Pew's Digital Twins Show Large Errors; Just Asking Frontier Models Is Far More Accurate at 1/1000th the Cost
paulnovosad · x · 2026-10-02
Reacting to a new Pew Research study, John Horton notes the AI-built "digital twins" of survey respondents showed large average errors vs. real people. Simply asking frontier models for the response distribution was far more accurate — at roughly 1/1000th the cost. "Bitter lesson comes for digital twins," as Arpit Rage put it: heavy simulation machinery may be beaten by plain model queries.
More from AGI Musings
- Personal agents dominate Machine Earning summit: 40% of digital work time within a year, Town CEO predicts — illscience · 2026-10-02
- Reddit user uses Claude to meta-analyze Awesome Jev repos; AI coding emerges as top application — PrincipleFar6835 · 2026-10-02
- Dev's Advice: Build While AI Mods Still Earn Your Dopamine, Before It All Gets Too Easy — EddyVGG · 2026-10-02
- Why this tech commentator wants OpenAI to win the AI race — VraserX · 2026-10-02
- Report: Anthropic pressed the Vatican to alter Magnifica Humanitas text on AI personhood — StewartalsopIII · 2026-10-02
- Americans now fear AI losing control more than AI taking their jobs — ben_j_todd · 2026-10-02