Claude hand-writes a neural net's weights to catch fake legal citations, beating trained models
141_1337 · reddit · 2026-11-04
A Reddit post shows a counterintuitive experiment: Claude wrote every weight of a neural network by hand — zero training — to detect fake legal citations, scoring 97.8% test accuracy vs 94.6% for a conventionally trained model (after 8,192 labeled examples).
The comparison table is striking: the hand-written version saw 0 labels, uses only 3,793 non-zero params, and hits 97.8%; one revision pass (5,062 labels seen) lifts it to 98.9%. The trained baseline needs 107k params and scores 76.5% at 512 labels, 94.6% at 8,192.
A follow-up demo (CiteNetWall) runs on Apple's Neural Engine: 2.27ms per 64KB chunk, 50μs per citation on M5 Pro, 1.8M citations/s across ANE+GPU, scanning 3.3GB of case law in 2.3s — catching all 31 planted fakes in the LePhantomCite holdout with just 1/983 false flags. Suggests LLMs can directly "program" small models instead of relying on training pipelines.
More from Models
- Ripple claims $32M compute for open-weight coding SOTA; skeptics question the math — teortaxesTex · 2026-11-05
- Gemini 4 Flash (codename Hydrogen) spotted A/B testing in AI Studio and LMArena — lyraxana · 2026-11-05
- Perceptron's new 3D hand model beats ACE-Ego-Hand by 12-25% on HOT3D and ARCTIC — gui_penedo · 2026-11-05
- 0.8B open-source HaloGuard 2.0 claims streaming-guard SOTA, beating models 5-10x its size — const_reborn · 2026-11-05
- MolmoAct 2 Takes Second to GPT-6 Astra in RoboColosseum's Real-World Robot Tests — DJiafei · 2026-11-05
- Claude Max subscriber at $100/month says Opus 5.5 'almost feels unlimited' — prasenx · 2026-11-05