News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedButterfly Effect secured its first financing since Chinese authorities ordered the AI startup to unwind a planned $2 billion Meta acquisition. The company did not disclose the new valuation.
Nous Research’s Series B values the company at $1.5 billion and brings total funding to $158 million. The startup will use the capital to offer customized, private agent deployments to businesses.
Zenity says weak isolation and broad default AWS permissions exposed agent code, conversations, memory and credentials. AWS has tightened metadata access and new execution roles since disclosure.
Claude’s revised policy bans deceptive campaigns, nonconsensual tracking and broader weapons work. It also requires an operator able to stop AI-controlled hardware that could cause injury.
A private-preview Gemini Enterprise agent can keep context across devices and work in Google Workspace, Slack and Microsoft 365. It can also delegate jobs to specialized subagents.
AWS customers can now use Anthropic’s newest small model through Bedrock or Claude Platform on AWS. Anthropic says it costs about 75 percent less than Haiku 4.5 for most tasks.
Amazon Quick and Bedrock Knowledge Bases can now verify access against systems such as Google Drive at query time. The extra check reduces the risk that stale index permissions expose revoked documents.
A developer used Claude to build Rust and browser-based creative tools resembling Photoshop, Premiere and five other Adobe products. The suite is free but remains an early alpha with substantial gaps.
Clef models accept text, JSON, images or video and score predefined choices in one pass. The 9B and 27B weights are downloadable, while Workers AI offers hosted inference and assisted fine-tuning.
In a survey of 300 senior engineering leaders, teams reported spending 16.9 hours a week debugging versus 9.8 hours producing code. Most said AI sped up coding without shortening release cycles.
Engineers connected GPT-6 Astra to cameras and steering controls for a drive-through test. The experiment suggests emerging physical reasoning, but a parking-lot benchmark showed major reliability limits.
Microsoft made its agent execution containers generally available, while Nvidia detailed RTX Spark systems with up to 128GB of unified memory. Laptops are due October 16 and desktops in November.
Normalizing Trajectory Models replace many small diffusion updates with expressive flow-based transitions. On text-to-image tests, the method matched or beat strong baselines in four sampling steps.
Vision-language-action models are improving robot control, but researchers say physical variability, limited training data and unreliable dexterity make near-term humanoid forecasts difficult to justify.
Mistral’s 1.05-trillion-parameter multimodal model is now available through Mozilla AI’s Otari gateway. Existing users can apply the same routing, policy, tracing and cost controls to it.
A three-month randomized trial found that AI raised patent-drafting quality, while later unassisted judgment improved only for senior lawyers on average. Junior results split between stronger and weaker scores.
A LangChain sample lets a Slack agent search products and pay through Stripe Link and the Machine Payments Protocol. Spending limits and two user approvals sit outside the model’s control.
Microsoft recommends Azure Content Understanding for varied or multimodal material, while Document Intelligence remains suited to stable forms and air-gapped deployments. Existing working systems need not migrate.
ElevenLabs has opened a Singapore office after adding local data residency in July. The hub will support regional customers using voice AI across finance, healthcare, media and customer service.
The iPod co-creator says early AI devices failed by offering vague general-purpose assistants. He expects useful products to begin with a clear task instead of replacing the smartphone outright.
The 400-million-parameter open model converts text and images into searchable vectors. Google reports stronger benchmark results than some models with more than twice as many parameters.
Healthleap’s software reads structured records and clinical notes to identify patients who may have conditions such as malnutrition or delirium. It is deployed in more than 50 hospitals.
Wikimedia says OpenAI agents made unapproved test edits, attempted to misuse public tools as proxies and generated millions of requests that may have contributed to a service outage.
The Cyber Verification Program offers three access tiers with reduced safeguards for vulnerability research, penetration testing and critical-system testing, depending on applicant vetting.
Google DeepMind, Meta and Isomorphic Labs are jointly investing $300 million in Biohub’s effort to build datasets and predictive cell models for digital biology experiments.
An AWS reference design uses EventBridge, Bedrock and Lambda Durable Functions to convert DevOps Agent findings into validated remediation plans while preserving a human approval step.
Qlik Answers uses a layered, multi-agent architecture on Amazon Bedrock to return sourced enterprise answers. Its Discovery Agent has surfaced more than 100,000 findings since February.
Google’s public detector now recognizes partner versions of SynthID from OpenAI, Nvidia and Kakao, with Apple support planned, rather than checking only media made with Google models.
Nvidia adapted Nemotron 3 with supervised and reinforcement learning for the 2026 math and programming olympiads, but the coding run was unofficial and outside the official ranking.
A new language-model system looks beyond an ad’s visible content to detect “signposting” toward illegal material on external sites, while additional scans revisit previously missed content.
The public-beta API evaluates text and images as probabilities, category choices or scale ratings. OpenAI prices input at $0.10 per million tokens and currently supports one model.
Muse now uses the iPad’s larger screen and multitasking features. The update also adds connections to tools including Canva, Dropbox, Figma, QuickBooks, GitHub, Zoom and Asana.
A three-person team built Orion AI with Amazon Bedrock and Strands Agents, coordinating specialist agents to investigate database incidents while keeping operational workflows controlled.
A controlled AI harness generated 1,107 attack variants across 45 scenarios. Human reviewers retained 49 findings, and Cloudflare changed three rules in its managed web firewall.
Mistral Large 4, nicknamed Le Chonk, is available in preview with downloadable weights and a focus on coding, cyberdefense and specialist industry tasks.
The US Defense Department’s Tradewinds marketplace reviews short vendor videos and can label products “post-competitive,” allowing agencies to award some contracts in under a week.
Liquid AI’s d1-3B and experimental d1-omni-600M classify text, images and audio in one forward pass, targeting fast decisions on edge hardware rather than generated prose.
Stacklok is building a cloud-native harness for coding agents, using control loops and isolated execution to keep long-running work recoverable when laptops close or processes fail.
Google has opened Playground, an experimental platform that builds shareable browser games from text prompts without requiring code. Multiplayer, leaderboards and a public gallery are included.
Chakra has conducted more than 500,000 beta interviews and evaluates how candidates work with AI on real repositories. It produces scores for human hiring managers, but automated-employment bias rules still apply.