Researcher: frontier agents build ugly frankenstein artifacts that beat anything humans design
DimitrisPapail · x · 2026-11-13
ML researcher Dimitris Papail argues that researchers love telling stories about elegant algorithms, but a frontier agent building bespoke solutions without human aesthetic biases will crush any human-made inference stack, optimizer, or model. Agents only care about hitting the reward, not the story—thanks to superhuman attention spans and tireless trial-and-error. He's repeatedly lost bets against this since March, and adds a spicy take: "the best optimizer is Opus 5.5."
More from AGI Musings
- Dean Ball frames 'gradual disempowerment' as resources shifting from human to machine demand — tszzl · 2026-11-13
- FAIR researcher Dimitris Papail: burn your TCS/math hat faster, we're entering a 'no ideas' era — DimitrisPapail · 2026-11-13
- Asterisk Essay: Humans Are Shifting from Narrative-Based to Game-Like Thinking — Scobleizer · 2026-11-13
- Lenny's late-2026 tech jobs report: PMs winning, engineers plateauing, design at record low — lennysan · 2026-11-13
- Benjamin Bratton's Antikythera launches Agentworld journal issue with MIT Press — round · 2026-11-13
- ML professor predicts agents will become compilers between your intent and FLOPs — DimitrisPapail · 2026-11-13