Explainers
Deep dives into AI systems, written for humans.
Agentic AI and the Autonomy Question
For most of its history, AI has been reactive. You type a prompt, you get a response. You click a button, the model classifies an image. The human acts; the machine answers.
AI and Consumer Surplus
When AI makes something cheaper, faster, or better, someone captures that value. In economics, the portion that goes to buyers is called consumer surplus: the gap between what you'd be willing to pay and what you actually pay. A flight worth $500 to you that costs $200 leaves you $300 better off. That's consumer surplus.
AI in Healthcare: Who Benefits from the Diagnostic Revolution
A clinic in rural Tamil Nadu has no radiologist. The nearest one is four hours away. A patient arrives with a cough that won't quit and a chest X-ray taken on aging equipment. An AI system reads the scan in seconds, flags a suspicious mass, and recommends follow-up. In a well-funded Boston hospital, the same AI system runs alongside a team of specialists who barely need it. Both sites use the same algorithm. The outcomes couldn't be more different.
The Economics of Training a Frontier Model
Training a frontier AI model now costs roughly what it takes to build a skyscraper. That number keeps climbing. Understanding where the money goes reveals who gets to build these systems, who profits from them, and who gets locked out.
How Language Models Actually Work
When you ask ChatGPT a question, what actually happens? Not the marketing version — the mechanism.
How Reasoning Models Actually Work
Standard language models generate text one token at a time, optimising for the most plausible next word. Reasoning models do something different: they spend compute thinking before they answer. The distinction sounds subtle. The consequences — for performance, for cost, and for who gets access to capable AI — are not.
Multimodal AI: When Machines See, Hear, and Read at Once
For most of AI's modern history, models were specialists. One system read text. Another classified images. A third transcribed speech. They lived in separate silos, each blind to what the others could perceive.
Physical AI: When Machines Learn to Touch the Real World
For the past three years, the AI conversation has been dominated by language. Chatbots write emails, generate code, draft legal briefs. The disruption has been real, fast, and concentrated in offices. But a different transformation is underway — quieter, more capital-intensive, and aimed at a different workforce entirely.
How Retrieval-Augmented Generation Changes Who Needs What
When a company plugs a language model into its internal knowledge base, something shifts. The model didn't get smarter. The company got more leverage from the intelligence it already had. This is Retrieval-Augmented Generation — RAG — and it is quietly rearranging who needs what in the AI economy.
Recursive Self-Improvement: The Feedback Loop That Keeps AI Leaders Up at Night
In 1965, the British mathematician I.J. Good described a machine that could design a better version of itself. That improved version could design an even better one. And so on. He called this an "intelligence explosion" and predicted it would be "the last invention that man need ever make."
Running AI Locally
Every time you send a prompt to ChatGPT or Claude, your words travel to a distant data center, get processed on hardware you don't own, and generate a response that's logged on servers you can't inspect. The transaction is convenient, capable, and entirely dependent on someone else's infrastructure.
Synthetic Data and the Ouroboros Problem
AI labs have a resource problem. The models keep getting bigger, but the planet's supply of high-quality human-generated text is finite. The entire indexable internet has already been scraped — books, articles, forums, code repositories, academic papers, social media posts, government filings. The well is running dry. So the industry has landed on an elegant-sounding solution: have AI generate its own training data. The trouble is, this solution contains its own undoing.
The Productivity Paradox
AI will make us all more productive. We've heard this promise before — about computers in the 1980s, about the internet in the 1990s, about smartphones in the 2000s. Each time, the productivity gains were real. Each time, most workers saw little of the benefit.
The Shape of AI Displacement
"AI will eliminate 47% of jobs." You've seen this headline, or some version of it, recycled every few months since 2013. It's become background noise — alarming enough to share, vague enough to ignore.
The Training Pipeline
Before a language model can answer your questions, it goes through a multi-stage process that shapes its behavior. Understanding this pipeline reveals where decisions get made and whose values get embedded.
What Fine-Tuning Really Does
A pre-trained language model is a general-purpose tool. It knows a lot about everything and not enough about anything specific. Fine-tuning is the process of taking that general tool and reshaping it for a particular job — teaching it your vocabulary, your standards, your domain.
What Model Cards Tell You (And What They Don't)
Model cards are standardized documentation for AI models. They're supposed to be the nutrition label for machine learning. In practice, they range from genuinely useful to artfully incomplete.
World Models vs. Large Language Models: Two Visions of Machine Intelligence
There is a quiet civil war in AI research, and the stakes are not just technical. On one side: large language models, the architecture behind ChatGPT, Claude, and Gemini, which have consumed hundreds of billions of dollars in compute and crowned a handful of companies as the gatekeepers of machine intelligence. On the other: world models, a fundamentally different approach that aims to learn how reality works — not by predicting the next word, but by building internal simulations of the world.