Mage is a 4B open-weight computer-use agent that scores 66% on OSWorld-Verified, compared to 72.4% for humans. It completes a spreadsheet task in 39 seconds, while Qwen3.6-27B takes nearly 10 minutes when limited to clicking.
Key points
Release: Microsoft released Mage, a 4B open-weight computer-use agent.
Benchmark: Mage scored 66% on OSWorld-Verified and completed a spreadsheet task in 39 seconds.
Sources
Qwenputer Qwen3.5-4B with RISC-V CPU inside weights
Qwenputer runs Qwen3.5-4B with a full RISC-V CPU inside its weights, using Gated DeltaNet layers as RAM. On DeepMind math it achieves 97.5–100% accuracy, compared to 55–90% base and 95–100% with a Python tool.
Key points
CPU: Qwen3.5-4B contains a full RISC-V CPU inside its weights, using Gated DeltaNet layers as RAM.
Runtime: It runs C and MicroPython in its forward passes, no external runtime.
Benchmark: DeepMind math results are 97.5–100% accuracy, versus base 55–90% and base + Python tool 95–100%.
Sources
Bruce Booth article: AI drug design crowds 37 targets
Bruce Booth article says AI speeds new molecule design but drives crowding on 37 targets, about a quarter of R&D projects. It notes many firms are making me-too drugs and patient recruitment is strained.
Key points
R&D projects: about a quarter focus on 37 targets.
Publication date: article published October 30.
Sources
New paper from our lab - neural geometry HD-EEG study
New paper from our lab uses neural geometry analyses on HD-EEG to examine coarse-to-fine versus fine-to-coarse object perception. Led by ayaka_hach and Thomas Baumgarten.
Key points
New paper from our lab
Led by ayaka_hach and Thomas Baumgarten
Used neural geometry analyses applied to object category information in HD-EEG
Sources
- Original post
- jneurosci.org - Linked in the post