Notes from inside the AI transition.

The agent era is creating a sharp question for every enterprise SaaS vendor: are you a UI or a platform? The answer is being decided right now, in real time. Salesforce just gave one of the most legible answers.

Synthetic content is everywhere now. The skill we used to call 'can you spot it' is becoming 'what does it mean that you can't.' A note on detection, skin in the game, and harm reduction.

All of the upstream chemistry, physics, and capital comes to rest on a single computation. One forward pass through the trained network. Out the other end: one token. How to reason about per-token energy and cost without hiding the assumptions.

Power, cooling, and capital around a training cluster. An H100 scenario with the assumptions and units visible.

Repeatedly patterning circuits on silicon. The process control, resource accounting, and yield behind an AI accelerator.

Why I want to spend the next decade inside the rooms where AI is being deployed into the real economy, and what it looks like from there.

From polysilicon chunks to a single-crystal wafer. Seed growth, slicing, and the mass balance behind chip manufacturing.

The first unit operation in the AI supply chain takes quartz sand and turns it into a semiconductor feedstock with tightly controlled impurities. The purification is the product.

Notes from a year of helping organizations ship agent systems into production. What actually works, what doesn't, and what surprised me.

Price's Law on AI steroids, Dunbar's number for AI-native orgs, and what happens when 10% of the people doing 50% of the work get tooling that's 10x better.

Anthropic shipped a protocol. Most people thought we needed a framework. The protocol is winning.

On consuming the internet the way models do, what it costs you, and how to stop scrolling long enough to learn something.

I've sat in on a lot of these. The technology isn't what kills them. The ownership question kills them, every time.

2024 was the year everyone tried to build a CEO out of LLMs. The frameworks were everywhere. The results were not.

Why optimizing for a metric tends to break the thing the metric was measuring, and how to design systems that survive it.