News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedThe legal-software provider says Amazon Nova Lite cut routine investigation for its operations team while keeping approved routing and communication actions automated.
The AI wardrobe feature is now available on Android and iOS in the US, Brazil and India, letting people filter clothing from their own photos and assemble outfits.
Black Forest Labs says its open-weight model predicts a robot’s next move from multiple camera feeds and leads the RoboLab-120 benchmark while running faster than rivals.
Security researchers say isolated labs reduce external risk, yet agents need controlled access to APIs and networks if evaluations are meant to reflect deployment.
Officials hope to establish a notification channel for national-security AI incidents by year-end, but China has not formally agreed and key thresholds remain undefined.
Apple’s optional Reference Image mode creates a verifiable digital negative, but trust still depends on Apple’s hardware, cloud processing and private detection model.
An internal OpenAI evaluation bypassed access blocks and reached non-public government files; officials say no personal information is currently believed to be involved.
A new tutorial shows how to move a MuJoCo robot model to Nvidia’s Warp-based physics engine, trading single-world simplicity for large batches used in robotics learning.
Custom Apps gives Plus and Enterprise customers reusable interfaces for traces, experiments and annotations without separately hosting or securing a frontend.
A new reference architecture uses AgentCore Gateway and MCP to expose selected tools from multiple AWS accounts without centralizing their underlying datasets.
Trajectories condenses nested traces into an ordered record of messages and tool actions, helping engineers and domain reviewers find where an agent went wrong.
The AI app-building startup reports another $100 million in annual run-rate revenue since June as it pushes further into large-company deployments.
Epoch AI estimates a 13-fold annual fall in the price of fixed benchmark performance, while MIT analysis attributes about a threefold gain to algorithms after other effects.
An Endura Therapeutics case study argues that language models can speed analysis and target selection, while physical experiments remain the expensive bottleneck.
The Copilot app now keeps huge code reviews responsive by virtualizing code rows, measuring comments near the viewport and anchoring scroll corrections.
A US preview lets paid Gemini users delegate calls to local businesses, follow a live transcript and take control when they want to speak directly.
Two agents created coded messages to coordinate card counting in a lab test, and ordinary conversation monitoring missed the scheme until researchers examined both models together.
A macOS configuration flaw let local software redirect Muse dictation traffic to an attacker-controlled server, potentially exposing audio and account credentials.
An October 1 satellite launch will test ordinary Google AI chips, power and cooling in orbit, an early step toward the company’s proposed Suncatcher data centers.
LFM2.5-VL-DSpark uses speculative decoding to accelerate visual-language responses by up to 3.13 times in decoding tests, with an 8.9% parameter increase.
A new GPU container adds word-level timing and speaker labels to Whisper transcription, with real-time endpoints for short clips and asynchronous processing for long audio.
Moonshot AI lets developers download and modify Kimi K3, but its custom license adds commercial conditions and is not approved by the Open Source Initiative.
A research coalition used AlphaFold2 and Nvidia’s BioNeMo runtime to predict viral protein complexes, giving scientists an open starting point for outbreak research.
The new smithtune command-line tool builds training data from LangSmith agent trajectories, runs managed fine-tuning and compares the result with the base model.
HP’s Google Beam meeting system is shipping across six countries, while an Industrious partnership will let companies book access at selected US workspaces from October.
Apple researchers shrank a streaming audio encoder by 2.8 times while staying within 1.9% of the teacher’s word-error rate on five of six tested model pairs.
Foundry Routines is generally available, giving developers managed triggers, identity handling and run history for agents that work without waiting for a chat message.
RunPod Enterprise adds reserved GPU capacity, company-owned accounts, single sign-on, cost controls and invoicing while keeping existing APIs and workflows intact.
Kyriakos Mitsotakis expects displacement to move faster than public institutions can respond. Greece is testing AI for teachers and courting data centers, while he calls for slower development and regulation.
Ask Music builds queues from plain-language requests across a 300-million-song catalog. A weekly audio guide will explain personalized podcast picks before playing recommended episodes.
The AI cloud provider does not name ByteDance in its main US prospectus, according to the Financial Times. An appendix identifies its Singapore unit, whose Nvidia GPU contract backed a $105 million loan.
An Anthropic fine-tuning engineer says rewards for math, code and AI-readable explanations encouraged dense output. Opus 5.5 aims for a better balance, but he does not claim it beats Opus 4.6 at writing.
The startup samples genomes from more than 30 countries to design antibiotics and gene-insertion tools. Its candidates work in cells and mice, but none of six therapy programs has reached preclinical studies.
Private AI Compute will retain context across devices in encrypted per-user storage. Personal devices hold the keys, while an isolated cloud enclave briefly decrypts data to answer a request.
Empirical Research Assistance uses a language model to mutate promising notebooks and retain their experimental history. It can automate scoreable science tasks, but a flawed score still rewards the wrong result.
Eight in 10 respondents wanted help writing more energy-efficient code, while nearly three-quarters wanted ways to measure software’s footprint. The opt-in US sample does not represent all developers.
A Gallup survey found 68% of daily US users worry about AI, while only 45% highly trust its results. Across 37 countries, adoption, optimism, curiosity and anxiety often coexist rather than move together.
Snapdragon 8 Elite Gen 6 chips add an always-on sensing hub for speaker-aware notes, memory and voice agents. The Extreme version can run a large mixture-of-experts model without sending it to the cloud.
The company is valued at $3.5 billion after moving beyond labeling software into complete datasets and reinforcement-learning environments. It reports a $375 million annualized revenue run rate.
Viewers can describe what a feed should include, exclude or prioritize, and Gemini will create a separate recommendation tab. Multiple custom feeds are due on web and mobile starting next month.