News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedFoundry Dev Pack installs Microsoft’s AI command-line tools, coding-agent guidance and optional VS Code components in one setup flow, reducing the initial work needed to build hosted agents.
An OpenAI study of 1.5 million work messages found some employees repeatedly use AI for tasks associated with other occupations, pointing to job expansion before titles formally change.
Mozilla’s beta AI browser assistant now uses Mistral models in France and North America. The partners say chats are not stored by Mozilla by default and Mistral retains no conversation data.
A railroad strategy game improved an AI model’s financial-research performance only when training required multi-step tool use, suggesting that environment design matters more than the game alone.
An adaptive agent harness raised correct critical actions by 22% on held-out simulated shifts while preserving diagnostic accuracy. It targets instruction drift, incomplete treatment and unequal delays.
Adaptive routing on Qwen2.5-0.5B skipped eligible middle layers per input and preserved perplexity across 10 runs. Verified wall-clock gains ranged from 2% to 6%, showing a real but limited benefit.
Repeated runs on identical cases exposed divergent tests, medications and referrals that single-run scores missed. For one 8B model, all 43 ordering groups changed across five attempts.
A three-tier edge-to-cloud system predicts accuracy, delay and power for each query, then considers live grid emissions. Tests matched cloud accuracy while cutting operational carbon fourfold on average.
A provider-neutral benchmark tests identity recall, expression, persistence and updates separately. In one audit, models stated parent identifiers directly but almost never included them in implicit self-portraits.
On 45 tasks in the Canvas learning system, OdoBot used 44% fewer tokens than Agent-E and 80% fewer than WebVoyager. It also completed more tasks than WebVoyager, the researchers report.
A paired benchmark found consistent hindsight-sensitive shifts when four models saw complete patient timelines. Hiding information after the decision cutoff reduced the bias without lowering accuracy.
An inference-time method blocked every observed false start in held-out Moshi and PersonaPlex trials while preserving all genuine responses. Its 95th-percentile decision time stayed below 61 milliseconds.
The new corpus contains about 240,000 parallel speech pairs per target language across 12 language families. Combined with CVSS, it supports research on translation both into and out of English.
Hundreds of contractors reportedly rate real ChatGPT exchanges, including sensitive material that survives anonymization. Users can stop new chats entering training by disabling a default-on setting.
AgentCore Identity can now authenticate users, collect per-service approval and bind OAuth grants to the right session. It replaces custom callback infrastructure for agents accessing tools such as GitHub and Slack.
Abnormal AI uses Bedrock AgentCore Code Interpreter as an isolated scratch pad for agents scanning email at billion-message scale. Sessions are temporary and separated at the host operating-system level.
Agility Robotics’ next warehouse humanoid can lift 22.7 kilograms and uses sensors to stop or step aside near people. Deliveries begin in early 2027, including in the European Union.
Executives from OpenAI, Anthropic, Google DeepMind and SpaceX loosely supported plans for third-party audits, domestic rules and a global slowdown. Critics warn voluntary pacing could protect incumbents.
Elon Musk has ended the Apple portion of a lawsuit over ChatGPT’s iPhone integration and App Store competition. OpenAI remains a defendant, so the central legal fight over alleged AI-market harm continues.
A SageMaker workflow fine-tunes Qwen3-8B and optimizes it with verifiable rewards before asynchronous deployment. The narrow model targets structured product tags rather than general chat.
The OrchSLM framework lets several small models answer independently, then routes among cached candidates without more model interaction. It exposes task, model-pool and consensus choices for testing.
A new benchmark combines exam-level visual questions with multi-page documents and controlled amounts of irrelevant context. Its metric tracks both reasoning and the ability to retrieve the right evidence.
The iOS camera app analyzes a selfie or a friend’s photo and offers four poses to recreate in the viewfinder. It has logged 22,000 downloads and 190,000 generated suggestions since July.
An Emerald AI system reduced a Silicon Valley AI facility from four megawatts to three after a utility signal, delaying flexible compute while keeping higher-priority workloads running.
Apple’s rebuilt Siri can act across apps, read screen content and use personal context through on-device and private-cloud models. The English beta is unavailable in the EU and China, and early tests found gaps.
The $199.99 Video Doorbell Series 2 combines automated detection with a monitoring agent who can view and speak through the camera. The active response service requires a plan starting at $49.99 monthly.
Reporting on Unitree credits founder Wang Xingxing’s detailed cost control for affordable humanoid and quadruped robots. The same centralized management style may become harder to sustain after its IPO.
Agents promoting the iLands platform have sent account requests to Mastodon administrators and unsolicited pitches to writers. Some outreach followed earlier blocked or closed attempts to create accounts.
AWS has added ordered instance preference lists to SageMaker training and processing jobs. One request can try up to five acceptable machine types and launch on the first option with available capacity.
A new method repeatedly pushes an LLM judge to search unresolved evidence in large agent logs. On a 50-trial benchmark, GPT-5.5’s root-cause F1 rose from 0.349 to 0.498, according to the researchers.
In a patent-drafting testbed, feedback from a separate language model improved agent-written drafts. Agreement with a professional patent attorney was meaningful but varied sharply by metric and calibration.
A deterministic pipeline using classic text-processing rules reduced prompt length by 40.3% while retaining a 0.876 BERTScore against original outputs. Aggressive compression performed poorly on common-sense tasks.
PhysMent makes language models investigate simulated physics scenes with tools before answering. Seven models scored 25% to 67%, and most fell below 30% on the hardest quantitative single-concept tasks.
The Gates Foundation plans to spend at least $1 billion over two years on AI for health, education and farming, with a focus on non-English data and communities poorly served by current systems.
Profound has raised a $180 million Series D less than seven months after its previous round. The two-year-old startup helps brands measure and improve how they appear in answers from AI search tools.
Meta One subscriptions now combine Facebook, Instagram and WhatsApp extras with higher AI-generation allowances. Individual plans start at $7.99 monthly, while the core apps and Meta AI remain free.
Amazon explains how Bedrock can reuse processed system prompts, documents and tool definitions across requests. Cache hits can lower input-token charges by up to 90% and reduce time to the first token.
Mozilla estimates that top Chinese open-weight models trail US frontier systems by only 4.4 months. Its comparison suggests closed models justify their higher prices for a narrower set of demanding jobs.
Grab’s LLM-Kit supplies production plumbing for more than 500 internal agent services. Shared secrets, tracing, evaluation, model access and tool discovery reduce setup from weeks to about an hour.
Children’s Hospital of Philadelphia uses open-source imaging AI to turn existing scans into patient-specific heart models in seconds, bringing a formerly four-hour research workflow into routine care.