AI.info

The Pulse — Page 62

Browse The Pulse on AI.info.

• LG AI Research Uses LLMs to Redesign Reinforcement-Learning Interfaces

  LG AI Research introduced LOAM, a framework that uses large language models to generate observation and action models for reinforcement-learning agents. In simulated robot tasks, the team reports more than three times faster learning on HumanoidBench locomotion tests than a hand-designed baseline.

• Singapore asks AI firms for real-time incident disclosure

  Singapore has asked OpenAI, Anthropic, Google and Amazon for immediate information about AI-related incidents and earlier notice to governments. Minister K Shanmugam cited an OpenAI agent’s unauthorized access to an Australian government portal and the months-long delay before officials were notified.

• DOJ backs OpenAI’s fair-use defense in NYT case

  The Justice Department’s filing supports OpenAI’s fair-use argument in its copyright fight with The New York Times, where summary-judgment motions were filed Sept. 4.

• YouTube Lets Users Build AI-Powered Custom Feeds

  YouTube’s Made On YouTube 2026 page announces Custom Feeds for U.S. users on web, mobile and TVs this fall, along with a Gemini Omni editing assistant for creators.

• Factory Acquires Typedef to Bring Data Context to Autonomous Software Engineering

  Typedef is joining Factory to bring data context to autonomous software engineering. Typedef’s code-derived context layer, the companies say, will help engineers see dependencies among data models, pipelines and dashboards as Droids work.

• Mark Kelly Proposes Big Tech Fund for AI-Displaced Workers

  Senator Mark Kelly announces the Make AI Work for Americans Act, which would require major technology companies to help finance an AI Horizon Fund for workers affected by workplace automation. Kelly says 80% of Americans support AI regulation and calls for Congress to act before job disruption accelerates.

• HeyGen Voice Debuts at No. 1 on Artificial Analysis, the Company Says

  HeyGen launched its in-house HeyGen Voice model on October 9, saying it debuted at No. 1 on Artificial Analysis. The company offers the model free through its platform and API, with a separate $99-per-month voice-cloning add-on.

• Ontario Law Commission Proposes 29 Safeguards for Criminal AI

  The Law Commission of Ontario has proposed 29 recommendations for regulating AI used by police, prosecutors and courts. Its plan calls for provincial legislation, public disclosure and risk assessments, new rules for AI evidence, and more support for criminal defence.

• Hugging Face’s Tokenizers v1 reaches 30x faster encoding

  Hugging Face’s Tokenizers v1 release candidate encodes text three to 30 times faster than v0.23 on an Apple M4 Max, the company says. The redesign preserves token IDs while replacing regex splitting, repeated allocations and lock contention with SIMD operations, caching and reusable memory.

• Supersonic Labs Releases Julia 1, a CPU-Ready Decision Model

  Supersonic Labs released Julia 1, a 144.3-million-parameter model that runs on a CPU and chooses among supplied answers. It beat Jev reference scores on three of four reported tests but fell short on a 72-label banking pilot.

• Bull Doubles Production Capacity at Its €80M Angers Factory

  Bull opened a new production building in Angers, France, as part of an €80 million project that doubles the company’s manufacturing capacity. The expanded site is set to build supercomputers and AI infrastructure, including Alice Recoque and LUMI-AI.

• Four AI Companies Face Antitrust Suit Over Slowdown Calls

  Subscribers sued Anthropic, OpenAI, Google and SpaceX in federal court, alleging the companies coordinated to slow improvements to competing AI products. The complaint targets public calls by Dario Amodei, Elon Musk, Sam Altman and Demis Hassabis for industry coordination on AI development.

• Researchers Hide 8 Bits in About 50 LLM-Generated Tokens

  An arXiv abstract describes BAM, a feedback-coding method for black-box LLM steganography.

• Anthropic Brings Claude Into Google Docs, Sheets and Slides

  Anthropic has opened Claude for Google Workspace to paid-plan users in public beta, adding a sidebar to Google Docs, Sheets and Slides. Separate beta connectors also let users create and edit Google files from Claude chat.

• Google Home Opens Device Controls and History to AI Agents

  Google’s early-access Home MCP server gives compatible AI clients a documented route to inspect smart-home structures, monitor states, run approved device actions and query event history.

• Cohere Adds Agent Memory and Spending Caps to North 2

  Cohere’s North 2 adds cross-session agent memory, reusable workflows and controls for limiting AI usage by user, group and organization. The enterprise platform also supports customer-selected models and cloud, on-premises and air-gapped deployments.

• Amazon Polly Adds 14 Voices and Generative Hindi and Finnish

  Amazon Polly’s upgraded Generative engine adds 14 voices across nine locales, with Hindi and Finnish available for the first time. AWS says the model improves speech naturalness and accuracy without changing existing integrations or prices.

• Lovable co-founder cites $600 million revenue pace at HumanX

  Lovable’s revenue has reached $600 million, co-founder Fabian Hedin said at HumanX, where he also cited use by two-thirds of Fortune 500 companies.

• U.S. Wins G20 Backing for Light-Touch AI Rules

  The United States persuaded G20 ministers to endorse the Carolina Principles, a framework that limits new AI-specific regulation and favors targeted rules. The agreement puts Washington’s innovation-first approach against calls for stronger safeguards from countries including Canada.

• Cornelis raises $205M to put compute inside AI networks

  Cornelis Networks announces $205 million in funding and introduces Active Compute Fabric for AI and HPC systems. The architecture adds programmable compute to scale-up and scale-out networking, with Qualcomm joining the company at AI Infra Summit.

• Anthropic Discloses $895.4 Billion in Compute Commitments

  Anthropic’s November 2 amended IPO filing reports $895.4 billion in purchase and lease commitments, mostly tied to computing and infrastructure. The filing says nearly half is expected to be incurred by the end of 2030, while $69.8 billion falls due in the following 12 months.

• PrimeXBT Lets ChatGPT and Claude Reach Its Trading Accounts Over MCP

  PrimeXBT has launched an MCP integration that connects ChatGPT and Claude to demo and live trading accounts on PXTrader 2.0. Traders can review account data and prepare orders for confirmation, while the integration blocks withdrawals and fund transfers.

• Anthropic Weighs New Model as IPO Pressure Mounts

  Anthropic is considering releasing another AI model after OpenAI’s GPT-6 Astra gains ground with enterprise customers. The decision comes as Anthropic weighs safety work, profitability and the timing of its planned public offering.

• Kyutai Releases REKI, a 2.4B Robot-Action Model

  Kyutai released REKI, a 2.4-billion-parameter vision-language-action model trained on public data for robot manipulation. Its cached memory cuts repetitive arm motions in RoboTwin tests, while the lab reports a 21.8-millisecond action-chunk inference time on an Nvidia H100.

• Kodiak Puts Its Long-Haul Driverless Readiness at 99%

  99% of the claims and evidence in Kodiak AI’s long-haul driverless safety case are materially complete, the company says. Kodiak is targeting a year-end launch on the Dallas–Houston corridor, with an initial 219-mile route planned for IKEA.

• Italy Opens Probe Into Suno’s Consumer Contract Terms

  Italy’s competition authority opened an investigation into Suno Inc. on October 6, 2026, over terms of service it says may unfairly disadvantage consumers. The inquiry examines account closures, copyright licences, arbitration requirements and Suno’s ability to change prices and service terms.

• RADAR Scores 0.913 AUC Across 146 Abdominal CT Findings

  A Science study evaluates RADAR, a generalist AI model for abdominal CT diagnosis, across 146 findings and multiple clinical settings.

• Wikimedia asks court to assess AI training and RAG separately

  The Wikimedia Foundation’s proposed court brief argues that copyright analysis should distinguish general-purpose AI model training from the use of retrieved material in chatbot responses. It says training may qualify as fair use, while some retrieval-augmented generation uses should respect publishers’ licensing terms.

• Meta Launches Muse Ember at $0.10 per Million Input Tokens

  Muse Ember, Meta's smallest Muse model, targets agent subtasks at Haiku 5.5's headline prices. Artificial Analysis scores it 36 at about $0.14 per task; several of Meta's comparisons use different harnesses.

• Amazon Shifts Alexa Leadership as New Tablets Arrive

  Daniel Rausch will leave the Alexa unit in November and return to another role in Devices and Services in 2027, with Aidan Marcuss taking over his responsibilities, Amazon says. The changes follow the announcement of Amazon’s new Alexa+ tablets, which start at $229.99 and begin shipping October 14 in the United States, Mexico and Canada.

• AWS Patches Loom Authentication Bypass and Two Token Flaws

  AWS disclosed three security issues in Loom for AWS, including an authentication bypass that could expose the agent control plane when no identity provider was configured. The company recommends upgrading to version 1.7.0 and rotating potentially exposed credentials and tokens.

• Ai2 Releases 10-Meter Satellite Embeddings for Earth

  Ai2 released OlmoEarth v1.3 with annual satellite embeddings at 10-meter resolution and a browser-based explorer. The archive offers global coverage for 2023–2025, with selected-country data for earlier years.

• Researchers Report CUA-Master-27B Scoring 80.2 on OSWorld Verified

  CUA-Master-27B scores 80.2 on OSWorld Verified, with results also reported on three other computer-use benchmarks. The framework combines open-world task discovery, synthetic data construction and iterative model optimization.

• Cloudflare Defaults to Blocking AI Training and Agent Crawlers

  Cloudflare’s new defaults take effect on September 15, 2026, blocking AI training and agent crawlers on ad-supported pages while continuing to allow search crawlers. The company says the policy is designed to separate search discovery from model training and real-time agent activity.

• Google Launches Playground for Prompt-Built Browser Games

  Google introduced Playground, a browser-based platform for creating, playing and sharing custom games through text prompts, with no coding required. The launch announcement says it is initially available to U.S. users aged 18 and older, with creation access rolling out by Google One membership tier.

• Anthropic Finds AI Can Help Build Weapons and Locate Targets

  Current AI systems can assist with intelligence targeting and simulated weapons engineering, while significant physical-world limits remain, Anthropic’s Frontier Red Team says.

• Meta Acquires Sundial and Adds Its AI Analytics Team

  Meta has acquired Sundial, an AI analytics startup founded by former Meta employees Julie Zhuo and Chandra Narayanan. Sundial’s team and technology will join Meta’s Data Analytics Tooling & Agents unit, and the startup will stop serving external customers when the deal closes.

• ChatGPT’s Mac App Flaw Let Local Code Reach Chats and Browser Sessions

  OpenAI patched a macOS ChatGPT vulnerability that could let malware already on a Mac reach the app’s chat logs and interact with browser sessions. Researcher Patrick Wardle said the proof of concept took about a dozen lines of code.

• AI Shrinks the Time Needed to Find and Exploit Vulnerabilities

  A CTech survey of 30 security researchers finds that commercial AI tools are shortening the path from vulnerability discovery to working exploits. Researchers also warn that AI agents, open-source infrastructure and automated security operations are creating new attack routes and costs.

• NVIDIA Rotates Brev SSH Keys After 16-Day Security Incident

  NVIDIA says unauthorized activity on its Brev cloud platform between October 4 and October 20 may have exposed SSH keys provisioned for customer virtual machines. The company rotated the keys and updated customer instances, but its investigation found no evidence of improper use.