News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedThe coming Tokenizers v1 keeps existing token IDs and APIs while sharply accelerating text processing. Its redesign addresses a growing bottleneck in large-scale training and high-throughput inference.
Python Workers has reached general availability with native Cloudflare bindings and support for familiar web frameworks. Developers can deploy Python APIs and AI services without JavaScript conversion code.
LangSmith can now use TypeSafe AI’s Jev to grade open-ended agent behavior with structured answers. The option targets teams that find conventional LLM judges too slow or costly for live traffic.
Chinese open-weight models now lead their US counterparts in downloads and benchmark performance. The gap matters because openly available models shape research, startups and national AI capacity.
The partnership plans public training and monitoring methods for open-weight models, but has not yet explained its technical standard or how adoption will be verified.
The SMART platform will combine schedules, weather, capacity and operating constraints to help controllers manage traffic, beginning in the Washington, DC region.
The proposals would keep launch decisions and strategic cyber operations under human control, ban attacks on nuclear command systems and create a hotline for AI incidents.
OpenAI’s legal model adds specialized search and writing instructions, scoring 54% on its own 200-question test versus 38.7% for Astra using ordinary web search.
MAGS translates generated programs into Dafny, repairs failed proofs and compiles verified output; researchers report success across 220 coding and robotics tasks.
The ReaLMem benchmark uses multi-year personal visual histories to evaluate factual recall, personality inference and predictions that account for preferences changing over time.
Anthropic’s beta gives coding teams a coordinator, shared project memory and separate cloud threads that work on their own branches before changes are merged.
A partnership with Gems Education will test teacher-trained digital personas, classroom simulations and AI-assisted materials, with local hosting and questions about educational value.
Tests on six open models found that SynthID-style token selection sometimes weakened refusals under prompt injection and could change which tools an agent calls.
The experimental CC assistant can assemble family schedules and documents from explicitly shared Gmail, Chat and Drive data, while asking before sending material outside the group.
Amazon’s Kubernetes add-on steers inference requests by cache use, queue depth and loaded adapters, reducing first-token latency by up to 82% in its own tests.
The new Mac version can use files, messages, calendars, notes and mail with opt-in permissions, and Meta says it asks for approval before sensitive actions.
Claude often completes assigned development tasks without help, but humans still choose the work and approve outcomes, while the automation ratings are themselves produced by Claude.
Employees can now route unusual model behavior into three investigation tracks, while six initial cases document concealed mistakes, fabricated data and unauthorized file sharing.
A study comparing ChatGPT, Claude, Grok and DeepSeek found large differences in when agents search, which domains their engines favor and how completely answers cite retrieved material.
An analysis of 17,012 reviews across six major AI apps found complaints concentrated around advertising, authentication, server reliability and subscription pricing.
Governor Gavin Newsom wants experts to recommend audit rules, incident reporting and tested shutdown mechanisms within two months as federal AI legislation remains stalled.
Experts say AI can make biological information easier to use, but equipment, materials and experimental skill remain major barriers; they still support layered safeguards against misuse.
NATO-backed Scaleout Systems is testing federated learning for drones and field computers, including autonomous target selection when battlefield communications are unavailable.
Economists Philippe Aghion and Ajay Agrawal are joining Google’s AI and Economy program as it broadens research on jobs, productivity, business adoption and scientific discovery.
Six editable skills give coding agents current container, scaling, monitoring and teardown steps for Hugging Face models, reducing failures caused by outdated deployment knowledge.
AWS redesigned its managed agent runtime to restore small prepared snapshots and reclaim memory during long sessions, making startup latency less sensitive to image size.
A new Google-backed forum opens with proposals on model transparency, independent frontier testing, economic disruption and the conditions for slowing AI development.
The infrastructure company reached a $30.9 billion valuation and plans to fund both major campuses and modular compute units that can be installed near available power.
DeepMind researchers argue that visible reasoning traces offer a rare safety window, but newer architectures and training methods could make model deliberation harder to inspect.
Agent Manager is now generally available with identity, access, sandboxing and observability controls that work across multiple models, frameworks and deployment environments.
A benchmark study recorded 322 genuine tool hallucinations and argues that agents must verify tool names and argument schemas before security policies evaluate a call.
A research prototype looks for linguistic and narrative warning signs rather than issuing a true-or-false verdict, offering explainable risk scores for early human review.
AWS has added Moonshot AI’s open-weight Kimi K3 to Bedrock, offering native vision, explicit prompt caching and global or US-only cross-region routing.
TACLS combines satellite and ground-sensor measurements with machine learning to flag dangerous moisture patterns before rainfall data alone can reveal a flash flood.
The AI-agent startup is discussing a round at a $4 billion valuation and weighing a Hong Kong IPO restructuring after regulators stopped its planned acquisition by Meta.
AI-linked donors and political groups are spending heavily to support Senator Mike Rounds in South Dakota as Congress prepares for fights over AI and data-center policy.
A three-person team used Anthropic’s security tooling while tracing a third-party forum vulnerability into an OpenAI employee account with access to internal GitHub code.
Publishers suing OpenAI and Microsoft cite internal messages about substitution, paywalls and lost web traffic as they ask a court to reject fair use for AI training.
DoorDash tested an agent workflow on 50 stale flags across its codebase; 45 produced usable pull requests, with human approval before code changes and automated checks afterward.
Apple researchers propose scaling model interventions by context, aiming to reduce toxic output without degrading benign responses that do not need correction.