AI autoresearch compresses MNIST classifiers 2-3x better than SOTA
DimitrisPapail · x · 2026-10-07
Dimitris Papail argues that models doing autoresearch have already compressed classifiers on beaten-to-death problems like MNIST 2-3x better than state of the art — a task requiring ruthless iteration, heavy tool use, and weeks-long experimental context, beyond any human expert. He sees this as proof such capability now exists, though not yet at the highest level. Ofir Press responds that a precursor might be rewriting ffmpeg or sqlite to be 5x faster with 2x less memory.
Related event: Researchers Debate Milestones for AI Coding: Rewriting FFmpeg 5x Faster(4 posts)→
More from AGI Musings
- Gary Marcus Slams OpenAI's Vague Math Proof Report: Zero Details, Won't Pass Peer Review — GaryMarcus · 2026-10-08
- Anthropic alignment head's doom estimate implies 800M innocent deaths, quips stats lecturer — wfithian · 2026-10-08
- Rotman Professor Sums Up His AI and Management Course in One Post — Afinetheorem · 2026-10-08
- eigenrobot on Academic Fraud: Cover-ups and Threats Are the Norm, Not the Exception — eigenrobot · 2026-10-08
- Researcher estimates 25% chance some current AIs are already conscious — cccalum · 2026-10-08
- A debate on whether AI abundance is distributed equally to everyone — LeviTurk · 2026-10-07