Programming as Archaeology

illustration

Long before neural networks learned to write code, I programmed in a way that many programmers still find irritating. I rarely started from an empty file. If a piece of a working system already existed somewhere that solved at least half of my problem, I was far more interested in figuring out how to adapt it than in proving to myself that I could write the same thing from scratch. Sometimes I dug into its insides, sometimes not, and usually as deep as was needed to get the next layer of the system working.

For a long time I regarded this as something like a bad habit best hidden from serious engineers. Then language models appeared, and this habit turned out to be perhaps the most useful skill for working with them.

---

In 1911 the English mathematician and philosopher Alfred North Whitehead wrote in an introductory book on mathematics a sentence I consider one of the most accurate ever said about technology: civilisation advances by extending the number of important operations which we can perform without thinking about them. He added that operations of thought are like cavalry charges in a battle: they are strictly limited in number, they require fresh horses, and must only be made at decisive moments.

Each of us lives most of our life this way. You drive a car without knowing at what moment the spark fires in the cylinder. You pay by card without imagining how many banks and processing centres your purchase of a coffee passes through in the two seconds the terminal is thinking. You open a website, and not a single cell of your brain is occupied with how the protocol works by which the page makes its way to your phone. Programmers in this respect live like everyone else: no normal developer reads the source code of network protocols before writing a program that sends data over the network. He knows what the protocol promises, and that is enough as long as the promise is kept.

---

archive

I probably got used to this way of working back in the early 2000s, at the St Petersburg company Nienschanz, where I worked on systems integration. We assembled large computing systems for clients out of other people’s components. One of those components was Oracle RAC, a database that can spread itself across several servers so that to the programs talking to it, it looks like one. Another job I remember was a computing cluster built on Intel Itanium processors for AvtoVAZ, the Russian carmaker, on which the LS-DYNA package was to run.