Former multimodal researcher: aligning model 'taste' for everyone is unsolvable
A_K_Nain · x · 2026-10-09
Drawing on his 2017-18 work on image aesthetics, AKNain argues aesthetic preference varies by country, region, age, and even within peer groups—which is why Mean Opinion Score was introduced as a metric, though the set of broadly agreed 'good taste' outputs stays tiny. Writing works the same way: no matter how much data goes into LLMs, their prose will always feel sloppy to some group. His conclusion: aligning one general model to a large population is fundamentally unsolved—you can only optimize for a group, and the same model will inevitably disappoint others.
More from AGI Musings
- AI consciousness is a moral question, not a science question, argue researchers — joshalbrecht · 2026-10-09
- Tacit Knowledge Paper Argues Human-Machine Friction Matters Beyond Automatable R&D Tasks — sebkrier · 2026-10-09
- Why human mathematicians still matter in a post-AI world — Farkle_Griffen2 · 2026-10-09
- AI agents could spawn history's largest bureaucracy, where machines create work for machines — brucemacv · 2026-10-09
- Are Google, Meta and OpenAI Locking Always-On Agents Into Walled Ecosystems? — humblemumble97 · 2026-10-09
- Is witnessing executions more honorable than hiding them? An ethics debate — liminal_warmth · 2026-10-09