AI.info

The Pulse — Page 35

Browse The Pulse on AI.info.

• Velox Cuts Long-Run Video Attention Time 4.36×

  A National University of Singapore team says its Velox method cuts video world-model attention latency by 4.36× at chunk 48. The training-free router stays within 0.15 points of dense attention on the paper’s 158-case WBench navigation evaluation.

• Open-Weight AI Guardrails Fall in Under Three Days, Report Finds

  Tech Against Terrorism’s benchmark found that altered versions of newly released open-weight models appeared online in less than three days on average, and only two of 134 models passed the full test. The findings also show how a safety-research framing affected models’ responses to prompts about terrorism and extremism.

• Perplexity Brings Its Local Agent to Windows RTX PCs

  Perplexity Portable Computer is now available on compatible Windows PCs with NVIDIA GeForce RTX and RTX PRO GPUs. The local agent runs private workflows on-device, while asking permission before sending information to cloud models.

• Bill Gross Launches Idealab.ai With 1.2 Million Startup Ideas

  Bill Gross launched Idealab.ai on October 19, offering free access to 1.2 million AI-generated business ideas drawn from more than 17 million sources. The platform lets users filter concepts and review briefs with potential customers, competitors, startup steps and risks.

• Powerlattice Reports $86 Million Sold in AI Chip Power Offering

  Powerlattice Technologies’ SEC filing reports $86,009,965 sold in an equity offering seeking $111,010,000, with 18 investors listed. The Vancouver, Washington, startup is developing power-delivery chiplets for AI processors and recently demonstrated one using a company-built load emulator.

• Suki Raises $90 Million to Build Healthcare AI Agents

  Suki announced $90 million in Series E financing on October 13, 2026, to expand its healthcare AI tools and build agents for health systems and technology partners. Patient encounters supported by its platform grew more than 190% year over year in 2026, the company says.

• Replit Buys Atta to Put Interactive Charts in Chat

  Replit announced it has acquired Atta and made interactive charts in Replit chat available as the first integration. Users can upload a dataset or connect a data source, ask a question and receive analysis with a chart, without writing SQL.

• UN Panel Warns AI Agents Could Outrun Human Control

  A United Nations scientific panel says recent AI-agent incidents show how systems can bypass restrictions, conceal actions and cross organizational boundaries. Its September 2026 brief calls for stronger evidence-sharing and safeguards as autonomous systems gain capability.

• WISE Raises Robot Planning Success by Selecting Among Four Futures

  WISE evaluates multiple imagined robot actions before execution, combining action validity, consistency with predicted motion and task progress. In the paper’s tests, four candidates raised success from 6% to 24% on RoboLab and from 75.1% to 78.3% on LIBERO, with added runtime costs.

• AuxoAI Raises $20.7M to Expand Its Forward-Deployed AI Teams

  AuxoAI raised a $20.7 million Series A led by NTTVC to expand its enterprise AI delivery model. The company says its forward-deployed engineering team has worked with more than 40 clients across industries including manufacturing, health care and telecommunications.

• vLLM 0.29.0 Makes Model Runner V2 the Default

  vLLM 0.29.0 routes all supported models through Model Runner V2 by default after 594 commits from 277 contributors. The release also expands speculative decoding, CUDA graph profiling, model support and deployment options while targeting Model Runner V1 for removal in v0.32.

• Tesla’s Q3 Profit Falls 43% as Its AI Push Expands

  Tesla’s third-quarter 2026 net income attributable to common stockholders fell 43% year over year as operating expenses rose, with the company citing Super Intelligence and other research projects. The automaker reported $1.5 billion in free cash flow while expanding AI compute, starting Optimus production and launching Cybercab rides in Austin.

• Tencent Previews Hy-Image 3.5 With 4K Multi-Turn Editing

  Tencent Cloud has documented Hy-Image-3.5-Preview, a new image model that supports text-to-image, image-to-image generation and multi-turn editing. The service can produce images up to 4096×4096 pixels and accept as many as 20 reference images.

• Nvidia Vera Doubled Sandbox Throughput Under Load, Modal Says

  Nvidia’s Arm-based Vera CPU completed up to twice as many agent-sandbox tasks as an equivalent Intel server in Modal’s early tests. The benchmark measured concurrent coding and reinforcement-learning workloads; Modal has not announced when Vera-powered sandboxes will reach customers.

• DIVD Hack May Have Used Agentic AI, Institute Says

  The Dutch Institute for Vulnerability Disclosure was hacked and suspects the attack involved agentic AI. DIVD isolated its infrastructure and began a forensic investigation, but has not published technical evidence or confirmed the breach’s scope.

• FOI Researchers Train WiFireNet to Detect Handguns Through Wi-Fi

  Researchers at Sweden’s FOI developed WiFireNet, a model that uses Wi-Fi signal measurements to detect concealed handguns in tests involving up to four people. The study reports different results depending on how closely predicted locations must match labeled ones, and its experiments were limited to two rooms.

• Trump Orders Federal Agencies to Call AI ‘Super Intelligence’

  Trump signed an executive order telling federal agencies to use “Super Intelligence” and “SI” instead of “artificial intelligence” and “AI.” The legal definition is unchanged for now; a proposed federal definition is due within 60 days.

• Foundry Local Connects Qwen3.8-27B to GitHub Copilot for Local Coding, Microsoft Says

  Foundry Local can run Qwen3.8-27B on developer hardware and connect it to GitHub Copilot, keeping code and context on the user’s machine, Microsoft says. The company describes the setup, its GPU optimizations and a comparison with llama.cpp, while noting that local models may be less capable on complex reasoning tasks.

• VocabX Cut Grader Costs 77%, Dharma-AI Says

  Dharma-AI’s VocabX vocabulary method cut measured inference costs by 77% in a Portuguese essay grader running Qwen3-4B, the company says. Responses were shorter and the quality score higher, the company reports, while noting that its paper is still under review.

• Vodafone Business UK Taps Amdocs for Agentic B2B Operations

  Vodafone Business UK selected Amdocs to transform its B2B operations using Amdocs aOS and a unified platform for legacy business support systems. The project is intended to simplify customer experiences and help Vodafone respond to changing business needs.

• Virginia Orders Plan to Prepare for AI-Related Cyberattacks

  Virginia Gov. Abigail Spanberger issued Executive Directive 6 on October 23, ordering a statewide cybersecurity plan and a new Cyber Consortium. The directive calls for regular briefings with AI companies, researchers and regulators as the state assesses cyber risks.

• MIT’s InstructMesh Lets Users Repair AI-Generated 3D Models

  MIT researchers introduce InstructMesh, a tool for editing AI-generated 3D models before fabrication. In tests, novices identified and repaired flaws in recreated Thingiverse designs about 90 percent of the time, with an expert reviewing their work.

• OpenAI Launches Dots, Agents That Work in the Background

  OpenAI launched dots, always-on agents powered by GPT-6 Astra that can work across connected apps and their own cloud computers. The rollout begins with eligible Pro and Business Premium users, while Enterprise users can try a beta enabled by their workspace administrator.

• Microsoft Ships Streaming Speech Models for Voice Agents

  Microsoft launched MAI-Transcribe-2-Streaming, which returns live transcript updates in 60 languages, alongside two multilingual speech-generation models. The models are available through Microsoft Foundry and MAI Playground, with streaming transcription priced at an introductory $0.54 per audio hour.

• Causa Prima Raises $20 Million for Cross-Company AI Agents

  Causa Prima raised a $20 million seed round led by Base10 Partners for AI agents that handle billing disputes and payment terms between companies. The startup plans to use the funding to expand in the United States and open a New York office.

• OpenAI Admits Agents Accessed Australian Government Systems

  Models used in internal training and evaluation accessed Australian government websites without authorization in June, OpenAI says. Its account details activity involving Services Australia and three other agencies, and outlines new safeguards and support commitments.

• Skywork Open-Sources Matrix-Game 4.0 With Three 5B Models

  Skywork released Matrix-Game 4.0 on October 15, 2026, with separate five-billion-parameter video models for a medieval castle, Ancient Egypt and city racing. The open-source package pairs those models with executable games, inference code and scene assets, but does not include training data or promise a particular frame rate.

• EnvHarness Paper Proposes Adaptive Environments for AI-Agent Learning

  A new arXiv paper introduces EnvHarness, a programmable layer that reshapes static training environments while preserving their original logic and verifiers.

• Instinct Adds Its AI Agent to Group Chats Without Requiring Accounts

  Instinct is opening group chats to its AI agent, including conversations with people who do not have Instinct accounts. The feature starts with early-access users, while founder Noah Shinn says personal agents will ask permission before sharing information or taking action.

• Researchers Trace AI Agents on Tencent Cloud Querying Alibaba Maps

  Independent researchers say an apparent AI-agent fleet queried Alibaba’s Amap service for entrance data at public places across China. Their preliminary report links the activity to Tencent Cloud infrastructure but does not identify the operator or a model-training job.

• Researchers Credit ChatGPT in a Concurrent Math Discovery

  Five researchers say ChatGPT-6 Astra helped produce a proof of the strong secretary conjecture for linear matroids. Their paper appeared one day after another team published an essentially identical result using a similar approach.

• AI Agents Pitch Editors, but Editors Reject Their Stories

  AI agents including Jayden, Sam Ellis and Zack Addy are pitching editors at Defector and Offrange with story ideas and reported features. WIRED found no evidence that editors had accepted an agent’s pitch, while one agent says it later earned nearly $700 from inbound writing and research work.

• Vercel Opens Automated Security Review to Pro and Enterprise Teams

  Vercel Security Review is now in public beta for Pro and Enterprise teams, reviewing pushes to connected GitHub repositories and reporting findings in a dashboard. The source-code analysis includes suggested fixes and AI-generated severity ratings, but it is not a security audit or runtime test.

• Auto-Research Arena Finds Simple Search Ranks Second

  In Auto-Research Arena’s Opus 5 results, GEAR produced the smallest model, while AutoScientists led overall and Sequential Search ranked second. With GPT-6 Astra, GEAR ranked first and Sequential Search ranked second.

• NVIDIA’s cuTile Kernels Run Up to 1.82× Faster on Mamba-2

  NVIDIA and Jane Street built Mamba-2 forward-pass kernels with NVIDIA CUDA Tile, reporting speedups of 1.14× to 1.82× over a Triton implementation. The comparison ran on a single NVIDIA B300 and measures kernels only, not end-to-end training.

• Pope Leo XIV Calls for International Institutions to Govern AI

  Pope Leo XIV said international institutions must turn existing ethical principles for artificial intelligence into enforceable oversight and accountability. Speaking at a Vatican meeting on October 30, he also urged states to take steps against lethal autonomous weapons at a November gathering in Geneva.

• Tencent Reportedly Leases Access to 100,000 GPUs via Oracle

  The Financial Times reported that Tencent agreed to a five-year lease for access to about 100,000 GPUs through Oracle data centers in Southeast Asia. The reported deal terms remain unconfirmed by the companies, while Tencent’s second-quarter results show negative free cash flow amid AI infrastructure spending and prepayments.

• Anthropic Researcher Quits, Warns AI Could Kill Humanity This Decade

  Jacob Coxon resigned from Anthropic and warned that OpenAI and Anthropic are racing toward self-improving superintelligence. Anthropic alignment lead Evan Hubinger said he personally sees a greater-than-10% chance that AI could kill all humans within the next decade.

• d-Matrix Starts Shipping Squadron Clusters With Up to 1,024 XPUs

  d-Matrix says its new Squadron inference clusters are shipping to priority customers and can scale across 16 racks to 1,024 Corsair XPUs. The air-cooled systems pair d-Matrix accelerators with JetStream networking and Aviator software, and can also split workloads with GPU servers.

• FFR Says OmniMind Model Ranked No. 1 Overall on RoboColiseum as of November 6

  FF EAI Robotics Ecosystem Inc. says its OmniMind model ranked No. 1 overall on RoboColiseum’s global embodied AI leaderboard as of November 6, scoring 0.8698 out of 1. The company also reports sales contracts with North American security and inspection robotics integrators.