When I was a kid I wanted to become an astronaut. But then I found out you have to be a pilot first and that seemed boring. So I decided to become an astronomer. But that involves working at night. But then I read in a science magazine aimed at children about how you could simulate planets coming into being using a computer. So I decided to do that. I fired up BASIC, implemented my understanding of Newtonian physics and the universe in motion.
It did not go well. To run a simulation of the Kant-La Place theory of the solar system formation you’ll need thousands if not tens of thousands of rocks orbiting the sun for thousands if not tens of thousands of years. My implementation on an 8 bit home computer could simulate at rough resolution a year in about a minute. With maybe 25 rocks.
But programming itself was fun, so I stuck with that. I learned that BASIC isn’t the fastest language so I dove into compiler building. Calculating gravity between every pair of rocks turned out to be really expensive and I learned about algorithmic complexity and to avoid O(N^2). I figured out geohashing which came in handy when I worked in mapping later.
And then many years later I put it all together and built a system that could actually simulate the Kant-La Place theory on a desktop computer. It was still quite slow and because it ran on a windows desktop had limited interactivity and you couldn’t easily share it. But it made a point. Meanwhile I had become a professional software engineer which probably ended up being a better career than either astronaut or astronomer would have been.
In the relatively short period since ChatGPT came out LLMs have moved on from being cute but often confused chatbots to being full blown coding agents. At Block I’ve been working on one. Each iteration is a little better than the previous one, and can do something the previous couldn’t. So last weekend I asked goose powered by Opus 4.5 to try its hand on my childhood dream.
The results have been mind blowing, fun and disconcerting. Mind blowing because it can actually do this. It needs help with some of the algorithms, but mostly it does ok. And it knows the physics from Newtonian mechanics to planetary escape velocity and the speed of hydrogen at 1000K. Fun because this is the way to build these small projects. Focus on what it looks like, the core algorithms and have the AI throw up a UI, keep track of stats or visualize gas depletion.
But also disconcerting. Historically the way to get good at hard things is to get good at the easy stuff first. It takes humans years to become better at writing code than the top AI models can today. And much like Achilles never catching the turtle, after those years of study, the AI will have improved even more. Why bother? And if nobody bothers, where do the future experts come from?
Maybe it’s fine. I imagine in the fifties die hard machine code writers sighing that compilers are all good and well but who will write them in the future if nobody bothers to learn how CPUs work. And at the current trend no human will outprogram an AI in a few years; it happened to chess and we still learn that. For now building stuff in software is more fun than it has been in a long while.
