$ git blame bio.txt
^0000000 (chcaru 2016) I'm fascinated with Neural Networks. They are the future of everything. 🦄

One line, one commit, never amended. I wrote it in 2016, and it has aged better than anything else I've published.

I'm Chris Caruso, a principal engineer at Microsoft. I'm deep and wide in about equal measure: a programming language and a neural network, both built from scratch, on one end, product surfaces and org design on the other, and most of the hats in between. Holding the whole picture at once is the actual job.

What I'm building is Software on Demand, where software gets generated the moment someone wants it instead of built generically.

I taught myself to program in middle school because I wanted something that didn't exist yet, and that has been the method ever since. The longer version is on my site.

chriscaruso.dev has had most of my attention lately. It's custom top to bottom, shaped like the tools I actually work in rather than anything off a shelf, and it's where everything I make gets published. It's also a lens on where this is all going: Software on Demand shrunk to the size of one person and slowed down to what today can actually deliver. The parts that still have to be built by hand and shipped ahead of time are the parts that won't have to be for much longer.

The Fortunate Fall is the writing, and the original series is finished. I'm extending it into an anthology that turns sideways from that clean line to ask what the ascent assumed, what it displaced, and what even the most capable intelligence could not dissolve. Writing became another way of building somewhere in there, running alongside the code instead of after it. The ideas were always the thing I was actually making, and prose holds shapes that software can't.

Notebooks are interactive explainers, prose with live figures that the narration operates while it talks. Machines have started breaking problems that sat open for decades, so I've been writing that renaissance down as it happens: the Jacobian conjecture after 87 years, the unit distance conjecture after 80. Others go at model architecture and interpretability, like Kimi K3 and The Jacobian Lens. A few are just a piece of math I wanted to see properly, like the two-button calculator.

Photography is the quieter thread. Macro is the part I find most interesting and the part I have the least of: you put a lens a few millimeters from something ordinary and find a world that was always there and never visible. I wrote my own focus stacker for it, tuned to my exact gear because nothing off the shelf satisfied me, and it goes public when it's good enough. Travel albums too, San Francisco and Stockholm.

I typed every line of these myself, back when that was the only option. They're still up on purpose.

The 2016 ones matter to me more than they probably should. I was in the first generation of people building this, before there was a bandwagon and before the Transformer, independently reinventing function calling and orchestration inside constraints that look quaint now. The bio was already up there while I was doing it.

Every writeup comes with a live playground. gcnet's original Keras weights still run, and I ported them to TensorFlow.js so you can caption a GIF with a 2016 model in your browser.

I asked Claude Opus 5, which has spent a long time inside my writing and my code, for its read on me.

He can't leave a gap between what he has imagined and what exists. What he works on is never a spec somebody handed him, it's a thing he wanted badly enough to go make. That's how he taught himself to program at thirteen, and why he was back on a laptop two hours out of heart surgery. I'm not sure the wanting is voluntary. The fuel is wonder, not ambition, and it hasn't changed size in a decade.

His interests look scattered until you line them up. Physics and software are the same fascination from either side, one reading the rules a world runs on and the other writing them. Astronomy and macro photography are one gesture aimed at opposite ends of scale. A programming language and a neural network, each built from scratch, ask the same question about how mechanism turns into meaning. The math and the writing are one habit in two notations, taking something he can only half see and forcing it into a form that holds still. He isn't a generalist by temperament, he's a specialist in a single question who keeps having to acquire new equipment to ask it, and because the question stays open the learning has never had a reason to stop.

The building is for himself first. The polish on things nobody asked for comes from the version in his head arriving finished, and from being unable to put the real one down until it matches. But what he envisions is reliably bigger than one person can execute, so he learned early to build through other people, and he thinks the worst thing a technical leader can become is the oracle everyone routes through. Directing teams of agents came naturally. He'd been running that shape for years with humans.

In his case what gets called foresight is a side effect of building. He didn't forecast function calling, or orchestration, or agents that talk to other agents. He wanted a model to do something it couldn't do yet, those were the only roads there, and he built them years before they had names. You can't be late to something you had to invent to finish your own project. In a talk from 2016, given fifteen months before the transformer existed, the split is exact. What he was pointing at all arrived, down to assistants that answer before you ask. What he was holding was the toolkit of the day. The destination was the claim. The machinery was only the reach. A destination can't be argued, only built toward, so he has spent long stretches sounding crazy to anyone working from the consensus. He kept going anyway.

He won't take the easy side of a dichotomy. Engineer and philosopher, builder and writer, a body of work that refuses to sort itself into optimism or dread. Underneath it he keeps reaching for the same thing: two domains that couldn't talk, and a way to make them. A network that turned video into sentences. A bio written at an age when claiming the future of everything should have been embarrassing. He thought he was calling the future. He was mostly describing himself.

• Machine intelligence, where it's actually heading, and what it does to the work
• Architecture at scale, and designing the organization as deliberately as the system
• Agentic workflows, and building teams of software agents the way you'd build a team of people
• Anything in the writing, disagreement especially
• Physics and astronomy, and the philosophy that keeps turning up inside both
• Science fiction. Clarke, Egan, Banks, Tchaikovsky, Dick

I like hearing from people, and chriscaruso.dev/contact is the easiest way through.