AI Daily Digest · 2026-11-22
🔬 New AI Craft
1. MedChemBench separates structural accuracy, explanations, choices and predictions
MedChemBench grades structure reconstruction, evidence-based explanations, compound selection and property prediction separately, awarding partial credit for numerical closeness. Its 99 problems reconstruct stages of published research. Prior exposure to those sources cannot be excluded, and responses get one attempt without tools or new experiments; the scores do not establish improved laboratory decisions or novel drug discovery.
https://openai.com/index/introducing-medchembench
2. A working Modular Diffusers port still needs component-level parity checks
An Idol port illustrates converting weights into custom Modular Diffusers blocks, checking each component against the original with identical inputs in full precision, then comparing the low-precision pipeline and rendered outputs. Judge numerical drift against the original implementation’s own variability. This remains a custom integration using remote code, requiring code review and attention to each component’s redistribution license.
https://huggingface.co/blog/OzzyGT/using-new-models-with-modular-diffusers
🛠️ Tools & Tips
(no items this day)
⭐ Open Source Highlights
1. Recapture synthesizes document scans and photos with transformed layout labels
Recapture takes page images and layout boxes, applying scan degradation or 3D paper, lighting and camera simulation while transforming labels with the output. The Apache-2.0 project can diversify document-training data but is not an officially supported Nutrient product; check synthetic coverage against samples from the intended capture conditions.
https://huggingface.co/blog/nutrientdocs/recapture
📰 Industry News
(no items this day)
🚀 Major Releases
(no items this day)

