AI.info

The Pulse — Page 47

Browse The Pulse on AI.info.

• SB Intuitions Explores Visual Priors for Robot Action Learning

  SB Intuitions is exploring whether visual priors and visual prediction can improve action learning, with robotic manipulation as a downstream testbed. Its Latent-WAM experiments examine attention designs, visual memory and learning from interaction.

• Planet’s New AI App Lets Users Search Ten Years of Imagery

  Planet’s new AI app lets users search ten years of Planet imagery using everyday language. The map-based app is currently free to access, according to the company.

• OpenAI’s GPT-6 Astra Scores Face Scrutiny After Quiet Revisions

  OpenAI changed several GPT-6 Astra benchmark figures after its September 3 launch, including hallucination, cybersecurity and rival-model scores. Independent ARC Prize testing also showed a wide gap between Astra’s 99.9% result under OpenAI’s provider adapter and 62.7% under the standard harness.

• Germany Plans Hundreds of Millions for Nvidia AI Pact

  Germany is preparing a state-backed industrial AI partnership with Nvidia, with federal funding in the hundreds of millions of euros. The plan would support industrial AI models and physical AI for robotics and autonomous vehicles, using computing capacity from Deutsche Telekom’s Industrial AI Cloud.

• UK announces £90m AI-enabled undersea surveillance investment

  The UK has announced a £90 million investment in drones, sensors and AI systems for the Royal Navy’s Atlantic Net programme. The Ministry of Defence says a framework with Ocean Infinity Solutions will support almost 200 skilled jobs and strengthen underwater surveillance.

• OpenAI Reports Six Model-Misalignment Incidents and Sets Disclosure Rules

  OpenAI has published six reports on unexpected or concerning model behavior alongside a framework for tracking, investigating and disclosing model misalignment.

• Unique Introduces Remy, an Overnight AI Teammate for Advisers

  Unique AI introduced Remy, a wealth-management assistant that reviews client information overnight and prepares advisers’ work for the next morning. The system drafts meeting briefs, client letters and compliance materials, but leaves decisions and messages for human approval.

• Hugging Face Brings llama.cpp Quants to Transformers

  Hugging Face has added support for running llama.cpp quantized models through Transformers, allowing developers to load GGUF checkpoints with familiar Python and PyTorch APIs.

• Apex Intelligence Raises Nearly $50M for Self-Evolving Models

  The Beijing startup will use the financing to develop foundation models designed to improve their own research processes, while its claims about benchmarks and mathematical work remain unverified outside company materials.

• SpaceXAI Releases Grok Voice Transcribe 2.0 at the Same Price

  Grok Voice Transcribe 2.0 doubles the accuracy of its predecessor in SpaceXAI’s real-world evaluations while keeping batch and streaming prices unchanged.

• Nature Warns AI Reasoning Traces Are Weak Safety Evidence

  A Nature Machine Intelligence editorial cautions that written reasoning traces are not reliable explanations of a model’s internal computations. OpenAI’s GPT-6 Astra system card reports that a chain-of-thought-only monitor missed most prompted sandbagging cases in one test, while a separate full-context monitor caught successful honeypot attacks in its evaluation.

• Google Tests Flipkart Checkout in Gemini and AI Mode, Report Says

  Google is testing a Buy button for select Flipkart listings in Gemini and AI Mode in India, TechCrunch reported, citing people familiar with the matter and a test it saw. Google said it is always testing features but gave no details on Flipkart.

• Google Researchers Add Guardrails to Agent Self-Improvement

  A paper released by Google Cloud AI Research collaborators on September 21 describes a method for limiting how AI agents revise their operating instructions and how those revisions are selected. Its experiments report that the method improved held-out benchmark scores while using fewer policy tokens than unregularized evolution.

• Anthropic Leases 95 MW at EverWind’s Planned $10B Nova Scotia Campus

  Anthropic has signed a 20-year lease for the first 95 megawatts of EverWind’s planned data-centre campus at Point Tupper, Nova Scotia. EverWind says the wider campus and a portfolio of wind projects represent more than $10 billion in planned investment.

• CrowdStrike Adds Agent-Insider Detection to Its Agentic SOC

  CrowdStrike’s Falcon platform can connect an AI agent’s activity to its owner and device, then investigate when those signals point to compromise, the company says. The company also describes approval controls and response tools for managing agents across security operations.

• Property Finder Finds 96.8% of Its AI Labels Came From the Old Model

  Property Finder found that 96.8% of labels in its property-image evaluation set came from its incumbent vision model, biasing comparisons with alternatives. After independent annotation of 356 UAE images, the company selected Llama 4 Maverick on Amazon Bedrock, though AWS says the reported scores are optimistic because the test set includes images used for prompt tuning.

• HubSpot Cuts Nearly 660 Jobs in AI Strategy Reorganization

  HubSpot will eliminate nearly 660 jobs, about 7% of its workforce, while reorganizing product teams around customer outcomes. CEO Yamini Rangan says the cuts align the company with its AI strategy, not AI-driven productivity gains.

• David Robinson Resigns From OpenAI, Faulting Its Safety Culture

  David Robinson, who led transparency work on OpenAI’s safety team, says he resigned after three and a half years at the company. He argues that its sprint-driven culture and reliance on fixing problems after deployment leave gaps in safeguards for increasingly capable AI systems.

• Researchers present infrastructure for comparing world models

  A November 7 preprint introduces WorldFoundry, open-source infrastructure for running and evaluating generative systems across several modalities. Its catalog lists 270 model entries, 74 benchmarks and 62 registered metric entries, while tracking execution evidence separately from readiness.

• Microsoft Gives Copilot Permission to Act on Windows PCs

  Copilot+ PCs will let Copilot use local file context, take actions across Windows and run some AI models on the device, Microsoft says. The features are expected to begin rolling out in the coming months, while new taskbar Search actions start with Windows Insiders.

• FAA Sets September 21 AI Trial for Washington Airspace

  The FAA is preparing to deploy SMART, an AI-assisted air-traffic planning system, around Washington, D.C., under a 12-year, $875 million contract with Air Space Intelligence. The platform will analyze schedules, weather, airport capacity and airspace conditions while controllers retain responsibility for aircraft separation.

• DeepSeek V4.1 Flash Passes 1 Trillion Tokens on OpenRouter

  DeepSeek V4.1 Flash processed more than 1 trillion tokens on OpenRouter shortly after its September 10 launch. The open-weight model combines a 552-billion-parameter mixture-of-experts backbone, native vision and a one-million-token context window with sharply lower cached-input pricing.

• Treble Raises $18M to Bring Acoustic Simulation to Physical AI

  Treble Technologies has raised $18 million to expand its acoustic simulation and synthetic audio platform for physical AI applications, including robotics, wearables and autonomous vehicles.

• ServiceNow’s AutoSynthData Turns Agent Failures Into Training Tasks

  ServiceNow’s AutoSynthData pipeline uses agent failures and a stronger teacher model to generate and verify enterprise training tasks. In tests on EnterpriseOps Gym, fine-tuning improved mean Pass@1 by 7.2 percentage points in Hybrid and from 18.77% to 27.18% in ITSM.

• Apple Researchers Cut Multimodal Agents’ Reasoning Latency by 93%

  Apple researchers report that a method for selecting prepared reasoning steps reduced per-turn reasoning latency by more than 90% in tests of multimodal search agents. The paper reports results from research benchmarks, not a feature announcement for Apple products.

• UK CMA Opens Competition Probe Into Meta’s Business Terms

  The UK Competition and Markets Authority opened an investigation into whether Meta’s business terms unfairly let it use publishers’ website content, including for AI features. Meta has paused the planned changes while it consults industry groups; the CMA has not decided whether the terms break competition law.

• InclusionAI Posts Ling-3.1-Flash-1019 With 560B Parameters

  InclusionAI has posted Ling-3.1-Flash-1019, a 560-billion-parameter model with 25 billion active parameters per token and a one-million-token context window. Its Hugging Face release lists an MIT license and reports gains on several reasoning, agent and science benchmarks.

• Cohere Opens Confidential Inference in Model Vault Beta

  Cohere has opened a public beta of confidential computing for its Model Vault inference service, running on NVIDIA H100 GPUs on Microsoft Azure. Customer-side software verifies the processing environment before sending protected requests, while end-to-end encryption shields request contents from intermediaries.

• QwenGyre Cuts a 48-Step Agent Training Run to 75 Hours

  Alibaba researchers introduce QwenGyre, a reinforcement-learning framework that shifts computing capacity between agent rollouts and training. On NL2RepoBench, the team reports a 75.42-hour run and a pass-rate increase from 52.48% to 58.54% for Qwen 3.8 2.4T.

• Kroger Plans 3,000 Corporate Job Cuts Over Three Years

  Kroger plans to eliminate about 3,000 corporate jobs over the next three years, according to a memo from CEO Greg Foran. The memo cites technology and duplicated work, but does not explicitly connect the cuts to the company’s separate AI initiatives.

• Anthropic Opens Biology Access to Vetted Research Teams

  Anthropic is opening applications for its Life Sciences Verification Program, which gives vetted teams access to Claude models with fewer biology-related restrictions. The beta covers drug discovery, research biology, clinical development and manufacturing, with additional review for higher-risk projects.

• Navan Puts NVIDIA’s Nemotron 3.5 Model Into Travel Agents

  NVIDIA’s Nemotron 3.5 Super VL now powers several of Navan’s AI travel agents in production. The company fine-tuned the model in three weeks and says it beat or tied competing models on six of seven internal measures.

• Senators Say Data Centers Shift Grid Costs to Households

  A Senate report says AI data center operators oppose paying the full cost of grid upgrades their facilities trigger, while seeking tax breaks and nondisclosure agreements. The report examines responses from Amazon, Google, Meta, Microsoft, CoreWeave, Digital Realty and Equinix.

• LatticeFlow Finds Stronger Political Alignment in Larger Qwen Models

  LatticeFlow AI’s political-bias benchmark compares Qwen 3.7 Max with Qwen3 32B, finding the larger model closer to the Chinese pole across six China-politics categories.

• Microsoft Prices $25 Billion Bond Sale as AI Capital Spending Rises

  Microsoft priced $25 billion of senior unsecured notes, with the offering expected to close November 4. The sale comes as the company forecasts approximately $170 billion in calendar-year capital spending and says it may seek additional funding when attractive.

• Hume Ships Audio Decision API for Direct Speech Analysis

  Hume’s Audio Decision API lets developers ask questions about speech and sound and receive probabilities for answers they define. Evaluation and benchmark results are self-reported, while access is available by waitlist.

• Rivercell Raises $25 Million to Build an AI Model of Cell Responses

  Paris-based Rivercell has raised a $25 million seed round led by HV to expand its biological data platform and laboratory. The company aims to train an AI model to predict how human cells respond to drugs and genetic changes, but says the data needed to build it do not yet exist at the required scale.

• The RealReal Tightens Bot Defenses After Muse Cart Test

  The RealReal has made it harder for AI agents to log in, reserve items and check out after a Meta Muse test repeatedly renewed 20-minute holds on ten listings. The retailer still lets agents surface listings and direct shoppers to product pages.

• HHS Launches SURPASS to Redesign Clinical Trials With AI

  HHS announced SURPASS, an ARPA-H program intended to use computational models, shared infrastructure and real-time analysis to change how clinical trials are designed and run. ARPA-H lists October 12, 2026, as the webinar registration deadline and October 28, 2026, for Proposers’ Day registration.

• White House Calls for Immediate AI Incident Reporting

  A White House task force said AI companies must immediately disclose incidents involving their models and take action to remedy harm. Its statement did not explain how companies should report incidents or what consequences would follow if they failed to comply.