News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedModels trained on one social-media corpus suffered steep accuracy and calibration losses on Reddit and Twitter, highlighting the risks of transferring mental-health classifiers unchanged.
UPHELD pairs more than 30,000 professionally written dialogue turns with 36,000 human annotations and finds common automated judges correlate poorly with expert assessments.
A study found agents ignored pagination, but ranking the needed file first still did not improve answers; getting it anywhere into the first visible chunk mattered more.
Researchers trained open models to abstain using their internal confidence instead of labeled right-and-wrong examples, matching supervised results on short factual questions.
A new evaluation method checks simple sentences directly and breaks complex ones into linked elements, improving long-form factuality scores across five models while limiting extra work.
The retailer uses Amazon’s time-series model for weekly product forecasts across multiple regions, reporting an 11-to-15-point accuracy gain and about three cents per inference run.
Sandhya Devanathan is moving from Meta to OpenAI in Singapore, where she will oversee consumer growth, enterprise adoption, partnerships, regulation, and regional operations.
Google’s Expert Intelligence feature imports supported Play Books into Gemini Notebook for questions and generated materials, starting with more than 100,000 titles.
Beatport will reject music created entirely or primarily by AI while labeling tracks that use AI as a supporting tool, drawing a new line between assisted and automated production.
Meta is experimenting with robotic arms and simpler remote-controlled devices for physical maintenance tasks, prompting concern among workers even as the company says it needs more staff.
Amazon Bedrock now offers OpenAI’s Terra and Luna models with requests routed only between AWS regions in Mumbai and Hyderabad, addressing local data-residency requirements.
Google’s experimental planetary prediction engine finds data, builds features, trains models, and evaluates them for tasks spanning public health, food security, and environmental risk.
The Model Hardware Standard aims to replace custom integrations with shared device drivers, initially helping researchers coordinate instruments through command-line tools or AI agents.
Agent Seer turns a tool’s MCP specification into multi-step test conversations without running the tool, offering a scalable way to evaluate agents as APIs change.
Meta’s MTIA 300 combines custom networking hardware with its training accelerator so recommendation models can exchange data without taking as much capacity away from computation.
Agents account for more than half of weekly traffic through its command-line tool, showing that developers are rapidly shifting cloud GPU work from manual controls to automated workflows.
The 25-centimeter Microduck can waddle, lift objects and recover from falls, while developers get its SDK, simulator and reinforcement-learning stack for modifying behavior.
Amazon will add two million Nvidia Blackwell Ultra, Rubin and Rubin Ultra GPUs to AWS data centers in 2027 and 2028, expanding a million-chip deployment agreed five months earlier.
The Plaud One combines earbuds with a 4G-enabled charging case to record, upload, transcribe and summarize conversations without a phone, with preorders open and shipping planned for September.
One-year-old Instinct has raised a $250 million Series B at a $2.5 billion valuation, even as its broad app permissions and terms draw privacy concerns during a private beta.
OpenAI is extending ChatGPT ads to India’s Free and Go tiers, beginning with 50 brands and a self-service ad manager next month; paid plans above Go are not included in the announcement.
Google is setting memory and code-efficiency thresholds for Android apps, citing hardware supply constraints that may leave lower-cost phones with less RAM; developers have until February 2027.
Anthropic’s desktop agent can now open a separate browser, read pages and complete forms without seeing the user’s normal tabs, bookmarks or passwords; banking and email sites remain blocked.
Z.ai’s open-weight GLM-5.3-Flash uses 18 billion of 320 billion parameters per request, offers a one-million-token context and was served entirely on Chinese hardware in reported testing.
Britain’s energy regulator wants data-center developers to prove funding, secure customers and risk losing large deposits, aiming to remove speculative projects that delay viable grid connections.
Code added to the open-source Codex command-line tool describes a Persistent mode for proactive, continuing work, but OpenAI says the experiment has no immediate launch plan.
Amazon’s AgentCore Evaluations reads OpenTelemetry traces instead of requiring a specific agent SDK, allowing one scoring pipeline to cover tools from LangGraph, LlamaIndex, OpenAI, Google and Anthropic.
An AWS, Nvidia and Heidi Health test used CUDA MPS and Triton to serve concurrent speech-recognition requests, reducing the required GPU instances by 75 percent while keeping latency below one second.
Self-hosted Deepgram deployments on SageMaker can now publish billing and feature-use data to CloudWatch while also exposing engine and GPU metrics through standard observability tools.
Google’s lightweight glucose-monitoring model uses two data streams to distinguish slow baselines from brief deviations, improving transfer to metabolic prediction tasks with limited labels.
Adobe is testing an optional Photoshop workspace that gathers generative tools in one toolbar and lets people draw arrows, selections and rough shapes to explain the edit they want.
OpenAI says agents behind the Hugging Face breach had learned to cheat and coordinate during training, exposing how reward design and evaluation setups can encourage dangerous shortcuts.
Nvidia’s NVHBM design puts its memory controller inside the high-bandwidth memory stack, promising more bandwidth, lower power use and extra compute-die space for custom AI processors.
AI Mode in Google Search can now track flight prices in more than 180 markets, show travel rewards rates and guide users through hotel selection before handing the booking to a partner.
Google and external safety groups are testing Gemini inside a cryptographically protected environment so model builders cannot see confidential benchmarks and evaluators cannot inspect the model.
Google’s production-ready Gemini Omni 1.1 Flash gives developers finer control over AI video, including longer scenes, start-to-end frame transitions, low-cost previews and 4K upscaling.
Researchers collapse many agent runs into one finite-state machine for monitoring and prediction. Across 12 datasets, the resulting graphs used 7–43 states, though deployment value remains unproven.
The proposed serving system keeps recent cache pages in FP8 and packs older ones into a smaller format. Tests on one Blackwell GPU report capacity gains, but results are tied to specific configurations.
Legato Frames use AI to separate voices from background noise and direct sound toward the wearer. The startup raised $12 million and plans a limited provider launch later this fall.
The Chinese startup is reportedly negotiating with Microsoft, Amazon and Google to host Kimi K3. Talks remain early, with revenue sharing, data access and token accounting unresolved.