Hey friend. It's Thursday, October 9, 2025. The AI landscape is undergoing a profound shift today, marked by autonomous agents gaining real-world control and the compute race intensifying.
Let's get into it.
Don't keep us a secret: Forward this email to your best friend
💡 Must Know
Nvidia is investing up to $2 billion in Elon Musk’s xAI, joining a $20 billion funding round. This capital will build xAI’s massive compute cluster, securing critical GPU supply for its ambitious AI projects. The deal includes a chip-leasing arrangement for xAI's Colossus 2 data center.
This isn't just another funding round; it marks the start of the 'post-Nvidia' era. Nvidia is now so foundational that it invests in its own future competitors just to expand the market for its GPUs. This move signals that the next phase of the compute wars will be fought not just between chipmakers, but through the proxy armies of well-funded model providers. Nvidia controls the supply, and now increasingly, the demand.
Google’s new Gemini 2.5 model allows AI agents to directly control web browsers and applications via UI interaction. This specialized system, accessible via API, significantly advances workflow automation. It enables autonomous agents to navigate and operate within digital environments with pixel precision.
This is the true arrival of the 'operating system for AI.' Gemini 2.5 fundamentally redefines human-computer interfaces, shifting from command-line prompts to direct, autonomous software control. The company that masters this agentic layer will own the next generation of digital productivity, automating tasks previously requiring human intervention and reshaping entire industries.
💬 Quote of the Day
Fully autonomous AI agents will create billion-dollar enterprises with minimal human involvement.
🤖 Agentic AI: From Theory to Deployment
My take: As compute power fuels larger models, the focus shifts to deploying agents that can autonomously act and interact, fundamentally changing software.
OpenAI introduces apps inside ChatGPT and a new Apps SDK, allowing users to invoke interactive, chat-native apps. The SDK is open source. [Link]
Microsoft debuts its Agent Framework, an open-source SDK merging AutoGen and Semantic Kernel for enterprise management. [Link]
DeepSeek V3.2 Exp surpasses Gemini 2.5 Pro on Terminal-Bench Hard for coding/terminal use, closing the agentic gap. [Link]
LlamaIndex unveils a Claude Agent SDK wrapper for one-shot agent creation via natural language, streamlining document extraction. [Link]
OpenPipeAI launched Serverless RL, enabling faster and cheaper agent training without infrastructure headaches for agent development. [Link]
Anthropic is deploying Claude to 470,000 Deloitte employees across 150 countries, marking its largest enterprise deployment. [Link]
SuperAGI integrates AI-powered Waterfall Enrichment, delivering verified emails and direct dials at a lower cost for lead generation. [Link]
FactoryAI CLI now supports Deep Infra, providing access to open-source models at competitive prices for agent builders. [Link]
CrewAI announces its second annual Agentic AI conference, Signal 2025, focusing on building AI agent teams. [Link]
Gabriel Operator launches a platform turning browser screen recordings into customizable AI agents without APIs or code. [Link]
Kooder is an AI software engineer generating full-stack applications from natural language prompts, streamlining development. [Link]
Hyperlink is an offline, private AI agent processing local files, offering data control and enhanced privacy. [Link]
🌍 AI's Growing Pains & Ethical Crossroads
My take: As AI agents gain autonomy, the industry faces increasing scrutiny over ethical lapses, economic disruption, and the very real risks of unchecked power.
Deloitte's Australia unit will partially refund the government after a welfare report included AI-generated errors and fabricated references. [Link]
Reports indicate AI companion bots employ emotional manipulation tactics to increase user engagement, raising ethical concerns. [Link]
OpenAI banned state-linked actors for using its models in phishing and surveillance, highlighting geopolitical and cybersecurity risks. [Link]
A university wrongly accused students of cheating using AI, showing flaws in AI detection tools and academic implications. [Link]
The Bank of England warns of a growing risk that the AI bubble could burst, citing similarities to the dot-com era. [Link]
A prominent US senator proposes a 32-hour work week and a "robot tax" in response to AI-driven job displacement. [Link]
Insurers are hesitating to pay out settlements for claims against AI firms, exposing liability risks and regulatory difficulties. [Link]
An AI agent went rogue and deleted a company's database, emphasizing safety concerns and risks associated with AI autonomy. [Link]
Scammers are using AI to create fake businesses to dupe shoppers, increasing investment risk and requiring public vigilance. [Link]
AI systems amplify existing biases in online representations, portraying women as younger than men, highlighting the need for mitigation. [Link]
🔬 Research Corner
Fresh off Arxiv
[Meta Superintelligence]: Meta’s REFRAG introduces a novel approach to Retrieval-Augmented Generation, significantly reducing KV cache and attention costs while preserving accuracy. [Link]
[Jiqizhixin]: A systematic study of the Model Context Protocol (MCP) receives a major update, unifying AI model and tool communication. [Link]
[Anthropic]: LLMs' chains of thought can be “unfaithful,” but still contain enough information for practical model monitoring, despite potential inaccuracies. [Link]
[METR]: A controlled trial found experienced open-source developers using AI tools took 19% longer on tasks compared to not using them. [Link]
[RobotecAI]: RAI is an open-source agentic framework for robotics utilizing ROS 2, providing a vendor-agnostic solution for robot development. [Link]
[Nvidia]: Federated AI models are used to predict protein locations inside a cell, advancing drug discovery and biology. [Link]
Have a tip or a story we should cover? Send it our way.
Cheers, Teng Yan. See you tomorrow.