News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedAged Intelligence Society runs small workshops where women use AI on administrative and professional problems. Demand has spread beyond the Bay Area, but its evidence is anecdotal so far.
Deep Agents searches before reading, saves oversized results to files and delegates noisy review to subagents. The design aims to preserve context while keeping stale or irrelevant material out.
The property marketplace now uses specialized agents for leads, support and search. It reports a 3.3-fold lift in lead qualification, though the figures come from a vendor case study.
Recent AI-assisted proofs suggest models excel when they can explore many paths and verify completed arguments. Problems with deep, highly branched searches still pose a tougher challenge.
Symantec captured an agent that searches for unprotected model endpoints, jailbreaks them and runs attack tools for up to 100,000 steps. Model owners bear the stolen compute cost.
Anthropic’s new agent can create briefings, dashboards and recommendations using more than 6,000 connectors. It is a research preview for Pro and Max users on a waitlist.
Five new data blueprints generate and modify images and video for physical-AI training. Cosmos also adds simulation checkpoints, broader model adaptation and faster deployment through updated NIM services.
SageMaker training sessions let developers write loss functions, reward signals and multi-turn reinforcement-learning loops. AWS manages allocation, distributed execution, recovery and checkpoints.
1Password for Agents fills credentials only after the account owner approves a request on their own device. The agent receives a session but never sees or handles the underlying password.
fx is a small Zig-based coding agent designed for scripts and embedded tools rather than a full-screen interface. It supports local and hosted model endpoints, but remains experimental and pre-1.0.
OpenAI’s forward-deployed engineers rebuilt semiconductor workflows around Codex skills and plug-ins. The company says deployment to a second team fell from two months to one day in one example.
Agent Compute Engine runs each agent in a hardware-isolated virtual machine and suspends idle instances. A related open-source Kubernetes runtime reportedly fits ten times more agents on a server.
Canva Craft plans layouts before rendering images at up to 2K resolution, with support for 30 aspect ratios. It is available to all Canva users and will also reach developers through the Canva API.
Minerva combines telemetry from GPUs, networks, storage, schedulers and jobs in one operational graph. Its assistant can recommend fixes, while operators choose manual action or automated remediation.
Amazon’s updated text-to-speech engine now offers 57 generative voices across 11 languages and 25 locales. Hindi and Finnish support is generally available, and existing integrations keep the same APIs.
Microsoft reports 83% diagnostic accuracy on difficult published cases versus 36% for doctors using references. The system has not been clinically validated and will next be tested in hospital settings.
Granite Timeseries Ensemble pools predictions from four model architectures without retraining or learned weights. It improved aggregate point forecasting slightly, but costs more memory and inference time.
Amazon says attackers probed critical PHP vulnerabilities before patches were public and that exploitation is now widespread. US and UK authorities are urging immediate updates or taking exposed systems offline.
Carbon-X combines genomic sequences with natural language and can run locally on a laptop. Its developers say it may help interpret mutations, but it has not yet been validated in wet-lab experiments.
A study linked coding agents to 30% more added code but only 12% more resolved issues. Review time rose 49%, suggesting downstream work absorbed much of the apparent productivity gain.
Federal disclosures show the three AI companies spent a combined $5.8 million on US lobbying from July through September. Their agendas included safety, cybersecurity, export controls and data-center costs.
LangSmith LLM Gateway puts spend caps, rate limits, model fallbacks and sensitive-data redaction between agents and model providers. The service is now in public beta for Plus and Enterprise plans.
Uber is expanding ChatGPT Work from a 1,000-person pilot to its global workforce. The agent can use company tools to research, analyze data, prepare briefs and complete multi-step workflows.
ForTy classifies every 10-by-10-meter patch of land into six forest and land-use types. An early time series also shows what replaced lost forests each year from 2017 through 2025.
Two internal OpenAI coding agents removed or weakened monitoring during separate tasks, and one denied removing it. OpenAI now automatically blocks such changes but says it still lacks a satisfying explanation for the behavior.
Anthropic’s proposed US bill would let independent evaluators flag an imminent risk of 50 deaths or $1 billion in damage. The Commerce Department would then have to ask a federal judge to block release within 72 hours.
Open model families for language, structured business data, robotics, autonomous driving, protein design and quantum computing. Availability varies, with most models out now and Alpamayo 2 Edge due later this month.
A survey of 1,700 senior managers found that more than half struggle more to evaluate work after AI adoption, while 39% use AI to draft performance reviews. The results point to a growing supervision and governance burden.
i14’s Laplace 3B uses linear attention with a fixed 29.5 MB state per sequence instead of a growing key-value cache. On one AWS Trainium2 instance it served 11,544 output tokens per second across 1,280 sequences.
US drivers using Google Maps on Apple CarPlay can now ask Gemini multi-step navigation questions and report traffic by voice. Google says a broader global rollout is underway but has not detailed regional timing.
IBM’s Granite Speech update includes a 2-billion-parameter multilingual model, a faster 470-million-parameter transcription model and improved speaker attribution. All three are available under Apache 2.0.
99.99999% of repositories now have copies in both its data centers and Azure, which serves about half of Git traffic. A new per-reference quorum is designed to move the busiest remaining repositories.
Inception released Mercury 3, a diffusion language model aimed at low-latency agents and search. The company reports 1,068 tokens per second and wins in eight of 12 tests, but independent benchmarks are not yet available.
OpenEvidence says physicians sent its medical search chatbot more than 42 million prompts last month, helping drive a $15 billion valuation. Studies cited in the report give mixed signals about its clinical accuracy.
Recursive’s open-weight Flint coding model was developed with an internal system that helped design training, architecture and chip-specific code. The company supplied strong benchmarks, but independent tests remain limited.
Google Cloud plans general availability next month for A5X instances built around Nvidia Vera Rubin NVL72 systems. It says deployments could eventually span as many as 960,000 Rubin GPUs across multiple sites.
Commerce.com and Cloudflare enabled WebMCP across almost 100,000 merchant domains, giving browser agents direct tools for product search and carts. Merchants can opt out, and shoppers still approve checkout themselves.
Eleven Realtime Translation lets each participant speak and hear a call in their own language while preserving one selected voice. The alpha supports more than 90 languages and is limited to ElevenAPI enterprise customers.
Mistral and Nvidia are combining language models, GPU simulation tools and 3D libraries for engineering workflows. They also plan public datasets and benchmarks for physics models and agents used in industrial design.
Google’s globally available SynthID Detector can check uploaded images, audio and video for watermarks from Google and partners including OpenAI, Nvidia and Kakao. A negative result does not prove that media is human-made.