Why 'taste' is the white whale of AI coding: LLMs can read design systems but can't extrapolate them
threepointone · x · 2026-10-18
An engineer's long thread on the tension in taste modeling for LLMs: models readily extract and structurally understand token-based design systems from live code, but still fail to extrapolate from them discerningly the way human designers generalize a fixed system through creativity and self-critique.
The author attributes the gap to underspecification (tacit design knowledge lives in human brains, not training data) and spec shape (overemphasis on fixed values like colors, underemphasis on relationships like padding-to-radius). A quoted follow-up calls 'making taste legible and portable' a recurring white whale: weeks of techniques all lost to just crafting something intently and asserting that intention repeatedly throughout the product.
More from AGI Musings
- Checking 56 'AI killed X' claims: 6 gone, 22 shrunk, 13 grew — Mastbubbles · 2026-10-18
- Dvaita Vedanta's Intrinsic Validity: A Ground No AI Model Can Mimic — Paimaamu · 2026-10-18
- Closed AI evals face criticism: opaque, single point of failure, and potential conflicts of interest — evijit · 2026-10-18
- Paul Graham: paywalls are invisible to AI — "captcha for humans," bots get through — dbasch · 2026-10-18
- "90% of people have never heard of Claude": AI insider on how early we still are — kimmonismus · 2026-10-18
- Higher bar for LLM introspection claims: talk at NYU metacognition workshop — tallinzen · 2026-10-18