (1996)
Jobs said that thirty years ago as an insult. In 2026 it reads like a job description. When anyone can generate a working app, a passable essay, or ten design variations before their coffee cools, the thing that was always quietly doing the work — taste — steps into the light as the only part that was ever scarce.
I've written before that AI makes the manifesto's discipline more essential, not less. This is the sharper version of that claim. Strip it down and the skill that survives the automation of making has three faces: verification, taste, and knowing what's worth building. They're the same faculty pointed at three moments — before you build, while you build, and after. And the middle one, the aesthetic one, is the one nobody wants to admit is real.
Taste is not decoration. It's a decision.
We treat "taste" like it's about whether the buttons are pretty. It isn't. Taste is the ability to look at a field of plausible options and know which one is right — meaningful for this problem, feasible in this context, appropriate for these people — and to feel the wrongness of the others before you can fully articulate why. It's compression: years of seeing what works, collapsed into a fast, defensible "not that, this."
A model has read more than any of us and has none of it. It will hand you a beautiful solution to a problem you don't have, in a house style borrowed from a million others, with total confidence. It generates the average of everything. Taste is the refusal to ship the average.
The three skills that survived
The grunt work got automated. What's left is the part that was always the value, now standing on its own:
- Knowing what's worth building. This is taste pointed at the future. Before a line is generated, someone decides this is the thing, for these people, for this reason — and, harder, decides what not to build. A machine that makes everything cheap makes this discipline priceless, because it will just as happily build the wrong thing beautifully. This is the manifesto's whole front end: draw the map, shape the work, state the hypothesis.
- Verification. Taste pointed at the present. Fluent wrong is more dangerous than obvious wrong because it disarms your scepticism — a hallucinated answer arrives in the same confident prose as a correct one. "It doesn't exist if it can't be tested" stops being a slogan and becomes the load-bearing wall. When generation is free, the test strategy is the bottleneck, and that is exactly where the bottleneck belongs.
- Taste proper. The aesthetic read — good versus merely plausible. It's the least teachable-sounding and the most valuable, because it's the one the model most conspicuously lacks. It's what lets a senior engineer glance at generated code and feel the wrong abstraction, or a writer cut the paragraph that's technically fine and secretly dead.
The data already shows the split
If everyone had the same tools, you'd expect everyone to ship more. That's not what's happening. The 2026 delivery reports tell a stranger story: output is up across the board, but outcomes have forked. The top slice of teams roughly doubled their throughput while the bottom quartile saw no measurable gain at all — same models, same access, wildly different results. Coding agents now run for tens of minutes unsupervised instead of a handful, so they produce vastly more — which only widens the gap between the teams who can tell good output from noise and the teams drowning in it.
The tool didn't level the field. It tilted it toward the people with taste and away from the people who mistook volume for progress. Cheap generation is a magnifier: it makes discernment more valuable and its absence more expensive, on the same brutal schedule as always.
Taste is learnable — which is the good news
Here's the part that rescues this from fatalism. Taste is not a birthright or a personality trait; it's a discipline, and disciplines can be practised. Paul Graham argued this a quarter-century before it was urgent: good design is real, it has properties you can name, and the way you get it is deliberate — consume the best, study why it works, make things, and look hard at what you made. The loop is the same one the manifesto has always described. Get it wrong cheaply. Go find out. Test the thing against reality. Bank the win and stop.
What's changed is only the stakes. When making was expensive, weak taste was survivable — the cost of production hid it. Now that making is free, taste is the entire remaining surface. The people who thrive won't be the ones who generate the most. They'll be the ones who know what's worth generating, who can spot the confident wrong before it ships, and who have the taste to tell a good thing from a plausible one.
That was always the job. The machine just finally cleared away everything that wasn't.
References
- Jobs, S. (1996). Triumph of the Nerds, PBS. — The "no taste" line, delivered as a jab at a competitor, aged into the clearest one-sentence brief for what human work becomes once the making is automated.
- Graham, P. (2002). "Taste for Makers." — The essay that insists good taste is real and, crucially, learnable — a deliberate practice of consuming the best and interrogating your own output, not a gift you're born with.
- Agrawal, A., Gans, J., & Goldfarb, A. (2016, November 17). "The Simple Economics of Machine Intelligence." Harvard Business Review. — Priced it early: when prediction gets cheap, judgement gets dear. Taste is judgement with a faster clock.
- CircleCI. (2026). "2026 State of Software Delivery." — The empirical shape of the split: AI lifted output everywhere, but the throughput gains concentrated at the top while the bottom quartile flatlined. Same tools, different taste.
