Latest

Microsoft folds coding and always-on agents into one Copilot app

The Arena

Microsoft folds coding and always-on agents into one Copilot app

Microsoft finally shipped the consolidated Copilot it promised in June — and the surprise is that an app sold to knowledge workers now writes and hosts internal apps. Two other stories worth your time: ICLR's second anonymity incident in two cycles, and Berlin's AI street cameras, which are not judging anything yet. Microsoft unveiled the Copilot "super app" it has been teasing since June: one app with three tabs — Home, Code and Autopilot — that merges Copilot Chat with Cowork and rebrands the

Peking University's 591-run RL study finds uniform sampling wins

The Frontier

Peking University's 591-run RL study finds uniform sampling wins

A Peking University group built the comparison platform, ran the experiments, and reported the boring answer: with everything else held constant, the data policy is not what moves the score. The sharper finding is that the benchmark summary you choose can flip which method looks best. Peking University's Data-Centric AI group ran 591 reinforcement-learning experiments to test whether cleverer data policies beat plain uniform sampling, and could not find one that reliably does. The paper, DataFl

Medicare's AI pilot pays contractors a share of every denial, records show

The Guardrails

Medicare's AI pilot pays contractors a share of every denial, records show

The administration's Medicare experiment pays private companies when they refuse care — and its own actuaries wrote that down months before a Senate hearing where the man running the program said they don't. Medicare's WISeR pilot pays its contractors a share of the money they save by denying care, and the administration's own actuaries said so in writing. The program — Wasteful and Inappropriate Service Reduction — launched in January in six states: New Jersey, Ohio, Oklahoma, Texas, Arizona a

Goncourt drops the prize novel accused of being AI-written

The Everyday

Goncourt drops the prize novel accused of being AI-written

France's literary season just produced its first casualty of AI, and nobody involved can actually prove authorship either way. The Académie Goncourt withdrew Thélyson Orélien's novel C'était ça ou mourir from its selection on Friday 25 September, four days after the book won the Prix du Roman Fnac. The academy said the decision was driven by "the plagiarism of which its author has been guilty over time" and by "the convergent result of several analyses by credible researchers and journalists" c

Alibaba's CUDA alternative opens to outside developers

The Stack

Alibaba's CUDA alternative opens to outside developers

Alibaba spent its Apsara conference showing off silicon. The more consequential announcement came the day after, when it opened the software that decides whether anyone can actually use it. T-Head, Alibaba's chip subsidiary, published a fresh round of open-source releases for T-Head SAIL on September 23 — the CUDA-style stack that sits between PyTorch and its homegrown Zhenwu accelerators — one day after unveiling the Zhenwu V900, which Alibaba CEO Wu Yongming called the most compute-capable AI

The Pentagon wants $30 million for an AI lie detector

The Guardrails

The Pentagon wants $30 million for an AI lie detector

The US military is polygraphing its own officers over press leaks, and now it wants machine learning to improve the instrument. Also today: the creator of Ruby on Rails retires from hand-writing code, and Gemini starts phoning businesses for you. The Pentagon has asked Congress for $30.3 million over five years to build an AI-powered successor to the polygraph, to be run by the agency that screens federal employees. The line item sits in the Defense Department's FY2027 budget justification for

Open Source Radar — September 25: memory, method, smaller models

The Stack

Open Source Radar — September 25: memory, method, smaller models

Today's trending list splits three ways: what an agent remembers, how it works, and whether any of it fits on hardware you already own. Hindsight (Python, ~28,400 stars, MIT) — Vectorize's memory layer is one of the most-trended repos on GitHub today, and the distinction it draws is the useful part: most agent memory is conversation history, while this is built so an agent's beliefs change as evidence accumulates. Facts go in through one call, get consolidated in the background into observation

A Chinese firm built a city traffic model on satellite imagery

The Arena

A Chinese firm built a city traffic model on satellite imagery

A Shenzhen-listed location-data company decided the sensor city traffic was missing was never on the ground. Its answer went on show in Hangzhou this week. Merit Interactive (每日互动), a Shenzhen-listed location-data firm, has released MUMO — Modeling of Urban Mobility — a city traffic model that fuses satellite remote sensing with ground-level traffic data. The company announced the model at the fifth Global Digital Trade Expo, which runs September 23–27 in Hangzhou and drew more than 2,000 exhib

Deep Dive — The US built an AI gate with no law and no staff

The Guardrails

Deep Dive — The US built an AI gate with no law and no staff

The White House asked OpenAI and Anthropic not to share new models with Britain's AI Security Institute until the US government has reviewed them first, and Anthropic has already complied. Politico's account rests on one person familiar with the matter and a senior administration official, both granted anonymity; the White House did not respond to a request for comment. The concrete part is the product note. Anthropic's Claude Mythos 5.1 shipped with an availability line saying it is "only avail

Copilot made Okta's engineers faster but not more productive

The Stack

Copilot made Okta's engineers faster but not more productive

A year-long field study inside Okta finds the individual gains are real and the organisational ones are not there. Plus: the Jev team ships a code reviewer built on the assumption that nobody reads a 230-file diff. A peer-reviewed field study inside Okta found that GitHub Copilot cut engineers' working hours and sharply lifted their motivation and perceived skill — while producing no statistically significant change in the code the company shipped. The paper, published in Communications of the

AI 101 — What is federated learning?

The Stack

AI 101 — What is federated learning?

Federated learning is a way to train one shared AI model across thousands of devices or organisations without ever collecting their data in one place — each participant trains a copy on data it already holds, and only a summary of what that training learned travels back. The name comes from the word "federation": a group of independent members cooperating under one agreed protocol while each keeps its own affairs. The data stays where it lives. The learning moves. Why it matters right now Tw

Shanghai lets banks put self-trained models in front of customers

The Guardrails

Shanghai lets banks put self-trained models in front of customers

Shanghai's banking regulator just published the rules it will enforce when a bank's own model talks to a customer — and Moonshot turned the browser into something an agent can remember. Shanghai's financial regulator has issued 16 measures that, for the first time, sketch a route for banks and insurers to put generative AI in front of customers. The document — 沪金发〔2026〕19号, signed September 18 and published on the National Financial Regulatory Administration's Shanghai bureau site on September

See all