Google has announced Gemini Robotics 1.5 and Gemini Robotics-ER 1.5, enabling robots to perceive environments and plan actions. The former handles vision-language-action, while the latter manages embodied reasoning to support complex multi-step tasks.
The macOS app Dayflow uses a local VLM to read screen content and automatically generate semantic timelines, such as recording PR reviews for 45 minutes instead of simply Chrome for 3 hours.
Security firm Veria Labs has reported a vulnerability where flaws in the authentication flow allow remote code execution when connecting to Claude Code or Gemini CLI via MCP servers.
OpenAI and NVIDIA have signed a letter of intent for a strategic partnership to deploy at least 10GW of NVIDIA systems as next-generation AI infrastructure.
On September 20, 2025, xAI announced Grok 4 Fast, a new model specialized for cost efficiency. It reduces average inference tokens by 40% and achieves benchmark performance equivalent to Grok 4 at a 98% lower price.
xAI has announced Grok 4 Fast, a cost-efficient reasoning model. The company states that it achieves performance equivalent to Grok 4 using 40% fewer thought tokens, reducing the cost for the same performance level by 98%.
Google has announced the early release of a preview version of Gemini in Chrome for Google AI Pro and Ultra subscribers in the US, enabling summaries and Q&A based on the context of open tabs.
The Cactus team has released Cactus, an AI inference engine designed for smartphones. Using a Qwen3-600m INT8 model, it achieves speeds of 50 to 70 tokens per second on the latest devices.
Demis Hassabis of Google DeepMind stated that the ability to "learn how to learn" will be the most essential skill for the next generation to adapt to rapid AI-driven changes.
Google DeepMind CEO Demis Hassabis spoke at an ancient Roman theater in Athens on September 14, stating that AGI could arrive within a decade and that the most essential skill for the next generation is "learning how to learn."
An investigation by The Guardian has revealed that contract workers engaged in evaluating Google's AI models are facing tight deadlines, low compensation, and a lack of transparency regarding their duties.
Simon Willison praised Google's new AI Mode for providing quality similar to GPT-5's search capabilities and fast response times, though he expressed frustration over the lack of transparency regarding executed search queries.
Anthropic has reportedly agreed to pay $1.5 billion to settle a copyright infringement lawsuit brought by a group of book authors.
A scientific team has announced the successful application of an AI method called Deep Loop Shaping to the control system of the LIGO gravitational wave observatory, reducing feedback loop noise by 30 to 100 times.
The tool gptvis, developed by rti, enables step-by-step tracking of attention mechanisms and information transformation in a small Transformer with two layers and 20-dimensional embeddings.
Siemens Digital Industries Software has launched PADS Pro Essentials and Xpedition Standard, bringing AI-driven automation and collaborative PCB design tools to the SMB market.
OpenAI announced usage resets after a September 25 outage, but recurring limit fluctuations and pricing shifts reveal a deeper struggle to scale. The company is balancing explosive user growth against compute costs through aggressive tiering and controversial limit models.
OpenAI has launched ChatGPT Images 2.5, introducing the Sketch tool for doodle-based generation, improved editing precision, and up to 50% faster processing speeds.
The AI model Union Alpha, provided by an anonymous source, is available for free for a limited time on platforms such as OpenCode. It is said to be built for agentic workflows, featuring a context length of approximately 260,000 tokens and support for image input. While claiming performance close to Fable 5.1 and Astra on Terminal-Bench 4.0, early user reports on speed and stability are mixed.