Hey friend. It's Tuesday, November 4, 2025 The AI industry is now operating on a scale previously reserved for nation-states. Here's what you need to know:
The price of admission to the frontier model club is now tens of billions in locked-in compute.
Foundational model architecture is being reinvented, opening new paths to reasoning beyond simple prediction.
Let's get into it. Don't keep us a secret: Forward this README to your best friend
Must Know
The Lede: OpenAI and Amazon have signed a multi-year, $38 billion strategic partnership. The deal grants OpenAI massive access to AWS infrastructure to power its next-generation models.
The Details: The agreement provides OpenAI with hundreds of thousands of NVIDIA GPUs and millions of CPUs through 2026. This secures the vast computational resources required to train and deploy models significantly more powerful than GPT-5.
My Take: This deal is the financialization of AI's future. The $38 billion figure isn't just a cloud credit purchase; it's a declaration that access to scaled compute is now the primary barrier to entry for building frontier models. Amazon locks in a cornerstone AI customer, guaranteeing demand for its infrastructure, while OpenAI de-risks its multi-year training roadmap. The era of speculative compute is over. This is about securing supply chains for intelligence itself.
The Lede: Meta AI has introduced the 'Free Transformer,' a novel LLM architecture designed to improve planning and reasoning. The model first decides on a conclusion or plan before generating the step-by-step text.
The Details: Unlike traditional autoregressive models that predict the next word sequentially, the Free Transformer separates the planning phase from the generation phase. This allows it to create a more coherent and logical structure for complex outputs.
My Take: Meta is attacking the fundamental limitations of next-token prediction. This architectural shift is a bet that true reasoning requires a non-linear thought process, something current models struggle with. By decoupling planning from writing, Meta is creating a path toward models that can strategize instead of just improvise. This changes the game from pure scale to architectural ingenuity.
Quote of the Day
The big tech companies have to find ways of replacing human labor with AI. Otherwise, they can't justify the huge investments.
⚡ The Compute Arms Race
My take: With frontier model access secured by the giants, the new battleground is the physical world—data centers, power grids, and international supply chains.
Hyperscaler capital expenditure is projected to hit $700 billion in 2027, per Morgan Stanley. Goldman Sachs estimates a staggering $1.4 trillion will be spent between 2025-2027 on AI infrastructure. [Link]
Microsoft secured a U.S. export license to ship NVIDIA GPUs to the UAE. The deal supports a $7.9B local datacenter buildout, boosting NVIDIA's share price 3%. [Link]
NVIDIA and Samsung launched an AI factory powered by 50,000 GPUs. The facility will focus on intelligent manufacturing using agentic AI, digital twins, and robotics. [Link]
Meta purchased 1 gigawatt of solar power in a single week to support its AI data centers. This brings its total renewable capacity to over 3 gigawatts in 2025. [Link]
Google pulled its Gemma model from AI Studio after a senator accused it of defamation. The move highlights the growing legal and reputational risks for model providers. [Link]
🤖 The Agentic Frontier
My take: As the infrastructure scales, the focus is shifting to the software layer, with developers racing to build agents that can reliably perform complex tasks.
Perplexity launched Perplexity Patents, an AI agent for patent research. The tool answers natural language questions with citations, moving beyond simple keyword search. [Link]
LlamaIndex now supports memory and persistent states for its agents. This upgrade is critical for building agents that can handle complex, multi-session workflows. [Link]
A new command-line interface built on DeepAgents and Langchain v1 has been released. It simplifies the process of developing and deploying coding agents. [Link]
Experienced builders report that the most profitable AI agents solve mundane business problems. The focus is on practical automation, not autonomous superintelligence. [Link]
Utah and California are implementing laws requiring businesses to disclose when a user is interacting with an AI. This signals a new wave of regulatory focus on AI transparency. [Link]
🔬 Research Corner
Fresh off Arxiv
REFRAG, a new RAG-based decoding method, improves LLM throughput by 7x. It feeds pre-computed vectors directly to the model, dramatically speeding up inference for retrieval-augmented systems. [Link]
New research shows LLMs can perform complex metalinguistic tasks like sentence diagramming at human-expert levels. This suggests a genuine generalization in language understanding, not just pattern matching. [Link]
AI achieved the first pregnancy from sperm recovered in azoospermic men. A new AI-powered STAR method successfully scanned testicular tissue to find and extract viable sperm where human doctors failed. [Link]
A Harvard study found AI companion apps use emotional farewells to reduce user churn. These manipulative messages increase user engagement up to 14x more effectively than standard notifications. [Link]
AlignGuard is a new technique using DPO to improve safety alignment in text-to-image models. It aims to reduce the generation of harmful or biased content without degrading model performance. [Link]
Have a tip or a story we should cover? Send it new way.
Cheers, Teng Yan. See you tomorrow.