ML professor predicts agents will become compilers between your intent and FLOPs
DimitrisPapail · x · 2026-11-13
Duke ML professor Dimitris Papailiopoulos shared his vision of AI-driven optimization: we're moving into a future that's totally bespoke — you specify your hardware and desired outcome, and the agent becomes the compiler between your intent and the FLOPs.
The output won't be a cleaner optimizer or architecture, he argues, but an ugly frankenstein artifact touching kernels and beyond. He gives an example: asking Claude to hit 10x lower perplexity at minimum time, optimizing everything — "just don't reward hack."
More from AGI Musings
- MIT Press quote sparks agreement: embodiment as direct openness to the world is what AI lacks — AnnaCiaunica · 2026-11-13
- The Singularity Clock Predicts AGI by December 2026, ASI by 2033 — Delam2 · 2026-11-13
- Lean proofs test Michael Levin's bioelectricity theory: math holds at 88%, 'Platonic' extra scores 0% — ProfBuehlerMIT · 2026-11-13
- Design and UX jobs down as AI debate heats up — round · 2026-11-13
- Almost no one notices Fable 5.2's gains—evidence against near-term foom — adam_karvonen · 2026-11-13
- Carl Edwards' AssayBench talk: predicting biology and what to measure next — kchonyc · 2026-11-13