News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedProject HydraFusion chooses whether one model, an escalation path or a critique-and-revision loop should handle each coding request. GitHub says offline tests improved quality while lowering estimated cost.
Shopify’s “gisting” method replaces long system prompts with trained token embeddings. In its Sidekick agent, latency fell and throughput rose without a reported loss in prediction quality.
Azure DevOps teams can now run Copilot reviews without moving repositories to GitHub. Reviews are billed through Azure, but costs appear 48 hours later and budgets only send alerts rather than stopping usage.
GPT-6 Astra coordinated model training, debugging and fleets of subagents during a 20-billion-token test. Its estimated single-agent cost was low, but heavy parallel use can quickly raise the bill.
Ukraine has given more than 100 companies and the UK government access to combat-drone records. The rare edge cases could improve autonomous systems, but the trade is developing faster than rules for consent and reuse.
The text-based assistant has completed a SOC 2 audit and says it neither sells user information nor uses it to train models. Its privacy claim still depends on the company’s controls and policies.
OpenAI’s chief called some planned compute expansion “unsustainable silliness,” arguing that certain neocloud projects lack customers or revenue. He says OpenAI’s own capacity is demand-backed.
The open-source tool indexes sessions from Codex, Claude Code, pi, and Hermes so agents can retrieve earlier decisions with provenance. Processing stays local by default.
H has released 260-million- and 800-million-parameter encoders for searching documents by text and images. The smaller model processed 51 pages per second in the team’s test.
OpenAI ended its Cursor relationship after SpaceX acquired the coding startup, citing concerns about Elon Musk’s companies and contract compliance. Cursor had been one of OpenAI’s largest customers.
A new guide runs LiteLLM on ECS between Codex and an OpenAI model in Amazon Bedrock. The gateway centralizes identities, budgets, rate limits, routing, and usage telemetry.
AWS has documented an Outlook integration that lets Amazon Quick summarize threads, draft replies, schedule meetings, and trigger workflows. Administrators must configure Microsoft and AWS permissions.
New reference implementations show Bedrock AgentCore generating database diagrams and coordinating a multi-agent security review. Both keep people responsible for critical decisions.
Three major AI assistants developed problems on Thursday, disrupting chat, APIs, coding tools, and other features. Providers reported separate infrastructure and capacity symptoms, with no confirmed common cause.
Google says its updated AI weather model produces a global forecast every hour at five times the spatial resolution of its predecessor. The largest claimed gains are in rain and snow prediction.
New Gemini-powered modes let users search email, shape documents, and dictate notes through spoken conversations. Rollout starts with Google AI Pro and Ultra subscribers in the US.
Google Research and HHMI Janelia published the largest brain wiring map to date, covering the male fruit fly’s brain and central nervous system. The dataset supports comparisons with the existing female connectome.
Nvidia’s PAIR software will distribute inference across compatible computers on a local network. The company also announced faster local runtimes and compact RTX Spark Windows PCs arriving in October.
Nvidia’s largest acquisition would give the chipmaker control of a central hub for open AI models and datasets. The companies say Hugging Face will remain open and continue supporting multiple hardware platforms.
Copilot now compresses repetitive command output and strips unused line prefixes while preserving source-like results. GitHub says one formatting change lowered daily inference cost per user by about 3%.
Developers can run several Copilot sessions against one project without mixing their changes. Each agent receives a separate Git worktree and keeps its own task context.
Its new Turbo tier processed pages four times faster than its Cost Effective mode, with a 3.7-second median and 0.84 F1 score on ExtractBench. It is available now in beta.
Schneider Electric, Vodafone, and monday.com found that agent prototypes are easier than reliable operations. Their production systems emphasize shared platforms, evaluation, and bounded tools.
A Nevermined integration gives LangChain agents spending-limited payment cards for buying services or refilling credits mid-task. LangSmith records the resulting payments.
Cohere’s new dataset covers about 696,000 public AI tools, yet only 2.6% could perform a recognized work task under a strict test. The findings show uneven automation across jobs.
The startup says its recurring revenue grew more than tenfold as companies sought protection for models, agents and their tools. Its new funding will support product expansion, sales and growth in Europe.
Atlas combines view generation, 3D reconstruction and simulation in one model. It can export spatial data and create sensor views for robot training, but access is currently limited to selected partners.
Researchers model scams as multi-turn interactions whose danger grows from casual contact to urgent payment requests. Their approach updates a risk estimate as each new message or call segment arrives.
GUI-CC tests whether generated interfaces remain coherent when agents interact over multiple steps. It addresses a gap in benchmarks that score only the next predicted screen in isolation.
Developers can now mark OAuth scopes as optional, so users need not grant every permission an agent might theoretically use. Apps must inspect the scopes actually issued and disable unavailable actions.
Ai2’s auditing method analyzes individual benchmark questions instead of trusting one aggregate score. Across 34,000-plus prompts, it recovered separate safety and reasoning dimensions without labels.
The Israeli-Dutch startup more than doubled its valuation in about six months. It is expanding from customer-service agents into a platform that connects models and workflows to company data.
Google’s latest budget model posts stronger agentic-coding results at the same introductory token price. Independent testing suggests it uses more output tokens, making tasks costlier than on 3.7 Flash.
As employers use voice agents to screen applicants, candidates are answering with AI assistants. One experiment produced a polite bot-to-bot interview but no human follow-up, exposing an accountability gap.
AWS customers can now use Claude Fable 5.1 for coding, research and long-running agent work. Standard access retains prompts and outputs for review, while eligible safeguards users can choose zero retention.
OpenAI says Astra will use added chain-of-thought monitoring, but researchers fear its reported architecture could hide more internal reasoning. The model’s release was delayed for additional safety work.
The research framework checks concrete agent actions before they change shared systems. It aims to apply consistent authorization and provenance rules across different models, tools and runtimes.
Trajectory-judge tests show why a correct final answer may hide broken or unsafe intermediate actions. The benchmark uses a deterministic support environment where each injected fault is known.
Swiggy’s in-house model ranks the likely long-term value of new users for ad bidding. Adding order count as a second task cut parameters by 63% while slightly improving reported performance.
Google’s Fairwind program combines Gemini 3.8 Flash Cyber with CodeMender to find and patch software flaws. Initial access is limited to selected governments, critical infrastructure and partners.