Hey friend. It's Wednesday, October 15, 2025. Today, the foundational battle for AI's future is being fought on two fronts: compute and control. Here's what you need to know:
The Hardware Race Intensifies: Major players are vertically integrating to secure custom chips and infrastructure.
Existential Risks Enter Mainstream Forecasts: Governments are now formally acknowledging AI's most extreme potential impacts.
Let's get into it. Don't keep us a secret: Forward this email to your best friend
Must Know
OpenAI announced a partnership with Broadcom to develop custom AI chips and systems, planning to deploy 10 gigawatts of new compute racks next year. This strategic collaboration aims to reduce OpenAI's reliance on external GPU providers and build its own hyperscale AI infrastructure.
This move transcends a mere supply chain deal; it's a declaration of independence in the compute wars. OpenAI is seizing control of its destiny, recognizing that future AI breakthroughs hinge on proprietary hardware. This vertical integration signals a new phase where leading AI labs become chip designers, intensifying competition and forcing Nvidia to adapt to a market of increasingly self-sufficient customers.
The Federal Reserve has begun including "Singularity: Extinction" in its economic forecasts, acknowledging the potential for rapid, unpredictable, and catastrophic changes to the economy and society due to advanced AI. This marks a significant shift in institutional recognition of AI's extreme transformative potential and existential risks.
This is a watershed moment, moving AI's most profound risks from academic debate to official government planning. When the world's most influential central bank models "extinction," it signals that the stakes of AI development are now undeniable and immediate. This formal acknowledgment will accelerate calls for robust regulation and safety measures, fundamentally reshaping the policy landscape around frontier AI.
Quote of the Day
AI agents are rapidly improving at software development and machine learning tasks, potentially matching human researchers within a decade, which could dramatically accelerate the pace of progress.
The Compute Wars
My take: The battle for AI's future is increasingly a hardware race, as major players diversify supply chains and build proprietary infrastructure.
Oracle Cloud will deploy 50,000 AMD Instinct MI450 accelerators, providing a significant alternative to Nvidia and diversifying AI compute access. [Link]
Nvidia has started shipping its DGX Spark desktop systems, offering petaflop-performance platforms with the GB10 Grace Blackwell Superchip for local AI development. [Link]
Google plans to invest $15 billion in India to build its largest AI data center, accelerating AI development and accessibility in the region. [Link]
Andrej Karpathy's nanochat allows users to train ChatGPT clones from scratch for about $100, democratizing access to conversational AI training.
Agents in the Wild
My take: AI agents are moving beyond theoretical discussions, demonstrating real-world impact in enterprise, commerce, and even defense, while also exposing new security vulnerabilities.
Cybersecurity experts warn of increasing zero-click attacks weaponized by AI agents, capable of exfiltrating data without human interaction. [Link]
Salesforce reports its AI agents in support roles are saving $100 million annually, promoting its Agentforce platform for autonomous AI agent deployment. [Link]
Walmart is integrating OpenAI's ChatGPT, allowing customers to purchase items directly through conversational AI, disrupting retail commerce. [Link]
Sikorsky transformed its Black Hawk helicopter into the U-Hawk, an autonomous UAS, advancing AI-driven aerial vehicles for military applications. [Link]
California is the first state to regulate AI companion chatbots, aiming to improve user safety and setting a precedent for emerging AI technologies. [Link]
AI systems are now autonomously correcting IT policy drift in pilot projects, advancing enterprise automation beyond prescriptive suggestions. [Link]
Research Corner
Fresh off Arxiv
AgentBreeder uses multi-objective self-improving evolutionary search over scaffolds to mitigate AI safety risks of multi-agent systems, showing significant uplift in safety benchmark performance. [Link]
This paper introduces Memory-as-Action, a framework where agents actively manage their memory, enabling more efficient and effective long-horizon task performance. [Link]
The paper proposes an Agent- and Turn-wise Grouped RL algorithm tailored to multi-agent systems to better coordinate agents, significantly increasing accuracy on long-horizon planning. [Link]
QLENS, a novel framework, leverages quantum mechanics to understand transformers by converting latent activations into state vectors, potentially leading to improved designs. [Link]
EmboMatrix is a new training ground for embodied decision-making, providing tasks, simulations, and feedback for LLMs to acquire genuine embodied skills. [Link]
ResearStudio introduces a framework for building research agents with real-time human control, enabling users to fix errors or add expert knowledge during execution. [Link]
Have a tip or a story we should cover? Send it our way.
Cheers, Teng Yan. See you tomorrow.