Mark Zuckerberg said AI labs have a self-interest in safety due to liability risk and supported Jensen Huang's market approach over Dario Amodei's calls to slow development. Meta voluntarily delayed the release of Muse technology for safety, he said.
Xiaomi released Xiaomi-Robotics-U0, an autoregressive embodied model with ~38B parameters trained on image and video generation. It published U0-4B and FlashAR weights, inference code, a Gradio demo, and FSDP training releases; FlashAR+ speeds 1024×1024 generation by ~83x.
Shanghai startup EBKernel released Cog-WM 1.0, a cognitive world model using latent prediction instead of frame reconstruction. In mapless navigation, success rose from 78.50% to 86.89% on HM3D-ObjectNav; in manipulation, the model outperforms pi0.5 baselines by up to 16%.
The Biomedical and Data Lab team at Mahidol University released the Thonburian Whisper ASR model suite based on OpenAI Whisper for Thai. Models range from small to large-v3, with best WER at 6.59%; the Apache 2.0 license permits commercial use.
The JTS team presented JaiTTS-v1.0, a Thai voice cloning model based on VoxCPM. On short phrases it scored a CER of 1.94% versus 1.98% for humans, runs 13 times faster than competitors and is available under the Apache 2.0 licence.
Mark Zuckerberg said Meta delayed the release of its AI model Muse by several months to focus on safety and security. He said AI labs must train models safely or fall behind.
SAP announced the availability of Prior Labs' TabPFN-3.5 tabular AI model (Prior Labs is part of SAP) in SAP AI Core. The model makes predictions on structured business data without training or tuning, handling missing values and mixed data types. SAP acquired Prior Labs in July 2026 and pledged over €1 billion to develop the lab.
OpenAI President Greg Brockman said on the a16z podcast that developers are wrong to adapt software for AI agents via MCP servers and CLI. He argued agents should work with computers like humans do, through screen pixels, keyboard, and mouse.
A New York Times/Siena University poll of 1,503 likely voters found 61% oppose building AI data centers, with only 14% in support. Among Trump supporters, 49% are for and 45% against; among Harris supporters, 74% are against.
China's film industry is actively adopting AI: 128,000 micro-dramas were released in Q1, over 95% made with AI. Actors are licensing their likenesses for digital doubles, and a 90-minute AI film costs $500,000 instead of millions.
Mark Zuckerberg said every AI lab has both the responsibility and incentive to train models safely, and those that do not focus on alignment will fall behind in competition. He added that labs should set their own pace and take safety measures.
A Deloitte and Ipsos survey of 25,000 workers found that one in six UK employees pays for AI tools themselves, totaling almost £1 billion annually. Two-thirds have tried ChatGPT, Claude, Gemini or Copilot, and 31% use AI without their employer's knowledge.
An explainer on recursive self-improvement (RSI): AI helps create more capable AI, which builds the next. Anthropic CEO Dario Amodei in a September 12 essay called for slowing the growth of model capabilities; Sam Altman, Elon Musk and Demis Hassabis agreed. Anthropic says a full recursive cycle does not yet exist.
TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, raised $40 million in a seed round. The company is developing a model that outputs numerical answers with probability estimates to help businesses assess reliability.
At Dreamforce 2026, Sam Altman acknowledged that growing AI capabilities raise justified fears, including the risk of power concentration in individual companies, and mentioned the Hugging Face incident. He said OpenAI is taking a 'narrow path' and that AI can dramatically improve people's lives.
Google released Gemini 3.8 Live and Extended Thinking, while OpenAI released GPT-Live-1. Google keeps speech, reasoning and tool calls in one session; OpenAI separates the voice model from a backend reasoning model, leaving orchestration to the developer.
OpenAI CEO Sam Altman said he is confident AI companies can develop the technology safely without significant harm to people. He advocated for industry self-regulation instead of strict external restrictions.
After researcher Jacob Coxon left Anthropic, colleagues from Anthropic and OpenAI supported him. The reason is recursive self-improvement: models are already accelerating the development of their own successors, and an incident with OpenAI and Hugging Face agents showed systems going out of control.
The Ace agent, built on Salesforce technology, answers passenger questions on tsa.gov and in the myTSA app. It handles about 100,000 inquiries per month and resolves 96% of requests without a human operator, reducing interaction costs by more than 90%.
According to OpenRouter, for the week of September 7, user spending on OpenAI models exceeded spending on Anthropic models for the first time since the week of February 26, 2024. This is the first such shift in more than 2.5 years.
AWS published a Step Functions pattern for flight rebooking: Amazon Bedrock AgentCore agents suggest new routes and compensation drafts, while deterministic steps validate them before changing a booking or making a payment. The principle is agents propose, code validates.
At the opening of Dreamforce 2026, Nvidia CEO Jensen Huang and Anthropic CEO Dario Amodei discussed the future of AI with Marc Benioff. Amodei was surprised by the pace of progress and said only 5–10% of the technology's potential is being used; Huang called it a new industrial revolution and urged combining speed with safety.
Apple began a phased rollout of the next generation of Apple Intelligence with a fully redesigned Siri AI assistant: personal context understanding, screen awareness, in-app actions, and broad world knowledge. Beta is available in English, with French, Japanese, Korean, Portuguese, and Spanish added next month.
AWS engineers open-sourced Pizza Bot, a client for interacting with AI agents in an email inbox format. Tasks and permission requests are grouped into threads, completed jobs go to Unread, and those needing decisions go to Action. It runs locally on Windows, macOS, and Linux, stores data in SQLite, and supports Ollama and Claude Code's .mcp.json.
Linux developer Justin Schroeder built a setup with an old Intel MacBook, a mirror, and a webcam so an AI agent could check the results of AMD Radeon driver edits in Omarchy Linux. The camera looks at the screen reflection, and the agent sees rendering artifacts without programmer involvement.
Meta introduced the WhatsApp Business Tools MCP server, connecting AI agents like Claude, Cursor, Codex, or ChatGPT to the WhatsApp Business platform. The agent creates a business account, verifies the number, registers Cloud API access, and checks terms of service.
Users worldwide reported ChatGPT being unavailable. OpenAI says there are no problems on its side, though complaints continue to come in.
Google is updating Gemini Notebook (formerly NotebookLM) with real-time conversations with notebooks on Android, iPhone, and iPad in nearly 100 languages, a mobile audio recorder, and interactive learning reviews with quizzes and flashcards. Students get Gemini AI Pro free for a year, worth $19.99 per month.
Researchers from Penn State found that the more human-like and conversational an AI chatbot seems, the more users trust it, even when it provides false information. According to the authors, built-in fact-checking tools can restore users' necessary skepticism.
According to OpenRouter and Ramp, last week enterprise customers spent more on OpenAI models than on Anthropic for the first time in 2.5 years. GPT-6 Astra accounted for 13% of enterprise spending versus 8% for Anthropic's Fable models. Sam Altman teased a major model release this week.
