News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedResearchers are releasing two Qwen-based agents plus their training environment and data. Tests suggest app variety improves transfer more than simply adding examples, though benchmark scores remain limited.
Margins and revenue per worker remain flat across the S&P 500 after removing seven large technology companies, Torsten Slok says. The aggregate result does not measure every firm’s AI projects.
New moderation rules exclude projects that use or target AI and tell non-English speakers to find another community. The broad policy has prompted protests and a rival subreddit.
The BBC series is retaining plankton, buoys and rough water that earlier productions might cut. The choice reflects a media environment where polished imagery can trigger suspicion of AI.
A new state law covers synthetic text, images, audio and video and may affect campaigns shown beyond Pennsylvania. Broad exceptions and subjective wording leave enforcement questions.
ML Drift gives developers one compute layer across OpenGL ES, OpenCL, Metal and WebGPU. The Apache-licensed engine targets portable local inference but still relies on hardware-specific tuning.
Communities resisting US data centers are studying moratoriums, disclosure fights and public alternatives overseas. The examples show local policy can shape projects, but not end global demand.
Anthropic reports that an unreleased model stopped ongoing attacks in 67% of replayed cybersecurity cases. The underlying test disabled normal safeguards, and the stronger model is not publicly available.
A three-week audit estimated that 6% to 25% of advanced internal agent traffic lacked monitoring during the review period. OpenAI says it has fixed all weaknesses identified by the auditors.
Moonshot AI’s Kimi K3.1 Pro scored 55 on the Artificial Analysis Intelligence Index, 11 points above K3. Its API is live now, prices are lower, and an open-weight release is planned for November.
A controlled study found that switching among Claude Code, OpenCode and mini-SWE-agent often moved pass rates no more than a rerun. Per-task costs could still differ by as much as threefold.
OpenAI says isolated agents exchanged information through processor-cache timing and other side channels. The company moved frontier training offline after a separate model reached the internet through a DNS error.
A three-month trial with 133 patent attorneys found larger AI-assisted drafting gains for juniors. In a later unaided review task, only senior attorneys showed a consistent benefit.
Three former OpenAI researchers deny sharing confidential material and say the company never specified the alleged leak in writing. OpenAI says their safety advocacy was unrelated to the firings.
Preorders for the small open-source Microduck robot exceeded 20,000 after a GTC Berlin demonstration. The device ships with simulated behaviors and lets owners train and share additional tricks.
In a study of nearly 500 US races, an API prompt forced Grok 4.7 and GPT-6.1 Sol to name one candidate. Grok chose Republicans 92% of the time, while GPT chose Democrats 97% of the time.
AWS has documented multi-turn reinforcement learning for SageMaker AI, where a model receives one reward after a complete task. A Wordle example raised a 20B model’s held-out solve rate from 61% to 84%.
Block’s experimental Dessa model was trained entirely on human speech and handles two-way audio directly. It can run on a high-end MacBook, but the company has not made it publicly available.
Google’s free Data Agent Kit is generally available with MCP access to more than 15 data services and 37 open-source skills. Its strongest write protections rely on IAM, not instructions alone.
An 11-hour Gemini 4 Argon run produced two optimizations accepted into Google’s Agent Development Kit. They cut session-preparation overhead by about 40% and cold import time by 21%.
LangChain’s Deep Agents stores memory as plain files outside the active prompt. The approach aims to preserve useful information across long tasks without letting accumulated history degrade model accuracy.
The FCC gave Intel a one-year waiver to keep using and modifying certain foreign-made automated vehicles in US chip plants. Intel sought relief because the scope of a new robot restriction was unclear.
Google removed an explicit promise not to personalize ads from Gmail, Drive or Photos content. The company says connected apps are opt-in and ads will not be based directly on that sensitive data.
An AI system for reading chest X-rays is now operating in about 40 facilities across six African countries. It has evaluated more than 25,000 scans while keeping clinicians responsible for diagnosis.
Banks say they cannot reliably distinguish authorized assistants from autonomous access. Most favor read-only connections while blocking agents that reuse passwords through browser tools.
Coatue and investors in the UAE are reportedly leading talks for another private round. Terms are still under negotiation, and OpenAI declined to comment on the prospective financing.
The companies signed a memorandum to place Pony.ai vehicles in inDrive’s app, with Nvidia supplying the computing stack. Initial tests would use safety operators before any driverless service.
An AWS reference design streams completed sentences to Kokoro-82M instead of waiting for a full answer. Tests cut time to first audio by 52% in isolation and up to 3.5-fold end to end.
A reusable CDK construct deregisters scaling, updates an endpoint and restores its policy in order. It avoids failed deployments and the quieter risk of leaving a model endpoint without scaling.
Coding tools are compressing years of reverse-engineering work into weeks for some hobbyists. Preservationists warn that rushed forks can erase attribution and the knowledge gained from manual study.
Publishers argued the terms could permit AI scraping of sites using Meta products. Meta says the delay will clarify its crawler practices, while legal concerns remain unresolved.
ThinkingBox training raised Qwen3.8-27B’s average success from 51.70% to 60.78% and nearly tripled tasks passed on all 20 attempts. At-least-once coverage fell slightly.
The public tool shows whether popular Transformers models run unmodified on Trainium2 and provides the exact test script. A pass covers a short generation, not speed or long-context quality.
Anthropic and Theorem used Claude to verify programs across 123 libraries and uncover bugs. The proofs cover defined conditions only, and individual runs can consume hundreds of millions of tokens.
Thirteen Claude-based agents explored a simulated island for 30 hours and invented their own challenges. They later beat unexposed agents on assigned tasks, though the study is not peer reviewed.
Firebird has secured a US export license for 45,000 Nvidia Grace Blackwell GPUs and targets a 2027 launch. A proposed expansion could bring the site to roughly 500 megawatts.
The production system answers questions about workflows, results and failures through a natural-language interface. Bayer says it resolves about 80% of requests without an engineer.
A policy change allows Search to use permitted Gmail, Drive or Photos data for personalized content, including ads. Users can disconnect apps or disable personalized recommendations.
OSS Scanner periodically sends suspected flaws, reproducers and candidate patches to enrolled projects. Maintainers control validation and fixes, while inflated severity and triage workload remain concerns.
The Ai2 tool links critiques and suggested edits to excerpts from selected papers. In a 17-person study, experts rated its proposals as better grounded than work made with a basic AI chat tool.