Taste Is All That's Left

The post says AI has crushed the cost of producing software and prose, so the scarce skill is no longer making things but deciding what deserves to exist and recognizing when a result is wrong even if it is superficially acceptable. In the author’s framing, “taste” is that fast, hard-to-formalize judgment built through friction, repetition, and failed attempts. The strongest discussion signal was that many readers accepted the core premise but rejected the word. They translated “taste” into judgment, discipline, tacit knowledge, architecture sense, or maintainability, and argued that the real bottleneck is still not keystrokes but steering, pruning, and enforcing standards over time. Several commenters grounded this in practice. LLMs are useful for throwaway code, small scoped work, search, classification, and fast iteration, but they still tend to over-explain, over-comment, generate bloated tests, and lose coherence across larger systems. The recurring fear was not that models are useless, but that they make it too easy to converge on average choices and to ship code that works today while quietly raising tomorrow’s maintenance bill.