Meta's Papailiopoulos: Training's Future Is Agents That Compile Intent Into Fully Bespoke FLOPs
DimitrisPapail · x · 2026-11-13
Meta researcher Dimitris Papailiopoulos lays out a directional thesis: training is heading toward a "totally bespoke" future where you give the hardware and the desired outcome, and an agent acts as the compiler between intent and FLOPs.
Key points:
- The output won't be a better optimizer or cleaner architecture, but one ugly frankenstein artifact touching kernels, optimizer, architecture, data composition, data order — and possibly hardware.
- Everything is built exactly for your ask and your setup.
- "Beat Muon"-style benchmarks are the wrong ladder: they freeze everything else and force the agent to win inside a constraint box humans drew. Old nanogpt speedruns are similarly bad tests.
- He sketches the future interaction: "Claude, get 10x lower perplexity than that previous run, minimum time, optimize everything, just don't reward hack."
More from AGI Musings
- Anthropic's alignment work paying off as OpenAI falls behind, argues researcher — haider1 · 2026-11-13
- Claim: Senior researchers at OpenAI and Anthropic have confirmed contested study result — NathanpmYoung · 2026-11-13
- Investor tszzl: Chinese open-source models have earned huge global goodwill and soft power — tszzl · 2026-11-13
- 'We Have a Right to Compute': Critics Warn US AI Regulation Follows Nuclear Energy's Playbook — nptacek · 2026-11-13
- As Keywords Collapse, Translating Intent Into Language Becomes the Core Skill — AiJohnAllen · 2026-11-13
- Jacques Thibs: AI's 1.5-year capability jump came mostly from unhobbling — JacquesThibs · 2026-11-13