News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedGoogle’s research prototype lets teachers generate guided STEM simulations from learning goals, using automated checks and teacher review to catch broken or inaccurate activities.
OpenRouter measured real charges for 20 image models and found default generations ranging from $0.006 to $0.134, while price alone did not predict text-rendering quality.
The UN System Data Commons combines official statistics in one searchable knowledge graph, with natural-language queries and tools that let research agents retrieve source-backed data.
A global fintech moved its GLM-5.2 coding assistant to self-service dedicated inference, letting engineers scale capacity and tune endpoints without waiting on support tickets.
The open-source framework defines text, code, images, embeddings and statistical fields in one declarative file. A preview loop lets teams inspect a small sample before paying to generate at scale.
CapMem contains 33.7 hours of annotated video and 1,000 questions. On clips longer than 20 minutes, caption-based memory beat direct video analysis for most tested models while using a reusable text record.
The system keeps competing plans in a tree and adjusts effort after each result. Across 12 AI research tasks, it reported 10.3% higher reward than AutoResearch while using 50.6% fewer attempts.
Across 448 English and Chinese conversations, some model configurations stopped entirely while others stayed available but withdrew from the task. Availability did not always mean useful work continued.
A study found that length-based complexity estimates classify equivalent African American English and second-language writing as easier. The routing choice compounds accuracy gaps already present across model tiers.
A benchmark across receipts and forms found that large open models extract key-value pairs well from clean text, but model differences shrink as character recognition gets noisier and corrupts numbers.
Causal tests across four models found that relation information, such as “capital of,” controls generation much earlier than the specific answer. Entity information exists early but reaches the output later.
AWS tested NVRx with PyTorch training across two to eight H100 nodes. Asynchronous checkpoints reduced idle time, while separate restart mechanisms recovered from software faults and hard process crashes.
AWS released markdown-based instructions covering 11 healthcare and life-science fields. In a 410-prompt evaluation, agents with the skills won 70% to 86% of comparisons against the same systems without them.
An AWS pipeline inserts people into real machinery images and labels them automatically. Tests reported up to a 160% gain in person-detection accuracy, though synthetic results still require real-world validation.
MRH Trowe combined an open chat interface, agent SDK and isolated runtime to reduce shadow AI. Its first-month deployment cost about $14 per seat and keeps tools inside a centrally managed environment.
The energy-data company built APEX to share identity, memory, observability and guardrails across internal and customer-facing agents, while allowing teams to change frameworks and models.
Specs Intelligence can connect accounts such as Gmail and Slack, surface upcoming decisions and organize travel details. An iOS preview is available now, with Mac early access and AR glasses integration planned.
Groups such as METR and Redwood Research are gaining access after a serious agent incident, but their influence still depends on cooperation from model companies and uncertain political support.
The planned Luna glasses would use six microphones and a physical button for Meta AI and Muse, without an integrated camera. The design could reduce visual surveillance concerns but still raises audio privacy questions.
Both labs say independent researchers could inspect more than finished models. Evaluators welcome access to checkpoints and logs but say funding, publication rights and legal authority will determine independence.
The former US vice president says new fossil-fuel plants remain a concern, but takes industry warnings about advanced AI more seriously. He argues climate impact should be judged against larger energy trends.
The Icelandic startup builds synthetic audio and virtual acoustic tests for models and hardware. Its latest funding brings the total raised above $40 million and supports expansion into robotics.
Two consumer AI assistants can now call businesses to book restaurants, join cancellation lists or resolve service problems. Instinct’s feature is in early access, while Muse is expanding a US beta.
An open map of 166,000 fruit-fly neurons is powering experimental apps, including a WIRED headline generator. The project shows how coding agents can turn scientific datasets into playful prototypes.
ERPBench evaluates screenshot-driven agents against records in a live enterprise system. Some agents saved forms in 85% of runs but wrote the correct value in as few as 3%, exposing hidden failures.
Researchers built one calibrated measure for comparing quantization and early-exit methods. Four-bit output matched the 16-bit baseline within test precision, while three-bit and math tasks degraded.
A Basel Action Network report estimates that equipment retired for AI could total 395 million to 617 million metric tons by 2050. Its broader count includes power, cooling and network systems.
Google, Nvidia, Anthropic, Emerald AI and utilities are backing demand response for data centers. They argue that shifting noncritical computing away from grid peaks could unlock more capacity.
Apple is reportedly developing an enterprise AI server with two or four future M8 Ultra chips for 2029. The project could reconnect Apple with a server market it left when Xserve ended in 2011.
AWS published a layered access-control pattern for MCP tools connected to Amazon Quick. It checks identity claims, location, roles and tool permissions before a request reaches business logic.
The SAFE benchmark tests whether models inspect optional risk evidence before acting. Severity and retrieval cost changed behavior strongly, while large probability changes had a surprisingly weak effect.
A controlled study across six Qwen3 sizes found supervised LoRA tuning strongest on familiar tool-calling tasks. Reinforcement learning had a small edge more often when transferring to different datasets.
A Pew survey of 42,151 people found that most countries expect AI to eliminate more jobs than it creates over 20 years. Sentiment is mixed overall, with concern outweighing pure excitement.
Huawei now expects its Ascend 960DT accelerator in the first quarter of 2027 instead of the third. The company is also building larger systems that link thousands of chips for training and inference.
OpenAI developer Eric Provencher says using more than two parallel sub-agents often creates duplicate checking and tool calls. He recommends isolated tasks that report back only when finished.
GPT-6 Astra is the first OpenAI model classified as Critical for cybersecurity capability. Tests found new browser and kernel flaws, while separate evaluations showed weaker visibility into its reasoning.
Plans for a White House-linked AI oversight body have been shelved, while competing bills lack votes and floor time. That leaves little prospect of broad US federal AI regulation before the midterms.
AWS compared OpenSearch, Aurora PostgreSQL with pgvector and S3 Vectors for Bedrock knowledge bases. The guidance emphasizes latency, hybrid search, existing data and storage cost rather than one default.
Apple researchers found that reset-free learning agents often become permanently trapped when actions cannot be undone. REVERSAL-BENCH measures that risk across eight tasks and five physics engines.
AIUC announced a $40 million Series A for its work on security and reliability standards for AI agents. The company aims to connect repeatable stress tests with insurance for enterprise deployments.