Taste

2025-09-12

A meditation on humanity’s last edge over AI.

Taste is art

Literally interpreted, taste is one of our five senses.

Taste is multidimensional: salty, sweet, savory, sour, fatty, mouthfeel, trigeminal, aroma.

Taste is balance within dimensions: not too much nor too little.

Taste is balance between dimensions: sweet paired with sour, fat with acidity.

Taste is necessarily subjective, as a visceral experience, yet also objective, as people can generally agree which of two comparable dishes is tastier.

Taste also extends to the other domains. Fine fragrance, painting, sculpture, architecture, music, poetry, calligraphy, and Go all share the multidimensional, balanced, subjective, and objective aspects of taste.

We say someone has taste when they are capable of discerning options that are more tasteful than others. All other things being equal, people prefer more tasteful options to less tasteful options.

Taste is parsimony

Claude’s coding output is functional, but hardly tasteful. It’s full of unnecessary comments, unnecessary features, unnecessary layers of abstraction, lacks organizing philosophy, and generates rampant conceptual debt.

Within mathematics, mathematicians share an understanding of The Book: an imaginary collection of the simplest, most elegant, most beautiful, most accessible, most tasteful proofs of every theorem.

Within engineering, engineers share an understanding of tasteful engineering: systems that fulfill their stated goals, but exceed expectations on resource efficiency, scalability, expediency, longevity, versatility, or maintainability, due to a tasteful simplicity in design.

These are just some of the fields I’m most familiar with.

Machine Learning brings us the concept of regularization - a penalty applied to all solutions according to their complexity, allowing us to tiebreak between otherwise equivalent solutions by choosing the simpler one. (Unfortunately, machine learning fails to deliver any deeper insight into “complexity” - “bigger numbers are worse” is the best we can come up with.)

Taste is the regularization function of human endeavor.

Taste is the field behind the goalposts

Effectiveness is complementary to taste.

Options can be measurably compared among various axes. Once you subtract these axes - taste is what remains.

The arts are unique in that there are virtually no measurable axes; they are entirely taste-driven.

One fascinating case study is Go. Go is a game with a very simple measuring stick for effectiveness: the player with more board area under control wins the game. Yet, Go is simultaneously a game with no discernible rules for good play. Go’s tastefulness is thus an emergent property of its strategic depth; the pursuit of taste in Go is the pursuit of victory. The measurable superiority of AlphaGo Zero has, in my opinion, destroyed much of what used to be tasteful about Go. I believe Lee Sedol’s rationale for retiring from Go is exactly this sentiment.

This brings us to my core hypothesis: Taste is the field behind the goalposts, into which the goalposts are constantly being shifted.

Humanity’s edge over AI is that we are constantly expanding the field behind the goalposts.

In the second half of this essay, I want to discuss how humans create taste.

Taste is cultural distillation

To develop taste, it helps to have a community of practitioners who are united in purpose. Community accelerates the development of taste by showcasing examples of good taste, by furnishing peers who can give targeted feedback, and by distilling a raw stream of content into a curated trickle of exemplars.

I’m a mostly self-taught software engineer and ML researcher, but I couldn’t have pulled it off without community:

I learned from the best resources, as per community consensus: SICP, Game Programming Patterns, Stripe’s API, Linear Algebra Done Right, Andrej Karpathy, Nielsen, and more.
I received and gave mentorship through Google’s Readability program.
I’m a longtime follower of Hacker News, a community whose collective taste creates a reliable filtering mechanism for content of interest, and a forum to debate the merits of each post with subject matter experts.
The Recurse Center is a community of curious programmers pushing their abilities.

What are the mechanisms by which a community distills good taste from the collective efforts of its members?

Today, we are perhaps most familiar with The Algorithm: a class of solutions that rely on making statistical inferences over user interaction data, like upvotes/downvotes, dwell time, watch time, clickthrough rates, links, shares, retweets, and more. These solutions have the advantage of scalability – there’s far too much video on YouTube for anyone to watch it all – but overall, tend to promote tasteless, lowest-common-denominator content.

Of all of the mainstream content platforms, I think YouTube comes the closest to algorithmically surfacing good content. Yet I find that it wants to veer into the tasteless, and it’s only through my application of taste in downvoting tasteless content that my YouTube recommendation feed stays clean.

Overwhelmingly, I prefer curated channels over The Algorithm. Channels – say, a newspaper, journal, magazine, blog, YouTube channel, newsletter, mailing list, or even just shelf space in a retail storefront – allow a curator to elevate content that they deem tasteful.

The interactions between its members’ taste, curators’ taste, and reputational effects are what drive communities forward. Members choose which curators to trust, and curators themselves are members of the community; if a niche channel is trusted by other curators, then that channel has influence far beyond its raw readership.

Who and what determines whether a curator is trusted? Unfortunately, taste is a self-referential problem: you need taste to pick the right curators to trust. Curators are also not infallible. One failure mode is “selling out” – trading your readers’ accumulated trust for money. Another failure mode is simply not keeping ahead of the curve, as the community collectively uplevels their taste.

It’s possible for communities to exist in high-taste and low-taste equilibrium states - a high-taste community is one in which tasteful curators are identified and elevated, and a low-taste community is one in which sellouts prosper because nobody has enough taste to tell the difference. The replacement of traditional curation with algorithmic distillation is likely responsible for pushing many online and even offline communities into low-taste equilibrium states.

Taste is governance

The collective taste of humans shapes our environment by determining our leaders, our policies, our priorities, the set of goods and services available for purchase.

What are democratic elections, but a measurement of the taste of the population? What is the free market, but a measurement of consumer taste? What is a conclave, but a measurement of Catholic cardinals’ taste? What is royal succession, but a measurement of the current King’s taste?

Any attempt to measure collective taste runs afoul of politics. Whose taste should be considered? Is it weighted, e.g. by headcount (democracy), by money (capitalism), by insider status (elitism), by bloodline (monarchy), by number of accounts (The Algorithm)? (Structurelessness is also a type of governance). What incentives are there to express ones’ thoughtful opinion? What incentives are there to prefer the collective good over personal greed?

Different systems of governance try to solve these problems, and all have their problems. These systems exist at all scales, from your local franchise restaurant, to entire institutions (e.g. “modern art” or “academia”), to countries. Over time, systems of governance, and their particular incarnations, prove their ability or inability to continually elevate curators with good taste. Unfortunately, these dynamics play out over the timescales of decades to millenia, making it hard to judge which systems are the best.

I see competitive destruction – whether it be market competition, war, or immigration – as the evolutionary force that selects for the best governance.

Refining your taste

As individuals, how can we refine our tastes? Some general advice:

Great artists have great taste. Being able to recognize good work is a prerequisite to producing it.
You grow your taste by actively analyzing why one option is more tasteful than another.
You develop taste by exposing yourself to more tasteful content. Finding a community is a fast-track to finding tasteful content.
You also develop taste by creating. Creation highlights the inherent limitations, constraints, and difficulties of the medium.
“What if” exercises – where you intentionally drop one ingredient and see what happens – are a great way to understand why something is necessary, if it is necessary at all!
Always be looking for more tasteful communities. The beginner community is rarely the same as the advanced community, and both have a place in your journey.
Not all “advanced” communities have taste. You need taste to judge taste. (e.g. in software, many communities revolve around one flavor or another of pedantry, which is not equivalent to taste.)
Taste is contextual. Yet, someone with taste in one domain can weakly judge taste in a different domain. This video of Chef Wang, an extremely tasteful Sichuan chef, trying a fine dining vegan restaurant is fascinating.

Conclusion

Having dissected the phenomenon of taste along many dimensions, I think it’s appropriate to ask: what remains? The answer is left as an exercise to the tasteful reader.