Doctor's Prescription OCR Test: Gemma 4 31B Now Reads What Models Couldn't a Year Ago
MaziyarPanahi · x · 2026-10-27
The author re-tested the same doctor's handwritten prescription that OCR models could barely read a year ago. Gemma 4 31B — the top model in his handwriting test — now correctly reads the scribbled "buen daay" as "seven days." He had previously tried the same image with DeepSeek-OCR.
He poses the question: would you let a model do the first read before the pharmacist? — raising the feasibility and risk of AI as a "first reader" in high-stakes settings like pharmacies.
Related event: Gemma 31B now reads doctors' handwriting that stumped OCR a year ago(2 posts)→
More from Models
- OpenAI claims its models disproved the Exponential Time Hypothesis, with Lean-verified proof — faraz0x · 2026-10-27
- OpenAI verifies cryptosystems safe from latest math result, trend alarming — IgorCarron · 2026-10-27
- DeepSeek Flash impresses dev: let it build its own tools, cheap enough to experiment — victormustar · 2026-10-27
- Reddit asks: Why do almost all major open-weight models come from China? — PhilosopherSweaty826 · 2026-10-27
- cyb3rops: open-weight AI models mean attackers can compromise 90% of orgs without new exploits — cyb3rops · 2026-10-27
- Study Finds GPT-6 Astra Only Partially Capable at CAD Workflows — ZhejiangUniversity · 2026-10-27