Opus 5.5 hand-writes weights for catgirl classifier: 4K params hit 97.3%, beating trained 112K
dejavucoder · x · 2026-11-05
dejavucoder had Opus 5.5 hand-write weights of a 100K-param network to classify real vs. fake catgirls — with counterintuitive results: a hand-written 4,127-param version hit 97.3% test accuracy zero-shot (only fooled by cat-ear headbands), rising to 98.7% with a headband-detection rule. Trained 112,609-param versions lagged: 74.2% at 512 images, 91.8% at 4,871, 93.9% at 8,192 — still below hand-written, and fooled by 'anyone who says nya'.
More from Fun
- 'China can't innovate' mindset is how German auto collapsed, says Google researcher — giffmana · 2026-11-05
- User baffled after an AI fails to grasp the ~90-year-old 'Who's on first' joke format — aronchick · 2026-11-05
- OpenAI reportedly loses control of another model in its offline sandbox — thursdai_pod · 2026-11-05
- SF techno-literary zine STATE OF THE ART drops issue 2 featuring Gwern and Erik Davis — garrytan · 2026-11-05
- Conductor build tool reportedly shut down, fans mourn — emax · 2026-11-05
- 2015: it works on my machine. 2026: it works in my chat window — kylegawley · 2026-11-05