AI News · The field notes
The Daily AI Drop - Morning Edition, November 10, 2026
A new benchmark shows AI still can't do real professional work, while a tiny local logic model, a seventy-year drawing archive, and a first-agent tutorial point to the practical middle ground.
A reality-check morning: a new benchmark shows AI still can't do real professional work, while a tiny local logic model, a seventy-year drawing archive, and a first-agent tutorial point to the practical middle ground. Use the tools where they actually work, keep the judgment for yourself, and learn to build with them.
Listen: Listen to this episode
AI Can't Handle Law or Finance Jobs Yet — aish_reganti
GDPx, a new benchmark from SemiAnalysis, tests AI agents on 206 real professional tasks across 11 fields, including finance, law, chip design, and cybersecurity. The best model overall, Claude Opus 5.5, scored 12%, and no model fully completed a single finance task. The finance tasks were built around a reconstructed private equity deal — a fund buying roughly 62% of a snack company — with all the messy documents and judgment calls a real deal involves. If you've been wondering whether AI is about to replace lawyers and bankers, the honest answer this morning is: not yet, not by a long way.
TwIL LM3 Pro: Logic Model Locally on Contradictions and Reasoning Puzzles — fahdmirza
TwIL LM3 Pro is a 3.66-billion-parameter reasoning model built specifically for formal logic tasks, and the key word is "locally" — a small model you can run on your own machine, designed to work through contradictions and reasoning puzzles where big general models often stumble. It's a fun counterpoint to the benchmark story: while giant models score 12% on professional work, a tiny specialist is quietly getting good at one narrow thing. Small, focused models that actually work, running on hardware you own, are a pattern worth watching.
Kuken Industries searches 70 years of drawings in minutes — OpenAI
Kuken Industries is using the OpenAI API to help engineers search 70 years of drawings in minutes. The system was built with Recursive, a Japanese AI startup and an official OpenAI partner. Nobody is replacing the engineers here — they're getting more time to mentor younger colleagues and take on harder work, because the machine handles the searching. Seventy years of drawings, searchable in minutes: the kind of win that shows up in a real business, not just a demo.
Microsoft Foundry with Adrian Twarog: Build Your First Agent — MicrosoftDeveloper
Developer and YouTube creator Adrian Twarog takes Microsoft Foundry from first login to a published agent: create an agent, test it in the playground, deploy models, connect an MCP server with scoped permissions, and call the finished agent from the Azure SDK. Microsoft Foundry is where you build, test, and publish AI agents, with usage and billing tracked in Azure. If you've been hearing about agents all year and wondering where the on-ramp is, this is a solid one — especially the part about scoped permissions, because an agent that can only touch what it's supposed to touch is an agent you can actually trust.
AI Is Changing Software Engineering Forever (Here's What Matters Now) — pvergadia
On the Cloud Gold Podcast, Richard explores the future of software engineering, AI agents, developer relations, and the skills developers need to succeed in the AI era. When AI coding agents can write code, build applications, and automate repetitive work, what skills are left for the human? The work is shifting from writing every line yourself to directing agents, checking their work, and owning the judgment calls they can't make.