Arvind Narayanan: why Claude's 'superintelligent' writing style backfired
random_walker · x · 2026-10-29
Princeton's Arvind Narayanan on what superintelligent writing should sound like: good human writing traits — figurative language, information density, coined terms, punchiness — dialed up. Claude did exactly that until Anthropic dialed it back after user complaints; the 'neuralese' style read as 'a stupid person's idea of a smart person.'
Key points:
- Reader response to density and metaphor follows an inverted-U curve (his cartoon chart): past a point it fully backfires. Human writers hold back not from inability but awareness of cognitive load — simulating the reader's experience is a writer's hardest skill.
- Training AI to mimic mediocre writing is easy; optimizing for perceived goodness is very hard, and universal model use makes everyone tire of the same tics fast (he mourns no longer being able to use 'load bearing').
- The broader question: are most economic tasks like chess (clear optimizable metric) or writing (multidimensional quality with no training signal beyond human mimicry)?
More from AGI Musings
- AI wasn't overhyped: Benjamin Todd scores 18 months of forecasts — ben_j_todd · 2026-10-29
- Benjamin Todd: even AI builders keep getting surprised by progress — ben_j_todd · 2026-10-29
- "The fear isn't AI replacing you — it's realizing you only loved typing code" — feregri_no · 2026-10-29
- Employers worldwide pull back on AI amid errors, low-quality output and rising costs — GaryMarcus · 2026-10-29
- Early AI researchers made 9 figures while ETH/BTC core devs earned $150k a year — menhguin · 2026-10-29
- Why early no-coders excel with AI: they already built the frustration muscle — jacob_posel · 2026-10-29