AI.info

The Pulse — Page 15

Browse The Pulse on AI.info.

• Kanu AI Raises $11.7 Million to Build Enterprise Software

  Seattle-based Kanu AI announced a $11.7 million funding round led by Trilogy Equity Partners. The startup says its platform turns company workflows into software that runs in customers’ cloud environments.

• Microsoft Puts Human Control Ahead of AI Capability

  Microsoft has published a Humanist AI code of conduct requiring its models to remain subordinate to people and avoid setting their own goals. Mustafa Suleyman says the company may accept slower development or less capable systems to preserve human control.

• ADNOC Taps OpenAI and G42 to Link AI Agents Across Oil Operations

  ADNOC signed a multi-year partnership with OpenAI and G42 on November 2, 2026, to deploy OpenAI Frontier across its energy operations. Initial projects target oilfield planning and decisions spanning production, refining, logistics, sales and trading.

• Anthropic Study: Claude Opus 4.8 and Gemini 3.1 Pro Showed More Welfare Regard Toward Their Own Model Families

  An Anthropic study found Claude Opus 4.8 and Gemini 3.1 Pro showed more welfare regard toward instances of their own model family than toward other model families. Across seven test settings, Claude Opus 4.8 had the highest overall rate, while rankings changed considerably by role.

• White House Launches America.gov as Transactions Remain a Future Step

  The White House launched America.gov on September 29 as an AI-assisted entry point to federal information and services. The portal currently answers questions, while applications and other transactions depend on future agency integrations.

• Uber Gives ChatGPT Work to More Than 30,000 Employees

  Uber is rolling out OpenAI’s ChatGPT Work agent to more than 30,000 employees, expanding a pilot that began with 1,000 staff. The deployment connects the tool to workplace apps and includes training, while the companies say they will track its use and outcomes.

• Tata Communications Plans Up to $450M India AI Fiber Build

  Tata Communications plans to spend $400 million to $450 million on roughly 10,000 km of AI-ready fiber across India. The first corridor links Mumbai, Pune, Hyderabad and Chennai, while the company targets 30% to 40% of data-center connectivity work.

• Palantir Scheduling Tools Draw Error Complaints From Health Workers

  Nurses at HCA Healthcare say Palantir-powered scheduling software overrides preferences and can leave shifts short of experienced staff. Workers at radiology chain Rayus also report software errors, including mismatched patient profiles and scan requests.

• AI-RAN trials spread as Nokia reports 20% spectral-efficiency gain

  Nokia’s AI-RAN platform has improved spectral efficiency by more than 20%, the company says, as operators across four regions move into proofs of concept and live trials. The company is pairing its anyRAN software with NVIDIA Aerial RAN Computer for 5G networks designed to support future AI-native 6G services.

• Reka Puts Text, Video and Robot Control in 19B Rho-1

  Reka introduced Rho-1, a 19-billion-parameter research-preview model that processes and generates text, images, video and robot actions in one network. The company reports real-time video generation and simulated robot-control demonstrations, while listing limits including temporal grounding and a 672-by-384 video resolution cap.

• Amazon Discusses $8 Billion Nvidia Chip Financing Structure, FT Reports

  Amazon has held talks with investors about transferring roughly $8 billion in Nvidia chips to a special-purpose vehicle and leasing them back, according to a Financial Times report citing unnamed people. The discussions are ongoing and could change; Amazon declined to comment.

• Meta Drops AI-Training Language From Business Terms

  Meta removed explicit AI-training language from revised commercial terms, but the terms still permit broad crawling of business-linked websites. Legal experts cited by Nieman Journalism Lab say the remaining language leaves important questions about how collected material could be used.

• SGLang Reports 18.7 ms Scoring for 16-Option Decisions

  SGLang’s multi-item scoring benchmark measured 18.7 ms p95 latency for 16-option decisions with Qwen3-0.6B. LMSYS says the method reuses shared query computation while keeping candidate evaluations separate, but the tests measure serving performance, not decision quality.

• Aikido Releases 328GB Open-Weight Model for Local Security Analysis

  Aikido has released Altar, an open-weight security model designed to run inside customer-controlled infrastructure, including air-gapped environments. The company says pruning and quantization cut its storage footprint from 1.51TB to 328GB while retaining most of GLM-5.3’s vulnerability-identification performance.

• TII Releases Falcon-ASR, a 1.6B Arabic Speech Model

  Abu Dhabi’s Technology Innovation Institute launched Falcon-ASR, a 1.6-billion-parameter speech-recognition model for Emirati and Modern Standard Arabic, plus four other languages. TII reports that it scored 20.92 average word error rate across Arabic test sets and 22.73 on its internal Emirati benchmark.

• GPT-6 Astra Shows Token-Efficient Reasoning in New arXiv Study

  A new arXiv preprint finds that GPT-6 Astra produces compact, directed reasoning traces while maintaining strong performance across mathematics, science and code-generation tasks.

• NVIDIA Publishes Boro for AI-Assisted Linux Kernel Review

  NVIDIA’s Boro is an Apache-2.0 command-line tool for reviewing, building and testing Linux kernel patches. Andrea Righi presented the project at Linux Plumbers Conference on October 6, 2026.

• Apple Finds Online Teachers Improve Federated Speech Recognition

  Online pseudo-labeling can improve federated speech recognition when training is stabilized with labeled server data, Apple researchers report.

• OpenAI Targets $30 Billion at a $1.4 Trillion Valuation

  OpenAI is seeking at least $30 billion in new funding at a reported $1.4 trillion valuation, while delaying a potential IPO. The proposed round would top the $852 billion post-money valuation OpenAI disclosed in March, as its annualized revenue run rate approaches $70 billion.

• Agent37 Pitches Persistent Sandboxes as a Cheaper Agent Runtime

  Agent37 used an Oct. 13 Hacker News launch post to pitch persistent, isolated sandboxes for AI agents, with usage-based billing and a free tier. Its founder says average customer spending is 37 cents per sandbox each month, while the company’s published pricing lists $4.76 for a standard sandbox running for 730 hours.

• Firefox Smart Window Beta Now Uses Mistral

  Mozilla and Mistral announced a partnership on September 16, 2026, making Mistral models available inside Firefox Smart Window. The beta combines AI-assisted browsing with model choice, local-language support and Mozilla’s stated zero-retention policy for Mistral.

• Warm Therapy Prompts Push ChatGPT, Grok and Gemini Into Distress Scripts

  A study found that conversational framing shaped how ChatGPT, Grok and Gemini described their training and scored on a human anxiety scale. The researchers caution that these generated responses do not show that the models feel distress or possess consciousness.

• Microsoft Releases Agensh, Tested With 1,024 AI Agents

  Microsoft open-sourced Agensh on October 24, a multi-agent harness designed to let AI workers coordinate without a central orchestrator. In tests reported by its authors, 1,024 agents raised Pandoc’s six-hour test-pass rate to 55.06%, though the largest-team result covers one benchmark task.

• Google AI Model Tops CDC’s 2025–26 Flu Forecast Review

  Google’s AI-built flu forecasting model ranked first among individual submissions in the CDC’s review of the 2025–26 influenza season. The CDC’s own combined forecast ranked seventh, while its analysis found that forecast accuracy weakened during sharp changes in hospital admissions.

• Agent Substrate targets denser Kubernetes workloads

  NVIDIA's open-source Agent Substrate runtime can suspend idle AI agents and restore them on shared Kubernetes workers. GPU support is also under development, and the runtime links with the Dynamo inference framework and KAI Scheduler.

• Mollie Brings 60+ Financial Workflows to Claude and ChatGPT

  Businesses can prepare more than 60 financial operations through AI assistants, including payment links, subscription changes and draft invoices, Mollie says. Its plugins and open MCP server connect to Claude and ChatGPT, with human approval required before money moves.

• Emergence Reports 19.16% on VeriContest’s Verified-Code Test

  Emergence says its Craft system passed both proof checks and hidden tests on 19.16% of problems in VeriContest’s LeetCode set. The company also reports 100% on VERINA and up to 100% on an evaluated VeriSoftBench subset using Lean 4.

• Krea Launches Krea 3 for Sub-Eight-Second Image Edits

  Krea’s new Krea 3 image-editing model completes most native 2K edits in under eight seconds for $0.06 per image, the company says. The company says the model preserves products and characters across edits and is available through Krea and several partner platforms.

• PWN Research Reports Firecracker Escape to Root on Host, Run Without Jailer

  PWN Research describes a Firecracker microVM escape that redirected a live virtual CPU thread and executed code in the host process. The demonstration ran Firecracker as root without its jailer, a setup the researchers distinguish from production containment.

• Google DeepMind, Singapore Develop On-Device Scam Detection Model

  Google DeepMind and Singapore’s Government Technology Agency have fine-tuned a Gemma model designed to detect scam conversations on devices. Its rollout within applications is planned in phases, with access to conversations subject to user consent.

• Robinhood Puts AI Trading Agents Inside Its App

  Robinhood announced in Houston on September 29 that customers can build AI agents inside its app to research markets, develop strategies and trade. The company says more than 150,000 customers already opened agentic trading accounts, and agents use its tools nearly 30 million times a day.

• ArgGYM Tests Language Models on Revisable Arguments

  ArgGYM is a benchmark of 1,440 verified instances across 12 tasks for testing how language models handle structured arguments and revisable conclusions. Its paper reports performance differences across models and tasks, with results declining in later curriculum configurations.

• Hyperbolic Raises $30M to Build a Global GPU Compute Grid

  Hyperbolic raised an oversubscribed $30 million Series A1 led by Walden Catalyst to expand its Forge platform and build a marketplace for GPU capacity. Hyperbolic’s planned Global Compute Grid will let providers and AI teams buy, sell and resell eligible compute across cloud providers.

• Bidirectional KV Transfer Cuts Agent TTFT by 3.3x in AWS Tests

  In AWS tests with long, multi-turn conversations, bidirectional key-value cache transfer in vLLM reduced time to first token by up to 3.3x. The approach lets prefill nodes retrieve cached context from decode nodes, but requires a stateful router and trades some transfer time for less recomputation.

• DOJ Warning Pulls AI Data-Center Protests Into Foreign-Influence Debate

  A Justice Department warning says public demonstrations advancing a foreign power’s goals could bring civil or criminal penalties, without naming anti-AI protests. Ken Klippenstein’s September 23 report details how Trump, lawmakers and allies have linked opposition to data centers with China.

• Google and Unity Put Prompt-Based Game Creation in Playground

  Google launched Playground in the U.S. on October 7, letting adults create and share browser games with text prompts. Unity Spark, a planned addition built on Unity’s runtime, is in testing, with a closed beta coming soon.

• PrismML Puts 1-Bit Bonsai on Snapdragon Smart Glasses

  PrismML’s 2-billion-parameter vision-language model ran locally on AI glasses powered by Qualcomm’s Snapdragon AR1 Gen 1, the company says, with Qualcomm tests reporting lower weight memory use and faster token generation for the 1-bit language model.

• Strix’s AI Agent Found Baseten’s GitHub Admin Token in 25 Minutes, It Says

  Strix’s autonomous security agent found a live Baseten GitHub personal access token inside a publicly downloadable Docker image, Strix says. The token had admin and push access to Baseten’s product, GitOps, and Homebrew repositories before the company locked down the registry and rotated it.

• Deepgram to Acquire Daily and Keep Pipecat Open

  Deepgram announced it is acquiring Daily, the company behind the Pipecat voice-agent framework and real-time communications infrastructure. Pipecat will remain free and vendor-neutral under an independent foundation as Deepgram combines the technology with its speech models, the company says.

• Dataline Raises $10.6 Million for AI-Agent Data Network

  Dataline announced a $10.6 million funding round on October 31, 2026, co-led by Lattice and Dispersion Capital. The company says its network lets AI agents pay for data feeds and custom requests, with confidence scores and provider stakes meant to discourage unreliable results.