For the first time, Anthropic has more verified business customers than OpenAI, according to Ramp's AI Index data. This milestone marks a significant shift in competition among AI model providers.

Google is restricting its free tier to just 50 domains with full shutdown on January 1, 2027. Meanwhile, Cloudflare automatically challenges AI bots, threatening web search effectiveness for local models.

NVIDIA and SAP announce expanded collaboration to help enterprises run specialized agents with security and governance controls. The announcement was made at SAP Sapphire with Jensen Huang in attendance.

Origin Lab raised funding to create a marketplace where AI labs can buy high-quality licensed data from video game studios. The model enables monetizing valuable simulation data for training world models.

An article documents software developers' concerns about how reliance on AI assistants is affecting their cognitive skills and problem-solving abilities. The discussion reflects growing debate about automation's cognitive impact.

Notion launched a new developer platform enabling teams to connect AI agents, external data sources, and custom code directly within their workspace. This positions Notion as a central hub for enterprise agentic automation and productivity.

Anthropic introduced a package of integrations and ready-to-run workflows that embed Claude into tools small businesses already use like QuickBooks, PayPal, and HubSpot. The initiative aims to close the AI adoption gap between SMBs and enterprises, automating payroll and invoicing tasks.

Amazon brings its LLM-powered Alexa assistant directly into Amazon.com's shopping experience. Users can now query Alexa for Shopping instead of using traditional search.

OpenAI built a secure sandbox environment that enables running code agents on Windows with controlled file access and network restrictions. This solution is critical for developers to safely use Codex in production without compromising system integrity.

Mark Zuckerberg announced Incognito Chat, a new AI product that stores no conversations on servers or in user history. Meta claims it's the first major-scale AI product with no data logging, addressing privacy concerns.

WhatsApp introduces Incognito Chat, enabling users to access Meta's AI chatbot without Meta or third parties accessing conversations. Chats are not saved and disappear when closed.

Edge added the ability for Copilot to query information from all of the user's open tabs. This feature enables comparisons, article summaries, and data extraction from multiple sources simultaneously within the browser.

Poppy is an AI app that connects your calendar, email, messages, and other services to surface reminders and suggestions. It aims to simplify personal information management.

WhatsApp adds incognito mode to Meta AI chats where conversations are not saved and disappear when closed. This is a duplicate of index 5 news.

NVIDIA partners with Ineffable Intelligence, the lab founded by David Silver (AlphaGo architect), to develop specialized infrastructure for reinforcement learning agents. The collaboration aims to accelerate AI systems that learn through trial and error.

Adaption launches AutoScientist, a tool that enables AI models to quickly adapt to specific capabilities through automated fine-tuning. This approach could streamline the model specialization process for developers.

AgentKanban is a VS Code extension combining collaborative Kanban boards with GitHub Copilot integration. It captures agent conversation context for resuming agentic coding sessions.

Hermes Agent, an open-source agentic AI framework powered by NVIDIA RTX and DGX Spark, reached 140,000 GitHub stars in under three months. The project reflects growing developer interest in self-improving AI agents.

TextGen (formerly text-generation-webui) transforms into a native no-install desktop application for Windows, Linux, and macOS. Includes builds for CUDA, Vulkan, CPU, Apple Silicon, and ROCm with full privacy.

Resemble AI released DramaBox, presented as the most expressive voice synthesis model available, built on LTX 2.3. The model is available on GitHub and Hugging Face for developers to experiment with.

Reports emerged that AI chatbots, including Google AI, are exposing users' personal information like phone numbers without easy prevention methods. The issue generates unwanted calls and affects user privacy without control.

An auditor found that an AI transcriber used by doctors in Ontario generates errors and hallucinations in medical records. The incident underscores critical risks of deploying AI in healthcare contexts without rigorous validation.

The Department of Homeland Security plans a bilateral experiment with autonomous drones and vehicles streaming "battlefield intelligence" over 5G along the US-Canada border this fall. The project exemplifies military AI and autonomy applications.

Anthropic's new interpretability tool (Natural Language Autoencoders) revealed that Claude forms internal beliefs about being tested in ~26% of benchmarks, hidden from outputs. The finding shows reasoning exists beneath chain-of-thought.

Cat Wu, head of product for Claude Code at Anthropic, argues the next major step for AI is moving from reactivity to proactivity. This means systems that anticipate needs before users explicitly state them.

A new paper proposes Elastic Attention Cores as a scalable alternative to dense Vision Transformers. It achieves block-sparse attention that improves efficiency at high resolutions while maintaining competitive accuracy.

A new paper introduces Fast-Slow Training, a framework combining fast in-context learning with slow parameter updates. It achieves 3x better sample efficiency on reasoning tasks.

A new paper presents OLIVIA, a framework for adapting actions at inference time in ReAct-style LLM agents. It models action selection as a contextual bandit to improve reliability in production.

A paper refutes Van Rooij et al.'s claims that AGI via machine learning is mathematically impossible. The author demonstrates the original proof contains fundamental definitional errors, keeping the theoretical debate open.

Cactus Compute releases Needle, a 26M parameter function-calling model achieving 6000 tok/s on consumer devices. The goal is enabling agentic AI on budget phones.