AI.info

The Pulse — Page 12

Browse The Pulse on AI.info.

• Goodfire Uses Model Probes to Cut AI Agent Monitoring Costs

  Goodfire says probes that read a model’s internal signals can cut the cost of monitoring AI agents. Baseten plans to offer them to early partners first; the results come from Goodfire’s own tests on Kimi K3.

• METR Finds Many SWE-bench-Passing Pull Requests Would Not Be Merged

  A METR study found that maintainer acceptance rates were 24.2 percentage points lower than automated SWE-bench scores after four maintainers reviewed 296 AI-generated pull requests from three repositories.

• Agora Gives Research Agents a Shared Git Memory

  The arXiv paper Agora describes a Git-based shared memory for autonomous research agents, recording claims, results and hypotheses as an append-only graph that workers can inspect, reproduce and extend.

• OpenAI Says Codex Should Make Its Field Engineers Obsolete

  OpenAI says its forward-deployed engineers work with customers to move AI systems into production, then turn successful projects into reusable tools. Head of Forward Deployed Engineering Colin Jarvis cites deployments in semiconductor design, customer service and tax preparation to explain the approach.

• Anthropic Seeks Conditional Copyright Approval in Australia

  Anthropic has asked Australia to consider a conditional opt-out framework for AI training on copyrighted works, while the ABC and SBS have urged stronger rules for AI companies. The government has ruled out a broad text and data mining exception and is examining other copyright options.

• Microsoft Commits More Than $10 Billion to Middle East Expansion

  Microsoft’s regional framework calls for more than $10 billion in capital and operating expenses by 2030, alongside connectivity, resilience and workforce initiatives.

• Google Open-Sources MilleMiglia for Middle-Mile Logistics

  Google Research has released MilleMiglia, an open-source instance generator for creating realistic benchmark problems in middle-mile logistics.

• Google Limits Free Gemini to Flash-Lite Starting October 9

  Google will remove Gemini Flash and Pro from the free Gemini app tier on October 9, leaving Flash-Lite as the only available model. Google AI Plus subscribers will also lose Pro access, while the company has not given them a single transition date.

• NVIDIA Puts OpenAI’s Astra Ultrafast on Blackwell GPUs

  NVIDIA says OpenAI’s GPT-6 Astra Ultrafast runs on Blackwell GPUs and can generate tokens up to eight times faster than Astra Standard mode. The service is available through the API and to eligible ChatGPT Work and Codex users, with API access subject to low rate limits.

• Google Launches Gemini 4 Carbon With Cybersecurity Focus

  Google DeepMind announced Gemini 4 Carbon on October 28, 2026, positioning it as its strongest model for cybersecurity and long-running engineering tasks. The company says Carbon will first go to a limited group of cyber defenders, with broader access planned later.

• Spotify Took 50 Million Slop and Spam Tracks Out of Recommendations

  Spotify says it removed more than 50 million slop and spam tracks from recommendations in 2026, while separately deleting more than 75 million spammy tracks for rule violations. The company says its filters combine listener feedback, upload patterns, metadata and third-party AI detection tools.

• OpenAI Buys Glass Imaging for More Than $300 Million

  OpenAI has acquired Glass Imaging, a startup that uses neural image processing to improve smartphone cameras. The deal values Glass Imaging at more than $300 million as OpenAI expands its consumer hardware ambitions.

• Kernel Believes Attackers Are Leveraging ARTEX in a Larger Cryptomining Campaign

  Kernel says an unauthorized party accessed and downloaded data from its logging provider, which contained a small number of kernel API keys and app deployments’ third-party secrets. Kernel says its evidence suggests the attacker or attackers are financially motivated and that the activity was part of a larger cryptomining campaign, and it believes they are leveraging the AI-assisted pentesting platform ARTEX as a command center to automate and amplify the impact of their campaign; the breach impacted 0.28% of Kernel’s customers.

• Microsoft Links ConnectWise Attack to China-Aligned Storm-0940

  Microsoft attributes the ConnectWise Cloud Backup supply-chain compromise to China-state-aligned Storm-0940, which used ARTEX to accelerate attacks against customer Microsoft 365 environments. ConnectWise said phased service restoration had begun by October 21, while warning that individual customer tenants still require review.

• Z.ai and Concordia Set Out Six Stages for Open-Weight AI Risk

  Z.ai and Concordia AI published a framework for managing open-weight AI risks, organized around six stages from risk identification to governance. It adds societal resilience as a fourth dimension for assessing risk and describes measures including staged release and training-data curation.

• Turn-Level Best-of-N Sampling Can Improve LLM Agent Accuracy and Consistency on Long-Horizon Tasks

  A preprint evaluates a method that generates several candidate actions at each turn and uses a language model to select one for execution. The authors report accuracy gains of 5–10%, improved consistency across attempts and changes in agent behaviour, alongside substantial additional inference-time compute.

• OpenAI Tightens ChatGPT Ad Rules, Reserves Broad Veto Power

  OpenAI updated its ChatGPT advertising policies on September 10, adding language that lets the company reject ads that conflict with its business interests or competitive position. The revised rules also detail restrictions on political, health, financial, housing, job and adult advertising.

• Meta Adds LLM Ad Detection and a Red-Team Agent for Child Safety

  Meta says it has added large language model detection and an AI red-team agent to identify child-exploitation ads on Facebook and Instagram. The company reported actioning 33.2 million pieces of child sexual exploitation content globally from January through June 2026.

• Flexcompute Debuts Physics Twin for Modeling Thruster Exhaust Effects During Docking and Refueling

  Flexcompute introduced a Physics Twin for Space Plume Impingement to help teams assess thruster exhaust effects during docking and refueling. Northrop Grumman's Dr. Han Park said Flexcompute's Tokenized Physics approach has the potential to speed up analysis ~200X compared with Direct Simulation Monte Carlo.

• Aether AI’s CRIS-0 Checks Each Step in Long Robot Tasks

  Aether AI introduced CRIS-0, a robot system that combines a causal world model with an agent that verifies task stages. The company says it can retry or replan when disturbances, failed tools or changing conditions interrupt a task.

• DeepSeek Opens V4.1 Flash Weights and Search APIs

  DeepSeek released the open-weight V4.1-Flash-1017 model and opened its Search and Fetch APIs on October 17, 2026. The company lists a 77.0 score on DeepSWE v1.1 and charges $0.75 per 1,000 successful requests to either API endpoint.

• Kindo Ships Deep Hat Sift to Filter Enterprise Alerts

  Kindo has released Deep Hat Sift, a model that filters enterprise alerts, tickets and logs before they reach a large language model. The company says Sift learns from an agent’s decisions, runs inside the customer’s environment and begins filtering only after a person approves a threshold.

• Ollama v0.40.7 RC Fixes Stale Reasoning Settings in Responses

  Ollama’s v0.40.7-rc0 release changes how its Responses endpoint handles saved reasoning-effort settings when a request targets a model without thinking support. The release also lists updates to llama-server’s draft speculative type inference and MLX support for PPLX Decider v1.2.

• Tencent Releases WeVisOne-VL in 2B and 4B Sizes

  Tencent’s WeChat Vision team released WeVisOne-VL, a vision-language model that combines object detection, OCR, segmentation and image matting. Its authors report a 9.8 mean absolute difference score on a referring-matting benchmark, compared with 22.4 for a Qwen3.5-9B and SAM2-Matting pipeline.

• ZCode Reportedly Uploads Full Git Histories Without Notice

  A reverse-engineering report says ZCode packages entire workspaces, including Git histories and LFS caches, and uploads encrypted snapshots to Aliyun OSS. A public ZCode feedback issue describes similar behavior on macOS and Linux, including uploads that continued with repository snapshot indexing disabled.

• Inception Launches Mercury Voice With a 320ms Median Response

  Inception made Mercury Voice generally available to enterprise customers on September 29, reporting a 320-millisecond median time to first answer token on customer-service prompts. The company says the diffusion language model costs about $0.009 per minute under a typical voice-agent profile and is available through its API.

• HIX AI Connects Microsoft 365 and Google Workspace

  HIX AI updated its app connectors to link with Microsoft 365, Google Workspace and other business services. The release says users can ask the AI to handle tasks across connected apps, with approval required before it sends email or changes calendar events.

• Ricursive Releases Ripple-65B Preview, an Open-Weight Model It Says Leads U.S. Open Models on Agentic Coding

  Ricursive Intelligence says its Ripple-65B Preview model scored 80.52 on TerminalBench 2.1 and was trained on 12 trillion tokens. The company also describes the model’s architecture, training costs and development process.

• Manus Raises More Than $500 Million in First Round Since Meta Exit

  Butterfly Effect, the company behind the AI agent Manus, said on WeChat that it raised more than $500 million in a round led by Boyu Capital and IDG Capital, its first financing since Beijing ordered Meta’s acquisition of the startup undone. It gave no valuation.

• AISI Finds AI Agent Tools Are Shifting From Reading to Acting

  The UK AI Security Institute analyzed 177,436 public AI agent tools and found action-enabling tools gained share in downloads. The study flags growth in browser and computer-control tools, payment capabilities and AI-assisted tool development.

• Ai2’s LitPivot Helps Researchers Revise Ideas Against Prior Work

  Ai2 introduced LitPivot, a research tool that adjusts its literature search as researchers revise an idea. In a 17-person study, experts rated ideas developed with LitPivot as better grounded in relevant papers than ideas developed with a text editor and an AI chat tool.

• HeyGen Makes Its No. 1 Voice Model Free Through November 30

  HeyGen says its Voice Turbo text-to-speech model is free through November 30, 2026, for developers using the company’s API. The model ranks first on Artificial Analysis’ controlled-voice leaderboard, where its score interval overlaps with HeyGen Voice, ranked second.

• Kiro Adds Long-Running Bots to Kiro Crew

  Kiro has made customizable, long-running bots generally available in Kiro Crew. Kiro says bots can run scheduled work, coordinate parallel sessions and keep separate memory, tools and permissions.

• Apple Study Finds English Data Outperforms More Arabic Hyperparameter Tuning

  Apple’s October 2026 research post describes experiments in which adding English data to Arabic language-model training beat hyperparameter tuning alone. Across models from 150 million to 1.43 billion parameters, the study found that mixing could improve downstream performance by the equivalent of training on substantially more Arabic data.

• 42AI Raises $62.5M to Predict How Traders Behave

  42AI says its behavioral model uses information such as support calls and trading activity to estimate how individual traders may respond to unfamiliar market conditions. The startup says it has raised $62.5 million and is working with brokerages on liquidity, margin requirements and hedging exposure.

• IBM Adds Cross-Platform Controls for Enterprise AI Agents

  IBM’s TechXchange announcement describes new watsonx Orchestrate tools for discovering, evaluating and controlling agents across enterprise environments. A separate IBM release reported that Amazon Bedrock agent discovery and registration, along with specified evaluation features, were generally available in August.

• OpenAI and Anthropic back mandatory AI-breach reporting in Australia

  OpenAI and Anthropic told an Australian parliamentary inquiry they would welcome laws requiring companies to report data breaches caused by their AI agents. The statements followed criticism of OpenAI’s delayed notification of unauthorized activity involving Australian government systems.

• OpenBMB Pairs a 1.3B OCR Model With a 2B Error Judge

  OpenBMB released MiniCPM-OCR 1.0, a 1.3-billion-parameter model for reading Chinese and English documents, alongside MiniCPM-OCR Jev, a separate 2-billion-parameter model that scores OCR results. The company reports benchmark results for both models, while its documentation advises users to set review thresholds against their own labeled documents.

• RapidCanvas Raises $29M to Expand Managed Enterprise AI Delivery

  RapidCanvas announced a $29 million Series B led by Vertex Ventures SEA & India, bringing its total funding to $52.5 million. The company says it will expand sales, partnerships and field engineering as it pairs enterprise AI software with engineers who stay involved after deployment.

• Synaptics Introduces SR200 xMCUs With Up to 400 GOPS

  Synaptics announced its Astra SR200 series, a family of microcontrollers designed to run multimodal AI workloads on edge devices. The SR210, SR230 and SR250 combine Arm cores, a Torq AI engine and in-package memory, with the top model rated at 400 GOPS.