Google DeepMind has introduced SynthID Bio, a family of watermarking methods that embeds imperceptible signatures into protein sequences and 3D structures to enhance biosecurity and scientific integrity.
Strata introduces a self-service analytics platform that provides a governed semantic layer and dashboards, designed to ensure AI agents receive accurate, governed data.
Google's pilot program, involving approximately 100 publishers, compensates sites based on content contributions to AI Overviews and Gemini, though reports indicate significant disparities in payouts and dissatisfaction among participants.
As AI models learn from human text, a reciprocal relationship has emerged where humans begin to edit their own writing to avoid sounding like machines.
Instagram is adding an AI-powered "creative assistant" to its Edits app, offering creators personalized advice based on account metrics to optimize engagement.
Cerebras Systems CEO Andrew Feldman will address the growing demands for compute, energy, and infrastructure, and how the company's wafer-scale computing approach addresses these constraints.
Restate, the startup building durable execution infrastructure, has raised $20 million in a Series A round to scale its services for the growing AI agent ecosystem.
While formal methods like TLA+ are vital for designing concurrent systems, they cannot express certain properties, such as reachability or hyperproperties, without complex workarounds.
Google, Meta, OpenAI, Anthropic, xAI, and Nvidia executives have signed the Joint Commitment On Frontier Responsibilities, an agreement for voluntary AI safety oversight.
Airbnb is rolling out an AI search function that uses text and voice prompts alongside new social features that allow travelers to see trips from their connections.
The DrivingBench project tested whether general-purpose AI models can navigate real-world physical environments using a Toyota Corolla, uncovering issues with safety refusals and response latency.
Nomura Real Estate Holdings has implemented Sansan's invoice management service "Bill One" across 10 group companies, digitizing approximately 320,000 invoice tasks annually. The Treasury Department, which oversees payment processing, reported that this has led to the creation of over 9,000 hours of work capacity per year.
OpenClaw v2026.9.7 introduces support for the OpenAI Agents API, allowing agents to run in both OpenAI-hosted and self-hosted environments with features like streamed replies and live web search. This release also enhances update safety by implementing mandatory database backups before migrations and provides several bug fixes for Windows, macOS, and various integration channels.
Tesla has opened $30 billion in credit facilities from Citi and Wells Fargo to support operating activities and upcoming capital expenditures as profitability has dropped in recent years.
PSSA is a small language model that utilizes a recurrent state-space layer and an episodic memory bank to achieve faster training and inference than standard Transformer architectures.
According to the "Survey on AI Utilization in Development (2026)," 65.0% of respondents from large enterprises reported that development is being carried out by non-engineers. Meanwhile, the findings highlight issues such as the failure to verify generated code and an inability to grasp occurring troubles.
Following a survey of over 600 respondents, the mathematical community has issued recommendations for AI labs to ensure the transparency and human understandability of AI-generated mathematical results.
Google analyzed thousands of entries from its AI Agents Challenge to identify successful architectural patterns for multi-agent systems, highlighting key strategies for agent interoperability and tool use.
Sam Altman stated that OpenAI will not go public until it can make reliable safety claims about its increasingly capable models, avoiding "additional pressure" from Wall Street.
Apple researchers find that efficient LLM steering methods often compromise generation fluency and note that instruction-tuned models are less responsive to activation steering than base models.