beat · 154 stories
A paired study of two small classifiers that decide which model, tool, or action an LLM agent uses next, and the three errors the authors found in their own analysis.
The program that pays outside researchers to find flaws in Google's open-source code is paused, but supply-chain reports and other bug-bounty programs (VRPs) remain open.
Stanford's 2021 paper on Homa, a Linux kernel transport for short-message datacenter traffic, resurfaced in a 2026 ai.engineer talk reframed for AI clusters.
Taiwanese battery maker ProLogium's Dunkirk plant, Europe's biggest battery project since rival Northvolt's collapse, is being certified to a Chinese spec and may be Europe's most realistic shot at mass production by 2030.
The AI Agent Accountability Act, introduced October 1, would extend the 1986 Computer Fraud and Abuse Act to cover autonomous software, with state attorneys general given an independent enforcement track.
MAI-Transcribe-2-Streaming, MAI-Voice-2.1, and MAI-Voice-2.
A decade of canine work at Michigan State shows adult retinas can rewire after a single virus-delivered gene correction, with the next test being whether the dog result carries to children.
DeepSeek's five new open-source libraries target Huawei's Ascend chips — Huawei's domestic AI accelerators — and sidestep Nvidia's proprietary CUDA programming layer, while Manus 2.
Ukraine has built 100,000 interceptor drones at $2,100 a shot and demonstrated remote control from 2,000 km away. Now Russia is deploying faster jet-powered attack drones that may break the cost-asymmetric model.
The 2nd Cavalry Regiment calls the in-house strike a service first, with soldiers assembling a commercial first-person-view (FPV) quadcopter and 3D-printing munition casings the unit calls 'Dragoon' bombs.
When the state stops paying for water towers and hospitals, data center operators step in. The deal terms are being written county by county, and rural counties have no comparable counterparty.
A 125-billion-parameter model running at human reading speed on a card a working developer already owns is the latest round of a quieter shift: each consumer GPU generation pulls more of the frontier off the API rails and onto the local machine.