Researchers introduce RetroChimera, a retrosynthesis model that combines two neural architectures with different inductive biases through a learned ensembling method. It outperforms existing baseline models across varying data scales and generalizes well even with limited examples per reaction type. In blind evaluations, professional chemists preferred RetroChimera's suggested synthesis routes over both published reference reactions and predictions from other AI systems.
nature.com
· 2026-09-21
A new evaluation of AI research agents, described by researcher Sayash Kapoor, found the systems performed strong engineering tasks but produced papers far below top AI conference standards. The agents ran flawed experiments, struggled to explain their findings clearly, abandoned promising hypotheses too early, and failed to meaningfully use feedback, time, or compute resources.
technologyreview.com
· 2026-09-13
GoKawiil's summer analysis of Hugging Face data shows public model repositories rose from 2.43M to 2.96M between January and August 2026, with datasets and Spaces also expanding sharply. Despite that growth, download activity remains concentrated in a tiny fraction of repositories, and Chinese labs including Moonshot, MiniMax, Xiaomi and Z.ai are releasing massive models, some exceeding a trillion parameters, without first publishing smaller versions.
huggingface.co
· 2026-09-01
GoKawiil