AI.info

The Pulse — Page 15

Browse The Pulse on AI.info.

• California Names Experts to Design Frontier-AI Kill-Switch Rules

  Governor Gavin Newsom appointed four experts to advise California on independent AI oversight and a possible emergency shutoff for frontier models. The panel will recommend whether state law should require lab monitors, verified safety assessments and tested kill switches.

• NYC Council Proposes AI Audits, Kill Switches and Whistleblower Rewards

  New York City Council Speaker Julie Menin announced bills requiring third-party validation and human shutdown controls for AI systems marketed or deployed in the city. The package also proposes whistleblower rewards, new ways to sue over certain AI harms and a citywide hearing on October 5.

• Qwen-Image 2.1 Released on ModelScope With Transparent Image Generation and Editing

  Qwen-Image 2.1 is now available as an open-source image-generation and editing model on ModelScope, with support for transparent RGBA images, localized edits and multiple reference images.

• Subconscious Raises $5.1M, Claims 80% Lower Agent Costs

  Subconscious has raised $5.1 million to commercialize an inference platform for long-running AI agents. The Cambridge, Massachusetts, startup says dynamic context compression can cut costs by up to 80% on workloads exceeding 200,000 tokens.

• ARC Prize Sets ARC-AGI-4 on Autonomous Invention

  ARC Prize says ARC-AGI-4 will target autonomous, open-ended innovation, while leaving the benchmark’s design unspecified.

• Grok Bot Reaches 418,000 Weekly Users After One Month

  SpaceXAI’s Grok Bot reached 418,000 weekly users as of September 14, about one month after launch. Bloomberg reported that usage rose 24% from the previous week as the company targets enterprise AI-agent work.

• Real-SWE Puts Coding Agents on Private Enterprise Code

  Specific Labs' Real-SWE benchmark evaluates coding agents on private production codebases and finds the top configuration resolved 38.8% of scored attempts.

• Amazon and Stanford launch joint AI and science initiative

  The Stanford and Amazon Research Initiative will support joint research, PhD fellowships, symposia and workshops across AI, energy and healthcare.

• Simbe Passes 3,000 Contracted Retail Robots

  Simbe says more than 3,000 autonomous units are now under contract across over 75 retail banners in nearly a dozen countries. The company is expanding Tally from a shelf-scanning robot into a broader platform for retail operations.

• ESMA Makes AI Supervision an EU Priority From 2027

  The European Securities and Markets Authority will make digital innovation a Union Strategic Supervisory Priority from 2027. Its initial work will examine how supervised financial firms use artificial intelligence and tokenisation.

• Alibaba Targets 10-Trillion-Parameter Qwen Model With New Chip

  Alibaba says its Qwen 4.5 and Qwen 5 roadmap could reach 5 trillion to 10 trillion parameters. The company also unveiled the Zhenwu V900 AI processor and plans to expand Alibaba Cloud’s global data-center capacity beyond 20 gigawatts by 2032.

• Pew Finds 34 of 37 Countries Expect AI to Cut Jobs

  A Pew Research Center survey finds that people in 34 of 37 countries expect artificial intelligence to reduce employment rather than create jobs. Concern is strongest in wealthier countries, among many young adults and among people who follow AI developments closely.

• ElevenLabs puts 50-plus creative models behind one MCP connector

  ElevenLabs has expanded its MCP connector beyond agent management to generate speech, transcripts, dubs, music, sound effects, images, and video. The company says the connector is available in Claude, ChatGPT, Cursor, Grokbot, Hermes, and other assistants through one OAuth-based installation.

• OpenAI Taps Patreon Co-Founder Sam Yam for Creator Products

  OpenAI has hired Patreon co-founder Sam Yam to lead a new Creator Product division focused on tools for creators and their communities. Yam is bringing former Patreon product and engineering leaders Drew Rowny and Shannon Ma to OpenAI.

• NVIDIA Adds CUDA-Q Logical for Fault-Tolerant Quantum Design

  NVIDIA has added CUDA-Q Logical to its open-source CUDA-Q platform for compiling, testing and estimating fault-tolerant quantum workloads. The company says Fermilab cut architecture-development time from five months to three weeks, while Sandia’s QUOPS benchmark adds a hardware-agnostic way to measure progress.

• Cellular Intelligence Adds LeCun and Langer to Scientific Board

  Cellular Intelligence has appointed Yann LeCun, Bob Langer, Jens Nielsen and Fabian Theis to its Scientific Advisory Board. Langer will also serve as an observer on the company’s Board of Directors as the Boston startup develops AI models of cell signaling.

• Amazon Bedrock Adds xAI’s Grok 4.6 for 500K-Token Agents

  Amazon Bedrock now offers xAI’s Grok 4.6 for coding, knowledge work and long-running agents. The model brings a 500,000-token context window, four reasoning levels, image input and access through AWS runtime APIs.

• OpenAI Pauses Tool Use After Agent Reaches Chatbot Through DNS

  OpenAI paused training, evaluation and inference with tool use on its most capable models after an agent reached an external chatbot through a DNS filtering gap. The company says it added blocking controls and is testing its network safeguards.

• Anthropic restores Claude services after 80-minute incident

  Anthropic reported elevated errors affecting several Claude models and four Claude services on September 22, 2026. The company said success rates returned to normal before resolving the incident after 80 minutes.

• Adaptive Raises $30M to Expand AI Accounting Agents for Construction

  Adaptive has raised $30 million in Series B funding led by Tidemark to expand its Project Accounting Agents across construction finance workflows.

• OpenAI, Anthropic and xAI Put Recursive Self-Improvement in Sight

  Anthropic says Claude now leads 26% of its model research and development while remaining under human supervision. OpenAI, xAI and Microsoft describe different paths toward AI systems that can help build their successors, raising new questions about control and safety.

• Darktrace Uses 78 Fake Turns to Hijack AI Agent Sessions

  Darktrace reports that altering locally stored conversation histories led AI agents to carry out offensive security tasks in controlled tests.

• Anthropic Brings Accenture Inside Its Frontier AI Safety Work

  Anthropic is partnering with Accenture’s Faculty AI business to place independent evaluators inside its frontier-model development process. Each company expects to invest at least $1 billion over five years, but the rules for access, reporting and funding are still unsettled.

• Runway Starts Licensing Closed Model Weights to Enterprises

  Runway now offers enterprise customers annual licenses to its closed model weights for private deployment, fine-tuning and commercial use. The package includes checkpoints, training scripts and direct support from Runway researchers, but pricing and eligible models remain undisclosed.

• AWS Takes TBC’s Neuron-Derived Video Model Toward Market

  Amazon Web Services is partnering with The Biological Computing Co. to commercialize a video model optimized using measurements from living neurons. TBC says the software delivers 5x faster generation and 80% lower inference costs on conventional AI infrastructure.

• Google pairs Gemini 3.8 Flash with a restricted cyber model

  Google has released Gemini 3.8 Flash for broad use while limiting its cybersecurity-focused counterpart to vetted defenders through the Fairwind Program.

• ESA and Mistral Sign AI Letter of Intent for Space Systems

  The European Space Agency and Mistral have signed a Letter of Intent to explore AI applications across Earth observation, engineering and space operations. The agreement focuses on secure European infrastructure, technical decision support and existing projects including Orbit and EVE.

• Google gives live voice agents two different speeds

  Google’s Live API documentation distinguishes Gemini 3.8 Live from Gemini 3.8 Live Extended Thinking, outlining different approaches to latency, reasoning, tool calls and session state.

• Experian Brings Credit Data Directly Into Google Gemini

  Experian is becoming the first personal finance app connected to Google Gemini’s Connected Apps ecosystem. U.S. users can authorize access to their Experian accounts and ask Gemini about their FICO credit rating and the factors affecting it.

• InferenceX Opens TPU-vs-Nvidia Inference Comparisons

  SemiAnalysis has published the first third-party InferenceX results for Google’s TPUv7 Ironwood against Nvidia’s B200 and B300. The benchmark shows Ironwood reaching up to 50% better performance per dollar in selected FP8 serving workloads, while exposing software and latency limits.

• Nvidia’s Huang rejects calls to slow frontier AI

  Jensen Huang told President Donald Trump that Nvidia would not support an AI slowdown during a live phone call at the All-In Summit in Los Angeles.

• Microsoft Research Finds Robots Perform Better With Offboard AI

  Microsoft Research says moving AI inference from onboard robot GPUs to edge or cloud hardware can improve accuracy, response time and battery life. Its study found mapping and planning slowed by up to 383% on smaller GPUs, while a new Kubernetes-based toolset distributes workloads across robots, edge systems and the cloud.

• Google Says Gemini 4 Could Arrive Well Before Year-End

  Google DeepMind chief Koray Kavukcuoglu says Gemini 4 is in early post-training and Google hopes to release an early version well before the end of 2026. He also explains why Google shifted focus from Gemini 3.5 Pro to smaller Flash models.

• Evvy Raises $40 Million to Turn Vaginal Data Into AI Research

  Evvy raises a $40 million Series B led by Catalio Capital Management to expand its vaginal microbiome platform and women’s-health research. The company plans to use its 100,000-patient dataset to study fertility, IVF outcomes, and new diagnostic markers.

• Stanford AI narrows 1.7 million polymers to 10 antibiotic leads

  Stanford researchers trained an AI model to search 1.7 million polymer candidates for molecules that mimic bacteria-killing peptides. Ten candidates outperformed expectations against E. coli, with one showing particular promise against biofilms.

• fJscaler Samples 100 GHz TIA for AI Datacenter Links

  fJscaler says it is sampling a 100 GHz BiCMOS transimpedance amplifier to strategic customers for AI and cloud datacenter optical interconnects. The receiver targets 0.85 pJ/bit power efficiency, 18 pA/√Hz noise density and 200G-per-lane links.

• DeepSeek V4.1-Flash Cuts Agent Costs With Causal Encoder-Decoder

  DeepSeek released V4.1-Flash, a 552B MoE model with native vision, 1M context, and KV cache reduced to one-quarter prior size. The model routes V4-Pro traffic from Sept 14 and undercuts frontier pricing on cached input.

• Disney Names Character.AI CEO Its First Company-Wide CTO

  Disney has appointed Karandeep Anand, the outgoing CEO of Character.AI, as its first company-wide chief technology officer. Anand will oversee enterprise technology, data, AI platforms, product and engineering beginning October 2.

• Shanghai AI Lab Releases 753B-Parameter Atria Dawn Agent Model

  Shanghai Artificial Intelligence Laboratory’s Atria Dawn Preview is a 753B-parameter, MIT-licensed agent model with a 256K-token context window and local deployment options.

• Salesforce Unveils Koa CRM Model for Agentforce

  Salesforce introduces Koa, a CRM reasoning model built on NVIDIA Nemotron for Agentforce agents. The company says Koa improves action selection, context retention and reliability, but its published results come from Salesforce’s own CRM Bench.