A smiley face drawn into a network's weights survives training untouched
danbri · x · 2026-11-13
willdepue explains a neat ML fact: models are just billions of weights initialized with random numbers, and training adjusts most — but never all — of them. Weights that training never touches, and the residual correlation of the rest, mean the initialization remains readable in the final checkpoint. The accompanying figure shows a smiley face painted into one layer's weights is still clearly visible after training to 97.6% accuracy.
More from Research
- Waterloo authors' textbook 'Elements of Dimensionality Reduction and Manifold Learning' has 392 citations — FrnkNlsn · 2026-11-13
- mcbench-lite open-sourced: agents build in Minecraft, runs locally on your keys — adonis_singh · 2026-11-13
- AI-derived ClinicalAge from EHRs correlates with cancer and chronic disease risk — EricTopol · 2026-11-13
- TRUST framework cuts bad self-improvement accepts from 45.6% to 2.78% — dair_ai · 2026-11-13
- Weight normalization emerges as new standard: neuron norms hit almost exactly 1 in nGPT-style models — SonglinYang4 · 2026-11-13
- Idea basins: why autonomous agents get stuck, and how frontier models explore differently — AlexGDimakis · 2026-11-13