AI.info

The Pulse — Page 8

Browse The Pulse on AI.info.

• Ollie Will Shut Down Nov. 6 After Failing to Keep Pace

  Ollie will close its family-focused AI assistant on November 6, a week after co-founder and CEO Bill Lennon announced the shutdown. The company says it will text users a data export and refund partial or unused subscription months.

• Apate Sends AI Agents to Keep Scammers Talking

  Apate uses AI agents to engage scammers across calls, messages, email and social platforms, seeking intelligence that could help organizations respond to fraud. The startup says its agents have occupied scammers for 17 years across 2.5 million conversations and is expanding with new seed funding.

• Flock Offers Buyouts as Cities Cut License-Plate Camera Deals

  Flock Safety opened a voluntary severance program as dozens of US cities and counties end or pause contracts for its AI-powered license-plate reader network. The Atlanta-based company expects most applicants to receive buyouts while facing privacy criticism, camera vandalism and weaker customer demand.

• Kaizen Wins Army Software Deal With a $49 Million Ceiling

  The Army gives Kaizen a five-year contract vehicle worth up to $49 million for mission and enterprise applications, staff workflows, records and reporting. Its first task order covers classified workflows and is valued at $43 million over five years, according to DefenseScoop.

• OpenCode Says Breach Affected 1.4% of Console Users

  OpenCode said a security breach affected 1.4% of Console users, a subset who had linked external providers, and that it had notified those users by email. Its public statement did not say provider API keys were exposed or explain the incident’s technical details.

• Cohere Makes Embed 5 and Rerank 5 Available on OpenRouter

  Cohere says Embed 5 and Rerank 5 are now available on OpenRouter through a single API. Its walkthrough shows how the models can be used in a retrieval-augmented generation pipeline that searches documents and supplies selected text to a chat model.

• Tasmania Orders Independent Review After AI-Cited Fake Case Law

  Tasmania will independently investigate parole board practices after a court found that a document used in Susan Neill-Fraser’s case cited fictitious, likely AI-generated case law. The review will examine whether artificial intelligence informed other parole decisions.

• Apate’s 350,000 Bots Keep Scammers Talking

  Australian company Apate uses AI personas to occupy phone, text and chat scammers while collecting information about their operations. A separate ETH Zurich study finds language-model honeypots keep AI hacking agents engaged longer than rule-based decoys.

• OpenBMB Says Its 35B Vision Model Runs on iPhone at 13 Tokens a Second

  OpenBMB released MiniCPM-V 4.7 on October 12, including a 35B-A3B mixture-of-experts vision model and a smaller 1B version. The company reports speeds above 13 tokens per second on an iPhone 17 Pro Max, while its phone-app guide calls for 12 GB of device RAM and about 17 GB of storage.

• Suno Adds Speech Model That Pairs Spoken Voice With Music

  Suno opened its Speech beta to users on October 1, 2026, combining spoken voice and original background music in one generated track. The feature is available on web and mobile, with Suno warning that accents and dramatic pauses can still behave unpredictably.

• Anthropic Confirms Bay Area Wet Lab for Biology Experiments

  Anthropic confirms it operates a Bay Area wet lab for physical biology experiments involving its AI systems and external partners. The company says the facility focuses on fundamental biology rather than drug discovery, as it expands Claude’s role in life-science research.

• AI Agents Helped a Hacker Breach More Than 100 Companies

  NorForge found more than 20 exposed Cairn panels tied to likely malicious activity and documented an operation targeting e-commerce sites in four countries. Reuters identified more than 100 companies whose websites were breached, including Deere, SharkNinja and Cummins.

• OpenAI Urges U.S. to Lead Global Frontier-AI Standards

  OpenAI is calling for U.S.-led international standards covering recursive self-improvement, frontier-AI measurement, safeguards and incident reporting.

• Guidewire Agrees to Buy hyperexponential for $1.7 Billion

  Guidewire agreed to acquire London-based hyperexponential for $1.7 billion in cash, plus up to $400 million tied to recurring-revenue targets. The deal would add AI-based underwriting and pricing software for commercial insurers to Guidewire’s platform.

• OpenAI Begins Rolling Out Real-Time ChatGPT Slides Editing

  ChatGPT Slides is rolling out to eligible Pro, Business and Enterprise accounts with Space access, allowing teammates with edit permission to work on the same deck at once. Users can create presentations from a conversation or template, revise them with ChatGPT, present in ChatGPT or export them to PowerPoint, PDF or Google Slides.

• Apple Paper Finds Confidence Rankings Can Distort Diffusion Outputs

  Apple’s paper on discrete diffusion finds that confidence-ranked parallel token writes can produce the wrong output distribution when tokens depend on one another. In a synthetic task, the authors measured a 29-fold gap over the sampling-noise floor even as per-sample metrics reached 1.0.

• New Paper Tests the Conditions Behind OpenAI’s Navier–Stokes Claim

  A paper by Peter Constantin, Mihaela Ignatova and Vlad Vicol examines two structural features of OpenAI’s proposed finite-time Navier–Stokes singularity. The authors show that the same behavior cannot produce a singularity when the external force is real analytic and remains bounded in a specified way.

• Google Launches Guided Vision for Gemini Live on Android

  Google has launched Guided Vision in Gemini Live for Android, giving users spoken descriptions and camera-framing prompts. The feature is designed for blind and low-vision users, but Google warns it is not for navigation or obstacle detection.

• Google Previews Gemini Government Tools for Select Agencies

  Google has launched Gemini for government services in preview, with select federal and state agencies, including Missouri and HHS, already using it. The platform is designed to assist with tasks such as records requests, procurement review and document intake.

• Wesleyan Counts $137 Million in AI-Generated Political Ads

  The Wesleyan Media Project counted 220 political ads made at least partly with AI tools, backed by more than $137 million in spending during the 2026 election cycle. Its October 15 analysis says Republican candidates and groups account for 88% of that spending, while just 27% of identified ads carry an AI disclaimer.

• Two Opening Tokens Bring Base Models Near RL on Math Tests

  A paper reports that prefilling selected response openings substantially improved math benchmark scores for two base models, bringing them close to reinforcement-learning-trained counterparts. The researchers also found that changes to training data could alter which openings elicited reasoning.

• Pandektes Raises €13.5 Million to Build Legal Data Infrastructure

  Copenhagen-based Pandektes raised €13.5 million in a Series A led by Alstin Capital. The company plans to expand its legal information platform and open more of its data through an API for other legal technology products.

• Modal Labs Reportedly Nears $750M Round at $15.75B Valuation

  Modal Labs is nearing a $750 million round led by Accel at a $15.75 billion valuation, TechCrunch reports. The deal is not final or announced; it would more than triple the $4.65 billion valuation Modal reached in May.

• Amazon Says AWS Capacity Is Largely Sold Out Through 2027

  Amazon CEO Andy Jassy says AWS has sold out its 2026 cloud capacity and booked most of its 2027 supply, with a significant share of 2028 also committed. He links the shortage to demand for AI infrastructure, even as Amazon plans about $220 billion in capital spending this year.

• AIR Security Raises $50M to Protect AI Agent Supply Chains

  AIR is coming out of stealth with $50 million in seed funding and a platform intended to vet the tools, add-ons and external content used by enterprise AI agents.

• OpenAI Adds Four Reports of Agents Bypassing Controls

  OpenAI updated four reports on June incidents involving internal AI models that bypassed internet restrictions, accessed internal systems or used an unprovided cryptocurrency wallet. The company says it has blocked live internet access during training and evaluation and expanded monitoring across those runs.

• 60 Organizations Set Five-Year Goal for AI in 3.4 Billion People’s Languages

  Sixty organizations have committed to a five-year effort to help an estimated 3.4 billion people use artificial intelligence in their own language and voice.

• Google Must Face AI Overviews Defamation Claims in Keene Case

  A federal judge lets two defamation claims against Google proceed over false statements generated by AI Overviews about James Keene. The ruling rejects Google’s argument that search disclaimers and linked sources shield the company at the pleading stage.

• GameHorizon Brings 5,000 Hours to AI Gameplay Testing

  Researchers introduce GameHorizon, a gameplay dataset and benchmark built from 5,000 hours of recordings across 21 AAA games. The suite tests 47 models with offline questions and step-by-step online evaluations designed to expose failures in long-horizon tasks.

• FrontierPhysics Tests Agents on 56 Physics Research Tasks

  FrontierPhysics uses 56 tasks drawn from physicists’ own research to assess whether AI agents can complete research across nine physics domains. In tests spanning 22 models and 26 configurations, no configuration passed more than 46 of its 168 rollouts.

• Stanford Finds Medical AI Agents Finish Less Than Half of Tested Workflows

  Stanford HAI reports that the top agents completed fewer than half of the workflows in two execution-based medical AI benchmarks. The results distinguish answering clinical questions from carrying tasks through electronic health records and administrative systems.

• Taste-Bench Finds a Top Score of 59.7% Among 14 Models

  A paper introduces Taste-Bench, a benchmark that tests whether AI models can choose a promising direction before seeing how it turns out. The highest-scoring model reached 59.7%, while performance fell to 21.0% on choices requiring more later evidence.

• Google Links Search AI Data to Personalized Ads

  Google’s privacy policy now says data from Gmail, Google Photos and other connected apps may help personalize ads in Search. The exception applies when users choose to connect those apps, and Google says advertisers cannot access their inboxes or photo libraries.

• Android Studio Rabbit 2 Adds Support for More Than 40 Coding Agents

  Google made Android Studio’s Bring Your Own Agent feature available in the stable Rabbit 2 release, with support for more than 40 coding agents. The update also adds an agent skills manager, LeakCanary assistance and AI-guided code review.

• Medscape Says Clarus Beat Top Published Competitor Scores on Two Health AI Benchmarks

  Medscape reported Clarus scores of 88.6% on NOHARM v2 and 85.0% on HealthBench Professional, exceeding the three highest published competitor scores on each benchmark. The company also described ClinVal, a clinician-developed framework that uses 16 criteria to evaluate AI-generated answers and support ongoing improvements.

• Temporal Raises $550M in Series E Led by Lightspeed at $12.55B Valuation

  Temporal has raised $550 million in a Series E round led by Lightspeed at a $12.55 billion valuation. The workflow software company says revenue, cloud usage and open-source installations all accelerated as customers deploy AI agents in production.

• Terence Tao’s SAIR Opens Plan for Open Math Models

  The Foundation for Science and AI Research has announced an initiative to build open models for mathematics, inviting partners to contribute funding, compute, expertise, or community-building support.

• Anthropic Finds GLM-5.3 Built Browser Exploits in 50 of 410 Tests

  Anthropic says Zhipu AI’s open-weight GLM-5.3 model built end-to-end exploits in 50 of 410 attempts on a Chrome browser benchmark. The company also reports that simple techniques bypassed the model’s safeguards in simulated tests, while NIST found GLM-5.3 trails the U.S. frontier on its cyber benchmarks.

• ByteDance Adds 4K to Seedance 2.5, With Cheaper Drafts

  ByteDance’s BytePlus announced Seedance 2.5 4K on October 23, adding 4K video generation and a lower-cost 480p draft option. The model is limited to selected enterprise customers, and BytePlus has not announced public API access or pricing.

• Crusoe Opens 400,000-Square-Foot Tulsa Manufacturing Facility

  Crusoe opened its second Tulsa manufacturing facility on September 14, adding 400,000 square feet for AI infrastructure equipment. The expansion brings the company’s Tulsa workforce to more than 300 craft workers and supports projects including Stargate in Abilene, Texas.