🔮 The Classified Frontier
In May 2025 OpenAI physically delivered the ChatGPT o3 model weights to Los Alamos National Laboratory in locked briefcases, because the Venado supercomputer runs on a classified, air‑gapped network. The post stresses that AI models are tangible collections of numbers stored on hardware, not ethereal cloud services. While creating frontier‑level weights requires billions of dollars and massive compute, copying or using those weights is low‑viscosity, allowing rapid spread via APIs or simple file transfers. This asymmetry raises urgent governance questions about who controls the weights and how to secure access.
🔮 Exponential View #569: What Gathers Around a Powerful Model?
Anthropic unveiled a preview of its Mythos model, prompting a coalition to assess how such powerful AI could distort risk pricing for critical digital infrastructure. A new genetic study explains why GLP‑1 drugs produce wildly different outcomes, pointing to DNA...
💸 Mythos and the Mispricing of Everything
Anthropic’s Claude Mythos preview demonstrates AI agents that can generate functional cyber exploits in hours, challenging the long‑standing belief that offensive capability requires scarce human expertise. The US utilities sector, a $1.5 trillion market priced at roughly 22× earnings, treats cyber...
🔮 Where We’re Taking Exponential View Next
Exponential View announced that Greg Williams has joined as its first Executive Editor, ending a two‑year search for a leader who can match the studio’s cross‑disciplinary ambition. Williams, a former editor of Arena, WIRED UK, and a novelist, brings a...
📈 Data to Start Your Week: The AI Capacity Trap
Cheaper AI has triggered a classic Jevons paradox, where falling token prices spur demand that outpaces supply. OpenAI’s API traffic surged from 6 billion to 15 billion tokens per minute between October 2025 and April 2026, a 2.5‑fold rise in five months. Google’s aging...
🔮 Autoresearch and the Experimental Society
Autoresearch is an open‑source framework that runs autonomous experimental loops, letting a human define goals while an AI iterates toward optimal results. Andrej Karpathy’s initial test trained a GPT‑2‑scale model two days faster and discovered 20 genuine improvements, a speed...
📈 Data to Start Your Week: The AI Buildout
The AI-driven data center boom is surging, with 89% of new capacity in North America already pre‑leased and a pipeline that grew 159% to 241 GW in a year. However, two‑thirds of that pipeline sits idle due to grid connection queues...
🔮 Exponential View #567: How AI Is Rewiring Work
Exponential View #567 examines how artificial intelligence is fundamentally reshaping the nature of work. The episode highlights NVIDIA’s strategic investment in Groq’s custom AI processors and the emergence of OpenClaw, a platform promising ultra‑low‑latency model inference. By accelerating compute, these...
📈 Data to Start Your Week
A recent data roundup highlights three pivotal trends: U.S. data‑center construction spending surpassed office‑building spending in December 2025, signaling a shift toward cloud infrastructure. Global semiconductor sales jumped to $82.5 billion in January 2026, a 46.1% year‑on‑year increase. Meanwhile, more than...
🔮 The Lantern and the Flame
In this episode, the host explores how pervasive AI tools are reshaping our thinking, distinguishing between helpful cognitive offloading and the dangerous trend of cognitive surrender—uncritical abdication of reasoning. Drawing on research, personal workflow experiments, and references to thinkers like...
The AI Race that Apple Is Winning
In this episode of Exponential View, host Azeem Azhar reexamines his view of Apple’s role in the AI race, highlighting a paradox: while Apple appears absent from AI model development and large‑scale data center investment, its hardware—especially high‑spec Mac Minis...
📈 Data to Start Your Week
The Strait of Hormuz closure removed roughly 20 million barrels per day, creating the largest oil shock in history and dwarfing the next five biggest disruptions combined. In the AI arena, Anthropic’s Claude has surged to about 70% of U.S. business...
🧐 The Boundary of Tedium
In this episode, the host details his AI agent R Mini Arnold (RMA), a Mac Mini‑based system that uses Anthropic’s Claude models to automate tedious knowledge‑worker tasks such as presentation assembly, personal CRM building, and note‑file organization. He demonstrates how...