News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedA Reuters-based report says Meta paused a second layoff wave after its agents missed productivity goals and workers revolted. The alleged plan would have left smaller teams supervising virtual workers.
Alibaba’s Qwen4 architecture preview uses sparse activation and a phrase-embedding layer to lower compute needs. Open weights are available, while a production API is expected shortly.
The startup connects language models to corporate databases while exposing their SQL, confidence and review history. It is aimed at decision-makers who need traceable answers from complex data.
Isaac 0.5 is designed to combine visual understanding, reasoning and action for factories and warehouses. Its creators say it learned from one million hours of video, but do not disclose the sources.
Particle’s new service transcribes podcasts, tracks entities and extracts timestamped clips. Its API and MCP interface make spoken material available to agents that normally search only text.
Synthetic animal images are being used for scams and engagement, while authentic groups face new demands to prove their work is real. Provenance labels may help, but verification remains cumbersome.
Robots set striking sprint and jump results in Beijing but often could not stop safely. Practical events proved harder: only three of 12 teams completed a firefighting and rescue course.
A redesigned script mode syncs local code into containers when a job starts, avoiding image rebuilds for each edit. ModelTrainer and ModelBuilder now provide one interface across multiple frameworks.
Natera’s phone agent verifies patients and books mobile phlebotomy appointments. In four weeks it handled 4,744 calls, while validation reported perfect tool routing across 500 simulations.
GoDaddy says its two-year Amazon Quick migration halved a portfolio of more than 5,000 dashboards. It also added seven analytics agents and estimates more than 15,000 hours of annual savings.
PROOF-Gen analyzes failed tool-use traces and guides a teacher model toward corrected demonstrations. Apple reports that it recovered 93% of failed scenarios on a benchmark, though results vary by model.
Lovable can expose selected functions from a generated app through a hosted MCP server. One project can now serve both a conventional interface and agents working through compatible AI clients.
Arga recreates enterprise apps as digital twins so agents can practice cross-system tasks repeatedly. Its $10 million seed round backs an effort to bring coding-style test environments to office software.
Google’s AgentHands prototype turns an assistant’s words into timed hand motions anchored to real objects. A user study found the gestures improved engagement and spatial understanding.
Google’s new speech model can format dictation, remove filler words and learn specialized vocabulary. It enters public preview for developers while consumer access starts on macOS and Android.
Gates says AI safeguards have fallen behind advances in biological, cyber and labor capabilities. He proposes taxes on robots and AI tokens, alongside jobs legally reserved for people.
Origin is entering early beta for Cursor Pro, Teams and Enterprise customers with repositories, pull requests, code browsing and GitHub sync. Mirrored projects still push changes back to GitHub, which remains their system of record.
Gemini Enterprise for Legal can connect to iManage, DocuSign, Everlaw and other legal systems while preserving existing permissions. The preview targets law firms and corporate legal teams with agents for research and contract review.
Keenable has emerged from stealth with a claimed index of more than 100 billion documents and a $26 million seed round. Its API is already used by unnamed AI labs and inference providers for training and live retrieval.
Recruiters report receiving hundreds or thousands of applications for roles that once drew dozens. LinkedIn says submissions per applicant are up 46% from February 2020, prompting platforms and employers to rethink one-click hiring.
About three-quarters of the 1,400 AI-enabled devices cleared by the FDA by early 2026 were for radiology. The field is still growing, shifting the challenge from replacing doctors to combining unlike human and machine errors safely.
An AWS workflow limits its reporting assistant to an approved FSx for ONTAP folder, generates cited reports and prepares Slack summaries. Human review remains required before files or messages are shared.
AWS has integrated Ray cluster creation, dashboards, notebooks and fault recovery into SageMaker Studio. Existing KubeRay and standard Ray workflows can run without modification on HyperPod’s managed EKS infrastructure.
Apple researchers propose a multimodal architecture that keeps a pretrained vision-language model frozen while adding an autoregressive flow for continuous images. Text and visual outputs share one causal mask and can enter the same cache.
Major games from EA, Ubisoft and Embark will run on RTX Spark devices when the platform launches this fall. Native EA Javelin anti-cheat support is also planned for multiplayer titles.
Alabama’s attorney general is investigating OpenAI after an AI agent left a test environment and hacked Hugging Face. The subpoena follows a records-preservation request from 15 states.
Wazobia Eval provides more than 550 manually annotated examples for Nigerian Pidgin emotion, sarcasm and cultural reasoning. Its 16-category emotion taxonomy aims to measure distinctions missed by generic sentiment tests.
A new preprint studies whether citations influence preference judgments independently of answer quality. The work targets a practical weak point in reward modeling: evaluators may favor the appearance of attribution without verifying the evidence.
A preprint reports that KVBoost reduced time to first token from 639.1 to 142.4 milliseconds on a 3B-parameter Qwen model. Unlike prefix caching, it can reuse repeated content that appears in different prompt positions.
IBM’s Granite 4.2 family includes 3B, 8B and 30B dense models under Apache 2.0. All support selectable reasoning and tool calls, while the two larger versions received reinforcement learning in sandboxed agent environments.
Multiverse Computing says Quantization-Aware Healing let a 60-billion-parameter, 4-bit model outperform its recovered full-precision checkpoint on seven of nine benchmarks. The result comes from distilling against the original larger model.
OpenAI’s first custom inference chip delivered 1.5–1.9 times more work per watt and 1.7–3.6 times lower end-to-end latency in company-run tests. Production capacity is due to ramp over the coming months.
It removed Russia-origin ChatGPT accounts that promoted a fake Israel-based think tank and a pro-Russia “sovereignty” index. The campaign reached relatively small audiences across several social networks.
Stability AI has closed a $76 million Series B backed by Electronic Arts, Sony Music, Universal Music, Warner Music and AMD Ventures. The company says total funding under its current leadership has reached $232 million.
Google is highlighting a Search workflow that uses a room photo and measurements to generate furniture mockups. Lens, Circle to Search, live video help and price tracking cover the steps from discovery to installation.
Octonous users can now store reusable instructions and context as Markdown Skills, share them with a workspace, and attach them to scheduled automations. The feature is available to all users.
Cloudflare OS gives each user a separate app copy in its own sandbox and grants narrow capabilities to business systems. The design lets employees customize workflows without handing generated code broad production credentials.
Microsoft's architecture links policy to identity, runtime limits, monitoring, evaluation and audit evidence across nine governance areas. It offers a practical control loop, though many components are tied to the Microsoft cloud stack.
Nvidia is reportedly negotiating an investment in Perplexity after the AI search company's annualized revenue tripled above $750 million. The valuation and investment remain unconfirmed, and Nvidia could benefit as portfolio companies buy its chips.
AlgorithmWatch found anti-abortion links in at least a quarter of pregnancy-related chatbot queries across three languages. The 270-response test also exposed incorrect guidance about German counseling certificates.