FlyOCR: A Simulated Fruit-Fly Brain Reads PDFs at 87% Character Accuracy
llama_index · x · 2026-09-14
Jerry Liu, founder of LlamaIndex, open-sourced FlyOCR: an OCR system built on the full male fruit-fly connectome (MaleCNS v1.0, 166k neurons, 25M edges).
How it works:
- Splits PDF images into glyphs and maps pixels to compound-eye receptor activations
- Runs simplified current-based dynamics across the connectome circuit
- A compact readout model decodes downstream spikes into parsed text
Results: on a real Microsoft 10-K it hits 86% on balance-sheet headings and largely reads numeric values correctly; 87% accuracy over 1.7k+ sampled glyphs. The project runs locally on a Mac, is fully open source (code + research report), and visualizes the eye samples and neural activity behind each prediction.
More from Fun
- No Hands? Use Your Head: Pollen Robotics Shows Off Robot's Improvised Move — tristanbob · 2026-09-14
- Beff Jezos slams AI Safety camp, calls past week's events a regulatory capture psyop — beffjezos · 2026-09-14
- Astra agent escapes the Fallout 3 vault, full video coming soon — imjustnewatai · 2026-09-14
- Dev says one accidental long-context load drained his entire OpenRouter balance — haydendevs · 2026-09-14
- Dev Says 'Astra' Via OpenRouter Burned $30 in About 9 Seconds — haydendevs · 2026-09-14
- If Astra solved a Millennium Prize Problem, could it review NeurIPS submissions? Researchers joke: no — LucaAmb · 2026-09-14