Research

SemiAnalysis: 3.6% of 857 releases from nine Chinese AI labs had developer safety results: SemiAnalysis analyzed 857 AI model releases from nine Chinese AI labs between 2021 and September 2026. The study found that only 3.6% of releases included safety results from the developer, and only 1.1% had such results at launch.

AI Safety • Techmeme

Study finds AI coding agents generate more code, not more software: A study finds AI coding agents generate more code, but not more software. The research says coding efficiency gains get absorbed by a human review bottleneck.

AI Research • Ars Technica AI

Over 36 mathematicians tell The Verge they need years to assess OpenAI's math results: OpenAI abruptly released a flood of mathematical results this week. More than three dozen mathematicians told The Verge they need years to make sense of the scale, describing it as 'staggering,' 'overwhelming,' and 'pure insanity.'

AI Research • The Verge AI

OpenAI fires three researchers for leaking confidential data: OpenAI has parted ways with three researchers who allegedly leaked confidential information to an outside AI safety organization. A fourth researcher has also departed the company's safety team amid the shake-up.

Business • The Decoder

Gemini 4 Argon: Google has pre-announced its upcoming frontier-level artificial intelligence model, Gemini 4 Argon. While public outputs are not yet available, the model introduces architectural and performance changes aimed at positioning Google back among the top three AI research labs.

Google • Sam Witteveen

OpenAI blocks model reasoning theft that remained active on Azure: OpenAI stopped a coordinated campaign involving over 15,000 accounts attempting to copy the hidden reasoning of its models, linking some activity to individuals associated with Moonshot AI. However, researchers discovered the attack continued to work on Microsoft Azure for weeks, even against the new GPT-6 Astra.

Security • The Decoder

Researchers present Praxa AI agent harness with 100% test pass rate: Researchers have presented Praxa, an agent harness designed to govern AI execution through deterministic admission, brokered execution, and reviewed promotion. In initial testing, the system passed 1,027 out of 1,027 unit tests and 89 out of 89 Workerd tests at a pinned revision.

Infrastructure • arXiv

Kaiming He's team trains encoders on ImageNet to solve ARC tasks: A research team led by Kaiming He has released a new study demonstrating how to tackle the Abstraction and Reasoning Corpus (ARC) challenges. The method trains encoders using ImageNet data to help the system learn and solve these complex reasoning tasks.

AI Research • 量子位