AINov 7, 2026The Verge
The US let Nvidia sell chips to an empty office in China
An approved H200 buyer whose China office is an empty shell with dangling wires is not a paperwork glitch — it is the export regime telling on itself.
Desk
Research
Papers, model cards, and lab research notes.
AINov 7, 2026The Verge
METR says AI is now driving most of Anthropic’s speedup, and the evidence is mostly redacted
After 11 weeks inside Anthropic, METR reports the lab’s models are improving about twice as fast as a year ago, mainly because of AI itself. Anthropic redacted most of the evidence METR found most informative, while still urging the industry to slow down.
AINov 5, 2026OpenAI
Measuring how AI agents handle everyday wet-lab science
OpenAI's BioBasicsBench, built with Ligo, scores agents on 100 messy wet-lab tasks — raw gels, bad controls, ambiguous images — and GPT-6 Astra leads at 66.7% while still failing ordinary judgment calls.
AINov 5, 2026NVIDIA Blog
The Wizard of Open Weights: Chris Alexiuk Makes AI Easy With Open Models
NVIDIA profile of Chris Alexiuk — YouTube's llm_wizard and self-styled No. 1 Nemotron fan — argues open weights, data, and recipes are the precondition for useful AI, not a nice-to-have.
AINov 5, 2026Microsoft Research
Formally verifying agent-written full-stack web apps with PDD Kit
Microsoft Research's PDD Kit turns intent into Verus specs so agent-written Rust — including a Pet Store login — is proved for every input, with tests only where proofs cannot reach.
AINov 4, 2026MarkTechPost
Google AI Research Releases Agile Deliberation Framework for Gray-Area Image Classifiers
Google Research open-sourced a human-in-the-loop way to turn fuzzy ideas like “healthy food” or clickbait into written policies that actually work as vision-model prompts — and beat zero-shot Gemini by about 10 F1 points.
AINov 4, 2026OpenAI
Franklin Templeton uses OpenAI to support investment research
Franklin Templeton is wiring Codex and OpenAI tooling into a "connected intelligence" layer so portfolio managers see diverse research views and risks — not a narrower feed of what they already like.
AINov 4, 2026NVIDIA Developer
Use NVIDIA NeMo to Turn Agent Failures into Tests and Harness Fixes
NVIDIA is pitching a disciplined loop for fixing agents: read the traces, write a test for the failure, change the harness, and prove it on tasks the fixer never saw.
AINov 3, 2026Microsoft Research
Learn the rules: The hacker mindset
A former DARPA hacking-contest lead says AI has broken the old math of cyberattacks, and his point about trust is the part security teams should hear.
AINov 2, 2026MarkTechPost
Deep Principle Releases Hydrogen-1.0: A 753B Open-Weight Model for Scientific Discovery
An MIT-licensed science model trained on real lab records and failed experiments is a meaningful gift to open research, if its self-reported wins hold up once outsiders test it.
AINov 2, 2026OpenAI
LifeVal
OpenAI built a benchmark around kindergarten forms, groceries and caregiving, a welcome shift toward real life that we can't yet check from the outside.
AIOct 30, 2026Microsoft Research
From synthetic enterprise data evaluations to real Microsoft Copilot AI confidence
Microsoft is testing Copilot against fake companies built from public data, a smart answer to the privacy problem that still leaves open whether fake offices behave like real ones.