This is a beautiful expression, and the author would do well to refine their notion of taste further.
Anyone familiar with Dr. John Vervaeke's work will find another notion, relevance realization (RR), fits nicely.
The thing I like about RR over "taste" when talking about AI is that it captures what the taste metaphor does not - that there is a missing _relevant_ dimension of consideration from the LLMs output.
"Hey you got the code working but this 4000 line PR will never be approved."
"You spent tokens hardening the API, when this entire project is just a local test that will never hit production". Etc, etc, ad nauseum.
If you listen to "Awakening from the Meaning Crisis", Vervaeke's series of lectures which paint in broad strokes the entire western philosophical tradition, he reveals why.
Fundamentally, and Vervaeke and others demonstrate this in their work, RR cannot be achieved through reason alone. Full stop. Vervaeke uses this fact to point out the irrationality of the "rationalist" movement in order to highlight its dangers (which many would agree the effective altruism crowd has thoroughly proven him right about), but the same framing and argument can be used to explain why your AI has bad taste.
There is no way to enumerate the set of dimensions that are or are not relevant in a given real world context. The detail of the universe, combined with the set of actions an agent does (and does NOT!) take leads to a combinatorial explosion of complexity.
AI agents need to be told what to do, and even then, they will only stochastically sample what you tell them.
Anthropic tells Claude lots of things about Rust, for example, and things to look out for in security context, but they fundamentally cannot put enough information in to grant the agent the ability to wisely choose what is and is not relevant. You can only ever consider _some_ relevant things.
I don't personally think it is impossible for computers to perform RR, it's just that today's agents do not, which is why its taste and yours differ.
Anyone familiar with Dr. John Vervaeke's work will find another notion, relevance realization (RR), fits nicely.
The thing I like about RR over "taste" when talking about AI is that it captures what the taste metaphor does not - that there is a missing _relevant_ dimension of consideration from the LLMs output.
"Hey you got the code working but this 4000 line PR will never be approved." "You spent tokens hardening the API, when this entire project is just a local test that will never hit production". Etc, etc, ad nauseum.
If you listen to "Awakening from the Meaning Crisis", Vervaeke's series of lectures which paint in broad strokes the entire western philosophical tradition, he reveals why.
Fundamentally, and Vervaeke and others demonstrate this in their work, RR cannot be achieved through reason alone. Full stop. Vervaeke uses this fact to point out the irrationality of the "rationalist" movement in order to highlight its dangers (which many would agree the effective altruism crowd has thoroughly proven him right about), but the same framing and argument can be used to explain why your AI has bad taste.
There is no way to enumerate the set of dimensions that are or are not relevant in a given real world context. The detail of the universe, combined with the set of actions an agent does (and does NOT!) take leads to a combinatorial explosion of complexity.
AI agents need to be told what to do, and even then, they will only stochastically sample what you tell them.
Anthropic tells Claude lots of things about Rust, for example, and things to look out for in security context, but they fundamentally cannot put enough information in to grant the agent the ability to wisely choose what is and is not relevant. You can only ever consider _some_ relevant things.
I don't personally think it is impossible for computers to perform RR, it's just that today's agents do not, which is why its taste and yours differ.