Taste as a Trainable Model
Taste is a virtual machine in your head that predicts whether your in-group will like an idea — built by reps with feedback.
- Difficulty
- Moderate
- Time to result
- ~months to results
- Steps
- 3
- Confidence
- 88%
Schoening reframes taste from an innate gift into a trainable prediction engine. Taste, in his definition, is the ability to run a virtual machine in your head that, given an idea, predicts whether a specific in-group will like it. You build that model exactly the way a machine-learning model is trained: repeated cycles of input (idea) and feedback (how people reacted), backpropagated over a long time in a specific domain. This demystifies taste and gives a concrete method — pick your in-group, maximize the frequency of feedback reps.
Origin
Schoening's own formulation, which he explicitly likens to backpropagation and model training; he uses Japanese craftspeople as the archetype of taste built through sheer volume of reps.
Core principles
- 01Taste is domain-specific and takes a long time to build, then can extrapolate to adjacent domains
- 02Taste = a mental virtual machine predicting a chosen in-group's reaction
- 03You must decide who your in-group is; being the only person who likes something is not taste
- 04The build loop (idea -> reaction -> learn) is structurally identical to training a model
- 05Frequency of reps with feedback is the lever — increase it
How to run it
- 1
Choose your in-group
Explicitly decide which group you are building taste for. You don't need to satisfy 8 billion people; taste is judged relative to a defined audience, and everyone is in the top of the class at something.
Pro tip Narrowing the in-group makes the prediction problem tractable and the feedback signal cleaner.
Watch out If you are the only person on the planet who thinks something is good, it isn't good — a private taste of one doesn't count.
- 2
Run high-frequency feedback reps
Repeatedly put ideas in front of the in-group and observe their reaction, then adjust. This is the training loop; the more reps per unit time, the faster the internal model improves.
Pro tip Increase the frequency of reps — like Japanese craftspeople painting the same bowl for years — rather than waiting for rare, high-stakes feedback.
Watch out There is no way to speed-run this; people with the best taste have simply done it for a long time.
- 3
Surround yourself with tasteful things and other people's ideas
Expose yourself to tasteful reference objects so your own work feels lacking by comparison, and to other people's tools and ideas as a source of new inputs. Schoening notes Notion names conference rooms after famous objects so people feel they must do better.
Pro tip The high-taste designers Schoening sees both ship end-to-end side projects and constantly tinker with new apps — exposure plus ownership.
Watch out Being the annoying person suggesting a 49th new tool can grate, but that restless exposure is a real source of taste.
In the wild
Schoening points to Japanese craftspeople who have painted the same bowl for however long — the taste comes from sheer accumulated reps over a very long time.
→ Illustrates that taste is a function of rep volume, not innate talent.
Notion names its conference rooms after famous objects — the first typewriter, the Macintosh, a Porsche 911. Sitting in one, Schoening feels that nothing he's doing amounts to that, prompting him to do better.
→ Deliberate exposure to tasteful things raises the internal bar for one's own work.
Common mistakes
Treating taste as innate or magical
Believing taste is a gift you either have or don't stops you from training it. Schoening frames it as backpropagation — reps with feedback — which anyone can run.
Confusing personal preference with taste
Judging an idea good because you alone like it ignores the in-group. Without a defined audience and their reactions, you have preference, not taste.
Is it for you?
Best for
Designers, PMs, and founders who want to systematically develop judgment in a specific product or craft domain
Not ideal for
People expecting a shortcut — the method is slow by nature — or domains where you cannot get regular audience feedback
From the transcript
“run a virtual machine in your head where given an idea, you can predict for a certain inroup whether they're going to like it or…”
“you just have to do reps. It's almost like training a model”
“you decide what your in-group is and then how good do you get at emulating uh how they will react to it”
From the episode
Why cultivating agency matters more than cultivating skills in the AI era
Max Schoening (Head of Product, Notion)