Curated insights and tools for curious minds.
๐
Date: 4 November 2026
โฑ๏ธ Read time: 5 minutes
1. ๐ฏ Highlight of the Week
- โจ Resource: Microsoft AI's diagnostic orchestrator, MAI-DxO, is now published in Science (22 October 2026). Microsoft AI announcement ยท Science paper, DOI 10.1126/science.aeh7930
- ๐ก Why it matters: Instead of acing multiple-choice exams, MAI-DxO works like a virtual panel of doctors that asks questions and orders tests step by step, weighing each test's value against its cost. On 304 complex NEJM cases, it reached 83% accuracy versus 22% for physicians without reference resources (36% with them), and on common US emergency-department cases, 80% versus 47%. I like the honesty in the fine print, though: costs are estimates, these are benchmark cases rather than real patients, and the real test (a prospective study with Beth Israel Deaconess) hasn't been reported yet.
2. ๐ Social Spotlight
- ๐ฌ Surgery with a sticky note instead of stitches: MIT spinout SanaHeal, founded by Hyunwoo Yuk from Xuanhe Zhao's mechanical engineering lab, makes a hydrogel adhesive that bonds to wet, uneven tissue within seconds, inspired by barnacles and spider webs. It already has regulatory approval to stop trauma bleeding and is set to enter human trials for surgical bleeding and tissue repair. As a materials person, I love seeing wet adhesion (a close cousin of tribology) turn into something that could make robotic surgery simpler. Source Link: MIT News, 20 October 2026
- ๐ฃ Google's AMIE meets real patients, in The Lancet: In a feasibility study at Beth Israel Deaconess Medical Center's primary care clinic, 98 patients chatted with the AMIE research AI before their urgent care visits. Its differential diagnoses matched the doctors' final diagnoses 90% of the time; clinicians said its summaries helped them prepare in 75% of cases, and no conversation had to be stopped on safety grounds. It's a first look, not proof: the authors say larger trials are needed. Source Link: Google announcement, 8 October 2026 ยท preprint, arXiv 2603.08448
3. โ๏ธ Tools & Resources
- ๐ง SDBench: The step-by-step diagnosis benchmark behind MAI-DxO is open source: the agent, gatekeeper and judge code and prompts, plus five synthetic NEJM-style cases you can run straight away. The full set of real NEJM cases requires a NEJM license, so this is more a research tool than a free dataset. Link: microsoft.github.io/sdbench
- ๐ ๏ธ NVIDIA Isaac Lab 3.0 and SimReady assets: Announced at GTC Berlin on 21 October, Isaac Lab 3.0 is now generally available as an open-source framework for training robot policies, from desktop to multi-GPU setups, with Newton and PhysX physics. Alongside it, the SimReady Central Library on Hugging Face offers more than 2,000 simulation-ready robot and factory assets from companies such as FANUC, Comau and Universal Robots. Free to use. Link: NVIDIA announcement
๐ป GitHub Repos of the Week
- nanochat (MIT licence, ๐ข safe in this week's review): "The best ChatGPT that $100 can buy." A small, readable codebase that walks through the whole chatbot-building pipeline, from tokenising text to pretraining, fine-tuning, evaluation, and chatting with your model. It's a brilliant way to understand how LLMs are actually made, even if you only read the code; actually training one means renting a multi-GPU machine for a few hours. Get it: github.com/karpathy/nanochat
- draw.io (GPL-3.0, ๐ข safe in this week's review): The free diagramming tool I use for figures, lab set-ups and system sketches. The easiest is the web app at app.diagrams.net; if you prefer the desktop app, keep it updated and only open .drawio files from people you trust, because a diagram can reference outside images and reveal your IP address. Get it: github.com/jgraph/drawio-desktop
4. ๐ Key Articles & Insights
- ๐ฐ LubZero: An active molecular generation framework for multiobjective properties of lubricant molecules โ By Luyao Bao, Rui Zhou, Xiangmeng Jia, Feng Zhou, Meirong Cai and Weimin Liu, Friction, published online 9 October 2026, DOI 10.26599/FRICT.2026.9441293 This is where AI meets tribology: the authors adapt AlphaZero, the game-playing algorithm, to design new ester lubricants from scratch, treating molecule building as a single-player game. Key contributions include:
- A graph neural network (GINE) combined with Monte Carlo tree search to assemble candidate molecules.
- A parallel self-play set-up on a shared GPU that explores faster than serial versions.
- The surrogate models predict candidates that will push the flash-point/pour-point trade-off beyond the range of existing experimental data.
It makes me wonder how many of these candidates survive the lab: the gains are predicted, not yet synthesised and tested.