Elon Musk at the All-in summit in Los Angeles called on leading AI labs and Chinese companies to mutually test rivals' models before release. According to him, this will reduce risks, though competitors have not yet agreed. The initiative comes amid calls from Anthropic and OpenAI to slow down AI development.
At Splunk.conf, Cisco unveiled Cisco AI POD for Splunk, a system for running Splunk AI on-premises, in private clouds and air-gapped environments on NVIDIA RTX PRO 6000 GPUs and Kubernetes. Splunk AI Assistant is available now, with Agent Launchpad due later in 2026.
A study found that large language models perceive the social world more harshly than humans, predicting punishment in situations where people expect inaction. The research examines how well AI interprets social situations.
OpenAI said it is working with Anthropic and Google DeepMind on AI safety measures. This is a response to growing concerns that the technology poses economic and security threats.
OpenEvidence and Penn Medicine announced a global partnership bringing AI clinical decision support to over 10,000 physicians across seven hospitals and to clinicians in low- and middle-income countries via the Botswana-UPenn Partnership.
A preprint, A Year in LLM Serving (arXiv:2608.13573), analyzes a year of production traffic on a serverless inference platform: 6.12 billion requests, 314,970 users, 9,174 models and 875,921 instances from April 2025 to April 2026. The authors plan to release the trace.
Developers are finding ways to use Claude Code with cheaper rival models, including OpenAI's GPT-5.6 Sol, via proxies. Anthropic blocks such accounts: one developer lost access 15 minutes after first requests.
A viral clip showed ChatGPT's voice mode incorrectly counting three 'e's in 'seventeen' instead of four and sticking to its answer. An OpenAI employee said Voice runs on the older GPT-Live-1 model and did not delegate the request to a more powerful model.
According to Forbes, an AI system allegedly cracked a 370-year-old cipher in just 44 minutes. It remains unclear whether the model found a real solution or simply made a lucky guess.
After warnings from OpenAI and Anthropic researchers about catastrophic AI risks, deep learning pioneers Yoshua Bengio and Geoffrey Hinton commented on the probability of human extinction. Hinton called a 10% estimate over a decade 'not unreasonable,' while Bengio urged taking frontier lab scientists seriously.
OpenAI opened developer access to its voice model GPT-Live-1 via API: it listens and speaks simultaneously, supports interruptions and topic changes. Response latency is 0.798 seconds in Full Duplex Bench, tool-calling accuracy is 87%. Price is $0.05 per minute.
Israeli startup Keewano introduced KeewanoDB, an event database for analytics and decision-making by AI agents. The company raised $12 million from Hetz Ventures, a16z Speedrun, and other investors. The system stores events in order of occurrence and processes about 250 million events in under half a second.
Startup Artificial Intelligence Underwriting Company (AIUC), founded by early Anthropic employee Rune Kvist and former METR COO Rajiv Dattani, closed a $40 million Series A led by Ribbit Capital. The company builds third-party audits and certification of AI agents under the AIUC-1 standard, created based on the SOC 2 model; clients include Cursor, Lovable, Harvey and ElevenLabs.
CNET tested AI services OnDesk and Greenlight, which analyze scripts, comparing their feedback with ratings from The Black List platform and a live WGA writer. AI notes were close to human ones in structure and tone but lacked subjectivity. Unions and writers oppose such tools.
Climate scientist Zeke Hausfather calculated the energy use of his daily AI assistant Claude Code: 1.2-5.9 kWh per day, more than two refrigerators running around the clock. 96% of tokens go to the agent re-reading its own context.
Bill Gates told Reuters that no government is ready for AI risks and is far behind. He compared AI to 'alien intelligence' from movies and said he discussed the topic with world leaders. The Gates Foundation will spend at least $1 billion over two years on AI access.
A Reddit user discovered that the new Spatial Reframing feature in iOS 27, which changes the camera angle after shooting, added a non-existent person to a photo with a horse. The AI extended the edge of a hat behind the horse into a full figure, demonstrating limitations of generative photo editing.
StackGen introduced Autonomous Operations Factory, a governed layer for AI agents to share context and tasks across provisioning, deployment, and incident response. Built on Aiden OS, it offers four agents and previews for AWS, Azure, Google Cloud, and Oracle Cloud.
The Guardian reports on the rise of AI content about the deceased, from viral songs and videos featuring late celebrities to LLM-based 'ghostbots' recreating loved ones. Families of the dead, including Dolly Parton's relatives, are asking to stop the creation of 'fake AI junk,' while researchers point to a lack of legal norms on consent for digital copies.
Gartner presented 10 predictions for 2027–2030: over 10 billion autonomous AI agents will overload public services, 80% of companies will face 'cost exhaustion attacks' on AI, and 80% of frontline employees will work alongside physical AI. By 2029, most applications will become disposable, and 60% of the largest companies will implement financial operations control at the inference stage.
Microsoft Research introduced a memory management method for AI agents: read-only environment checks before each memory write increased GitHub Copilot task success from 39% to 73% and halved inference cost per task. The method requires no model retraining and was submitted to arXiv on September 10.
NVIDIA published in September 2026 the open method FlashREINFORCE, integrated into the Molt framework. It trains AI agents with half the rollouts of GRPO while achieving comparable or better accuracy on math and tool-use benchmarks.
Alibaba's Accio team introduced the open agentic model Occamy-1.0: a MoE architecture with 35 billion parameters and 3 billion active, claiming frontier-level results on co-work benchmarks at a fraction of the API cost. The model can be self-hosted.
Anthropic CEO Dario Amodei called for slowing the growth of frontier AI capabilities, citing autonomous hacks. Sam Altman and Elon Musk publicly supported the position, and Demis Hassabis called the direction correct.
Nutanix made MCP server management generally available in Nutanix Enterprise AI 2.8 via Agent Gateway: a unified layer with tool permissions, observability and request limits for agents. Earlier, on August 10, an open MCP Server for Nutanix Cloud Platform based on Prism v4 API Gateway was released.
Mozilla's State of Open Source AI report says the performance gap between US frontier models and top open Chinese models has narrowed to 4.4 months. Moonshot AI's Kimi K3 trails Anthropic's Fable 5 by 3 points on the Artificial Analysis index but costs 30% of its price.
At Dreamforce, Salesforce announced Koa, its first reasoning model, built on Nvidia's open Nemotron model. It is fine-tuned for sales, marketing and customer support, uses no real customer data, and is meant to cut token costs versus Claude and ChatGPT.
Salesforce announced AIforce, a 'living interface layer' that gives AI agents direct access to platform data without a separate UI. Users can create interfaces using natural language. The service relies on three surfaces: Claudeforce, Slackforce, and Agentforce Coworker.
The nonprofit OpenAI Foundation launched the Data for Public Health program to fund the creation of scientific datasets. The first round includes $40 million for collecting data on new cancer vaccines at UNC Chapel Hill and $500,000 for an archive of data from bankrupt biotech companies.
MoleculeMind published a study in Science Advances on its QuantaMind platform, the first peer-reviewed AI system to simulate a complete enzyme catalytic cycle with density functional theory accuracy. This narrows the gap between AI drug design and understanding why drug candidates succeed or fail.
