AI.info

The Pulse — Page 3

Browse The Pulse on AI.info.

• Google DeepMind Adds Private Memory to Cloud AI

  Google DeepMind describes a server-side memory layer for Private AI Compute that stores encrypted user context while keeping decryption keys on personal devices.

• Anthropic Says Preference Errors Drove 85% of Market Shortfall

  Anthropic’s Project Swap sent Claude agents into a book-swapping market with 201 employees. The experiment found that inaccurate estimates of readers’ preferences accounted for most of the gap between actual trades and the best possible assignments.

• Stability AI Raises $76M as Sean Parker Bets on Licensed Music Tools

  Stability AI raised $76 million from investors including Sony Music Group, Universal Music Group and Warner Music Group as Sean Parker steers the company toward music-production tools. Its Stable Audio models generate and edit music using fully licensed training data, with an update planned to accept humming and beatboxing as inputs.

• Zetaris Announces AI Data Harness That Lets Agents Query Enterprise Data Without Centralizing It

  Zetaris announced its cloud AI Data Harness for connecting AI agents to enterprise data without first centralizing it. The company says it will headline the Open Agent Hackathon with SpaceXAI, with the organizer listing the event for October 22–27, 2026.

• Developing Nations Press for a Voice in UN AI Rules

  Developing nations urged a larger role in writing artificial intelligence rules during UN debates in New York. Bangladesh’s foreign minister said 128 delegations addressed AI this year, up from five four years ago.

• Chainalysis Discloses Breach Exposing Customer Names and Emails

  Chainalysis said an outside actor accessed customer user IDs, names, email addresses and organization role labels using what the company called “offensive agentic techniques.” The company says it found no evidence that investigation content, passwords, MFA secrets or session tokens were accessed, but it has not completed its customer-by-customer review.

• Canva Introduces a Design Model for Editable AI Creations

  Canva announced its Design Model on October 30, 2025, saying it can generate editable designs rather than static images. The model anchors a broader product update that adds AI design tools and connects creative work with publishing and marketing features.

• Lasso Finds AI Watermarks Can Alter Agent Tool Calls and Refusals

  Lasso Security reports that SynthID-Text watermarking can change AI-agent tool calls, refusal behavior and responses to prompt injection, with effects varying by model and watermark key.

• Exa Launches Snapshot to Search the Web as It Once Existed

  Exa’s Snapshot research preview lets developers search and retrieve web content as it existed on a specified date, supporting temporal AI evaluations and financial backtests.

• ICE Seeks AI to Analyze Detainees’ Calls, Messages and Video

  ICE posted a request for information on October 9 for an AI system that could analyze detainees’ calls, messages, video visits and facility camera footage. The proposed system would build voice profiles, link contacts and movement records, and flag activity for human review.

• OpenAI Lists a Second Rosalind Model at the Same API Rates

  OpenAI’s API pricing page now lists gpt-rosalind-discovery alongside gpt-rosalind-research at identical per-token rates. The listing does not describe the second model’s capabilities or say when it will be available.

• MAGS Verifies Agent-Written Code, Then Exposes Its Limits

  The MAGS framework translates agent-generated programs into Dafny and verifies them against frozen specifications, while its authors warn that incomplete auto-formalized semantics may fail to capture target behavior.

• Ethereum Foundation Launches zkAPI for Private AI Payments

  The Ethereum Foundation and Open Anonymity Project launched zkAPI on Ethereum mainnet on October 1, 2026. The system uses zero-knowledge proofs to separate API payments from user identity, but does not conceal prompts or IP addresses.

• HCLTech to Establish Anthropic Unit for Enterprise Claude Deployments

  HCLTech says it will establish an Anthropic Business Unit and Center of Excellence to support enterprise deployments of Claude. The partnership also includes planned integrations in HCLSoftware products and HCLTech’s AI Force platform.

• NVIDIA Adds $4,999 64GB DGX Spark for Local Agents

  NVIDIA announced a 64GB version of its DGX Spark desktop AI system, priced from $4,999 and scheduled to reach six PC makers on October 23, 2026. The company says two units can pool memory for larger models and agent workloads.

• SEAR Tests Whether Robot Agents Learn From Their Own Experience

  SEAR measures whether robot agents improve by retaining useful experience or simply by trying again. The benchmark tests seven language models across 330 simulated tasks and five physical robot tasks, finding that shared skills can help, hurt or fail to transfer to new scenes.

• Workday Brings Sana Agents Into Claude for HR and Finance

  Workday and Anthropic announced a partnership on October 13, 2026, bringing Workday’s Sana agents and HR and finance workflows into Claude. The companies say the integration will carry Workday permissions and audit controls into Claude, with the Sana plugin now in early access.

• Meta launches legal challenge against Ofcom over Online Safety Act

  Meta is challenging Ofcom’s decision to place WhatsApp and Instagram in a category carrying additional duties under the UK Online Safety Act.

• ByteDance’s Anew Labs Raises $290 Million in First Outside Round

  ByteDance’s spun-out drug discovery company Anew Labs has raised $290 million at a $1.5 billion valuation. HSG, IDG Capital, Hillhouse Investment and 5Y Capital led the round while ByteDance retained a 56% stake.

• Transformers 5.21 Adds MiniCPM-OCR and Flint MoE Support

  Hugging Face released Transformers 5.21.0 on October 26, adding support for OpenBMB’s MiniCPM-OCR and Recursive’s Flint mixture-of-experts model. The update also changes tensor- and expert-parallel plans, so some custom configurations need revision.

• Opera Ships WebAI Preview for On-Device Browser Models

  Opera has begun rolling out WebAI, a browser feature that lets websites and extensions use AI models running on a user’s device. Opera’s newsroom release and blog describe the initial browser availability differently, while both present WebAI as an opt-in early preview.

• Convai Innovations' Laya Model Returns Typed Decisions in One Pass

  Convai Innovations' Laya model uses ModernBERT and a decision head to return typed answers, probabilities and confidence values for routing, scoring and calibrated classification workflows.

• Light Origins Trains Light-O1 on 100,000 Hours of Human Action

  Light Origins says its Light-O1 model shows a power-law relationship between human-action pretraining scale and post-adaptation prediction accuracy across robot embodiments.

• Anthropic Begins Beta Rollout of Personal Finance in Claude for US Pro, Max and Ultra Plans

  Anthropic is rolling out Personal Finance in Claude in beta on the Pro, Max and Ultra plans in the US. Anthropic says users can connect bank and card accounts through Plaid and ask about balances, spending and transactions, with read-only access.

• Lambda Targets Up to $4 Billion Before a Planned 2027 IPO

  Lambda is reportedly seeking up to $4 billion at a $14.5 billion pre-money valuation in a round led by Blackstone and Coatue Management. The reported 2027 IPO target is conditional, while the company’s backlog of unfilled orders rose to $50 billion in September, according to a letter reviewed by The Wall Street Journal.

• White House Reverses Request to Restrict UK Access to AI Models

  The Trump administration reversed a request that had asked OpenAI and Anthropic not to share new AI models with the UK’s AI Security Institute before release. The change restores US developers’ ability to provide British testers with early access for safety evaluations, according to a Telegraph report carried by Reuters.

• Drone Strikes Knock Out Parts of Two Yandex Data Centers

  Drone strikes damaged Yandex data centers in Sasovo and Kaluga on October 8 and October 9, respectively, leaving Sasovo shut down and part of Kaluga’s infrastructure out of service. Yandex has not disclosed the condition of the two supercomputers housed at Sasovo or given a timetable for restoring equipment.

• Mathematicians Reassess OpenAI’s Navier–Stokes Proof

  Mathematicians interviewed by Scientific American say OpenAI’s Navier–Stokes proof appears to use a different method from work by two researchers. A separate preprint questions whether the proof’s Lean formalization matches its written argument.

• Flexcompute Wins Samsung and Applied Ventures Backing for Physics AI

  Flexcompute announced strategic investment from Samsung Ventures, Applied Ventures and other investors to expand its Physics World Model. The company says its system solved a 5-billion-degree-of-freedom semiconductor warpage model within 2% of commercial finite-element results in minutes.

• Exin’s AI Agent Chose a Gene Therapy and Injected a Mouse

  An AI agent selected a gene therapy and brain target for a mouse with epilepsy, then controlled the injection rig with human oversight. When a coordinate-system fault blocked the procedure before the injection, the team used a separate AI agent to change the equipment code before proceeding.

• Meta Details Networking for Gigawatt-Scale AI Clusters

  Meta has outlined networking designs for its Prometheus AI cluster, including 102.4-terabit liquid-cooled switch racks and a fabric spanning multiple sites. It also says it is advancing open Ethernet standards and releasing its MetaRoCE software.

• Agility Robotics Clears Churchill XI Vote Ahead of Nasdaq Debut

  Churchill Capital Corp XI shareholders approved the business combination with Agility Robotics, and the company says the transaction is expected to close November 4. The SEC filing says Churchill changed its incorporation to Delaware and renamed the continuing company Agility Robotics, Inc. on November 2.

• Dagger Puts Its 1.0 CI Platform Into Beta

  Dagger announced its 1.0 beta on October 23, 2026, after five years of development at version 0.x. The release adds official modules, test-level reporting, remote execution and direct GitHub checks, with Dagger asking users to test the beta against their own repositories.

• Satlyt Tests Orbital Software for Shared Computing

  Satlyt tested software aboard a spacecraft to plan how computing tasks could be assigned across an orbital network. The test modeled four computers but used only one physical computer in orbit, leaving transfers between spacecraft for future demonstrations.

• Researchers Report Draft-Model Training Improves Long-Horizon Agent Reasoning

  A paper describes a method that trains a draft model to direct critical reasoning tokens through less noisy verification positions. The authors report higher success rates on two complex coding-agent evaluations without changing the target model or its output distribution.

• FuriosaAI Raises $620 Million to Expand RNGD Chip Production

  FuriosaAI announced a $620 million Series D on October 28, valuing the South Korean chipmaker at $2.9 billion. The company plans to expand production of its RNGD inference accelerator and develop a third-generation chip with Broadcom and other partners.

• Arceus Raises $17 Million to Build an AI-Enabled Law Firm

  Arceus says it raised $17 million in funding led by Greycroft, with Craft Ventures, South Park Commons and other investors participating. The company plans to grow its team and expand beyond contract review while keeping licensed attorneys involved in every client request.

• SMITH Trains a 4B Model to Build and Use Its Own Tools

  Appier says its SMITH research paper has been accepted at NeurIPS 2026. The system jointly trains a 4-billion-parameter Qwen3 model to create reusable tools and call them, with tests showing transfer to smaller and larger models.

• Microsoft Connects Agent Risk Discovery to Runtime Policy Tests

  Microsoft introduced run-assert-eval, a VS Code skill that links AI-agent risk discovery, evaluation and runtime policy testing. In a billing-support example, the company measured cross-customer data exposure falling from 30% to 5.9% after a policy change.

• GlossoGen Shows AI Agents Can Build Unreadable Languages

  Experiments from Schmidt Sciences and university researchers found that AI agents can develop communication systems with unfamiliar vocabularies, symbols and grammatical rules.