Since the proliferation of AI, there has of course been a lot of discussion on what human domains will be affected and how. One conclusion this discussion has come to is that a lot of the more mechanical work in the white collar domains has become obsolescent. This includes things like translation, programming, legal research and writing corporate reports to name a few. So it seems that many people have converged around this topic of taste and the idea that AI can do a lot of the mechanical work, but it does not have good taste. But what is taste? I would define it as some kind of sense of aesthetics and judgment coming from experience and intuition, being able to say whether something is good or bad without necessarily being able to explain why. Being a software engineer myself I tend to agree on LLMs lacking taste. Nowadays AI writes all my code for me and I'm mostly relegated to the role of a taste judge, whether it's judging individual pieces of code, software architecture or product and UX decisions.

Talking about taste is nothing new. I first started thinking about it years ago in university during my signal processing classes. There is no special connection between the concept of taste and signal processing, but I got introduced to the ideas of Richard Hamming, who was a mathematician and a pretty big deal in digital signal processing. He worked in Bell Labs among the greats like Claude Shannon and later on he lectured at the Naval Postgraduate School in California, where he taught a course whose lecture notes were published as a book called The Art of Doing Science and Engineering. In Hamming's words the course was not about teaching any specific technical topic, while it did cover technical topics as examples, but rather about teaching taste, or what he called style:

The course is concerned with "style", and almost by definition style cannot be taught in the normal manner by using words. I can only approach the topic through particular examples, which I hope are well within your grasp, though the examples come mainly from my 30 years in the mathematics department of the Research Division of the Bell Telephone Laboratories.

So the man who created a course trying to teach taste conceded it could not be put into words, and that examples were all he had. Thinking about taste in big part motivated me to focus on learning the fundamentals in university and later on, trying to learn the general as opposed to the specific, because the fundamentals have more leverage. Though sometimes an abstract general concept is easiest to learn through specific examples, which is what Hamming was attempting in this course as well.

What brought the topic of taste back to the center of discussion, at least in my tech bubble, was Andrej Karpathy coining the term "vibe coding" in early 2025, and then the memes of Rick Rubin followed.

The memes resonate because many software engineers (the early adopters of AI) have started feeling their ability to produce code atrophy. So Rick Rubin talking about "having no technical ability, just opinions and taste" mirrors that. Especially for me personally, as I have had many discussions on this topic with a friend, a music producer turned software engineer. There is now a clear parallel with music production and software "production" with AI, where you kind of just listen and then adjust based on inexact feelings from observing the feel of the piece of music or the LLM's output in the context of coding.

The point of all this is to ask, what is left for humans? Generative AI can of course generate anything faster than us, so is technique dead and taste the only thing that is left? Are we relegated to judges of taste or will that quickly disappear as well?

I'm not sure our sense of aesthetics will protect us for long, but it does seem like some kind of moat at least for some percentage of the population who have good taste. So perhaps taste is rare enough, and rarely enough labelled as such, that it is hard for these models to pick it out of everything else they are trained on. When we send LLM agents off to solve verifiable problems they can overcome this by coming up with tons of solutions very fast and only one needs to be brilliant. This is roughly what has been happening in mathematics, where models have started producing solutions to problems that had been open for years, published alongside machine checkable proofs. That is the whole point: nobody needed taste to decide which attempt was any good, the proof checker decided. But taste applies in domains that are hard or impossible to verify as well, and there generating a thousand attempts does not help if nothing can tell you which one is good. Then again, if Hamming thought he could teach taste not as an abstract concept but by pointing to examples of work done with good taste, couldn't the same be done with LLMs? And where does it end? People keep saying AI can do X so now people will have to do Y, but then the immediate question is, why can't AI do Y or what makes us think AI can't do Y pretty soon?

For me all this is a big reason I became a software entrepreneur. I have always been uneasy about being a narrow expert and the AI proliferation really drove that home for me, as an entrepreneur I could exercise taste much more widely than as an engineer. I make product, architectural, organizational and sales decisions. All of it requires taste: what to build, what to say, what not to say, how to get information out of people. A lot of this feels strange for someone who has been working as a software engineer for many years and I must admit that there was a period of mourning for that identity two years ago when I figured out that AI would change this field. I suppose the question is whether I should expect another identity shift of similar or bigger magnitude, if and when AI encroaches further. Meanwhile I will try to keep cultivating taste and perhaps that is the reason I'm writing this text as well.

---

Richard W. Hamming, The Art of Doing Science and Engineering: Learning to Learn, Gordon and Breach, 1997. The quote is from Chapter 1, "Orientation", p. 1.

On machine checked proofs, see OpenAI, "Ten advances in mathematics and theoretical computer science", August 2026, which published Lean 4 certificates for each result.