AI 新闻 —— March 26, 2026
PulseAugur 当天浮现的 20 条头条故事 —— 综合实验室、论文及开发者社区的信号进行排序。
-
Judge blocks Pentagon's 'punitive' AI supply chain risk label on Anthropic
A federal judge has blocked the Pentagon's attempt to label Anthropic a supply chain risk and sever government ties, ruling the move violated the AI company's constitutional rights. The judge found the designation, which would have required other companies to prove they weren't …
-
Anthropic wins preliminary injunction against U.S. Department of War
Anthropic has secured a preliminary injunction against the U.S. Department of War. The court order prevents the Department of War from taking further action in a case brought by Anthropic. Details regarding the specific nature of the dispute and the grounds for the injunction ar…
-
Anthropic updates subprocessor list, impacting data handling practices
Anthropic has updated its subprocessor list, detailing the third-party services it uses to operate its AI products. This transparency aims to inform users about data handling practices and compliance with privacy regulations. The changes reflect Anthropic's ongoing efforts to ma…
-
Anthropic updates session limits for Claude models
Anthropic has announced updates to its session limits for Claude, its AI assistant. The company is implementing new measures to manage usage and ensure a stable experience for all users. These changes are intended to prevent abuse and maintain the quality of service.
-
OpenAI's Spud and Anthropic's Urgency Model Emerge Amidst ARC-AGI 3 Benchmark
OpenAI is reportedly developing a new AI model codenamed "Spud," with CEO Sam Altman shifting responsibilities to focus on its development. Concurrently, Anthropic is preparing a model that is expected to prompt governmental urgency, potentially in response to the challenging AR…
-
Anthropic's Claude Code limits adjusted down to 5 hours
Anthropic has reportedly reduced the daily usage limits for its Claude Code assistant. The adjustment appears to have decreased the 5-hour limit for code-related tasks, though the exact nature and duration of this change are not fully detailed. This modification may impact devel…
-
New 'AI;DR' acronym reflects growing distrust of AI-generated content
A new internet shorthand, AI;DR (AI; didn't read), has emerged to describe the growing tendency to dismiss content suspected of being AI-generated. This reflects a societal shift from judging content by its length to questioning its human authorship, externalizing responsibility…
-
Microsoft Research launches AsgardBench to test AI agents' visual planning adaptation
Microsoft Research has introduced AsgardBench, a new benchmark designed to evaluate the ability of embodied AI agents to adapt their plans based on visual feedback. The benchmark consists of 108 task instances across 12 types, requiring agents to revise their actions as tasks pr…
-
Judge questions Pentagon's move to blacklist AI firm Anthropic
A federal judge is scrutinizing the Pentagon's decision to label Anthropic a national security risk, potentially impacting the AI company's ability to secure government contracts. Judge Rita Lin questioned whether the government's actions, which extend beyond simply ceasing to u…
-
Microsoft Research develops benchmark for robots to plan and execute tasks with spatial grounding
Microsoft Research has introduced GroundedPlanBench, a new benchmark designed to evaluate the ability of vision-language models (VLMs) to perform long-horizon task planning for robot manipulation. Current VLM-based robot planners often struggle with complex tasks due to ambiguit…
-
Two Minute Papers explores a new algorithm that evokes strong emotional responses.
This YouTube video from Two Minute Papers discusses an algorithm that reportedly evoked an emotional response. The video's creator expresses that the algorithm made them cry, suggesting a significant emotional impact. Further details about the algorithm and its capabilities are …
-
Mistral AI releases new developer tools and resources
Mistral AI has released a new video showcasing their latest advancements, though specific details about the model or its capabilities are not provided in the announcement. The video appears to demonstrate new features or performance metrics, hinting at progress in their AI devel…
-
AI agents leverage CLI tools for complex task automation
Large language models are increasingly being integrated with Command Line Interface (CLI) tools to enable agents to perform complex tasks. These agents can execute sequences of commands, such as organizing and resizing files, or interacting with services like Stripe and AWS. By …
-
Percepta Research builds LLM-Computer by converting programs to transformer weights
Percepta has published a blog post detailing their work on constructing an LLM-Computer, which aims to transform traditional programs into transformer weights. This approach seeks to bridge the gap between symbolic programming and the neural network architecture of large languag…
-
NVIDIA GeForce NOW adds five new games including Honkai: Star Rail and Screamer
NVIDIA's GeForce NOW cloud gaming service has added five new titles for streaming, including the retro-style racing game "Screamer" and the latest update for "Honkai: Star Rail." Other additions include "King's Quest," "BATTLETECH," and "Despot's Game." These games are now acces…
-
Schizophrenia genetics reveal tradeoff vs. failure components, applicable to human conditions
Scott Alexander's Astral Codex Ten explores the concept of "tradeoffs versus failures" as a framework for understanding negative outcomes. He argues that many complex situations, from psychiatric conditions to poverty and even physical illnesses, can be analyzed as either a resu…
-
rses CLI tool bridges Claude, Codex, and OpenCode for coding tasks
A new command-line tool called "rses" has been released, enabling users to seamlessly switch between different coding models like Claude Code, Codex, and OpenCode within a single session. The tool reads session data from Codex, constructs a structured handover, and initiates Cla…
-
Lobsters re-packs Linux distros with ZFS, WireGuard, and NVIDIA drivers
kldload has released version 1.0.4, a tool that allows users to assemble custom Linux distributions with ZFS on root, WireGuard, and eBPF pre-baked. The system pulls packages directly from vendor repositories, ensuring that the resulting distribution is stock and unmodified. It …
-
Data scientists' core skills are essential for AI harness engineering and evaluation
The role of data scientists is evolving with the rise of large language models, shifting from direct model training to a focus on the "harness" that guides AI systems. While foundation model APIs reduce the need for traditional predictive modeling, crucial tasks like setting up …
-
METR red-teams Anthropic's agent monitoring systems, finds novel vulnerabilities
METR collaborated with Anthropic to conduct a three-week red-teaming exercise on Anthropic's internal agent monitoring and security systems. The collaboration, which involved providing researchers access to internal systems, identified several novel vulnerabilities that have sin…