Azeem Azhar's Exponential View

A deep dive into why inference is overtaking training and what that means for company budgets. A new OpenClaw-style harness is framed as a turning point for deploying agents at scale. Token consumption skyrocketing from millions to hundreds of millions per day and how organizations should govern and budget for it. Practical examples of agent workloads, model choice strategies, and shifting economics of compute.

A rethink of Apple’s place in AI, tracing how Mac Mini demand and OpenClaw use shifted perceptions. Discussion of China’s rapid agent adoption and Perplexity’s Mac Mini product. Exploration of Apple’s hardware, privacy enclaves and App Store as capture points. Examination of on-device models, hybrid device-cloud setups, and why ‘good enough’ local inference wins for sensitive daily tasks.

They debate whether intensive AI use sharpens thinking or quietly erodes it. He explains custom AI tools for signal detection, synthetic personas to surface diverse insights, and an argument engine that stress-tests ideas. The episode highlights practices for quiet, handwritten incubation, speaking drafts aloud, and a Stylometer for final wordcraft.

A deep dive into a personal AI chief of staff running on a Mac mini and managing a ten-person equivalent team. Hear how overnight agent crews research, debug code, and draft long scripts in parallel. Learn about the hardware, software stack, personality spec soul.md, costs and security checks. Practical steps for delegating tasks and why individuals can outpace big firms with agent teams.

Nita Farahany, law and philosophy professor specializing in cognitive liberty and mental privacy. Eric Topol, medical researcher and digital medicine pioneer. They debate whether AI controls our decisions, how AI changes medical practice and skills, the risks of de-skilling, and what we lose when creative work is ceded to tools.

Rohit Krishnan, writer and builder behind Strange Loop Canon and an engineer-economist-operator, joins to talk practical AI agents. They explore building and running persistent agents like Morpheus, tool stacks such as OpenClaw, and why writing remains the toughest task for LLMs. Conversation also zooms out to the infrastructure, identity, and exchange needed for a future of trillions of agents.

Matt Robinson, financial journalist and founder of AI Street, probes AI company finances. Hannah Petrovic, research lead with an astrophysics background, reconstructed OpenAI’s public unit economics. Jaime Sevilla, founder of Epoch AI, analyzes scaling laws and compute trends. They examine OpenAI’s tight operating economics, model depreciation, infrastructure bottlenecks (GPUs vs energy), rising compute demand from persistent agents, and ad-driven revenue paths.

Mustafa Suleyman, co-founder of DeepMind and CEO of Microsoft AI, is an AI leader focused on strategy, safety and policy. He warns about AI that seems conscious and how it can hack our empathy. They discuss a possible new class of being, market forces making AIs more engaging, engineering limits to prevent personhood, and the risks of open, persuasive systems.

Live dispatches from Davos capture political tension and a sense that the rules-based order is fracturing. European leaders push technological and political sovereignty. On the ground, conversations focus on practical AI adoption in enterprises, open source models, startup innovations, and energy transition challenges.

Peter McCrory, Head of Economics at Anthropic and lead author of their Economic Index report, discusses AI's current impact on productivity and labor. He highlights distinct usage patterns of Claude, emphasizing the need for businesses to extract tacit knowledge to unlock AI's full potential. McCrory warns of risks for junior workers as automation changes job skill dynamics while estimating that AI could boost productivity growth by 1.0-1.8% over the next decade. They also explore the complexities of implementing AI and the importance of developing cognitive endurance alongside technology.