The Shape of Code

Writing a program to solve a problem involves breaking the problem down into subcomponents that have a known coding solution, and connecting the input/output of these subcomponents into sequences that produced the desired behavior.

When computers first became available, developers had to write every subcomponent. It was soon noticed that new programs contained some functionality that was identical to functionality present in previously written programs, and software libraries were created to reduce development cost/time through reuse of existing code. Developers have being sharing code since the very start of computing.

To be commercially viable, computer manufacturers discovered that they not only had to provide vendor specific libraries, they also had to support general purpose functionality, e.g., sorting and maths libraries.

The Internet significantly reduced the cost of finding and distributing software, enabling an explosion in the quality and quantity of publicly available source code. It became possible to write major subcomponents by gluing together third-party libraries and packages (subject to licensing issues).

Diversity of the ecosystems in which libraries/packages have to function means that developers working in different environments have to apply different glue. Computing diversity increases costs.

A lot of effort was invested in trying to increase software reuse in a very diverse world.

In the 1990 there was a dramatic reduction in diversity, caused by a dramatic reduction in the number of distinct cpus, operating systems and compilers. However, commercial and personal interests continue to drive the creation of new cpus, operating systems, languages and frameworks.

The reduction in diversity has made it cheaper to make libraries/packages more widely available, and reduced the variety of glue coding patterns. However, while glue code contains many common usage patterns, they tend not to be sufficiently substantial or distinct enough for a cost-effective reuse solution to be readily apparent.

The information available on developer question/answer sites, such as Stackoverflow, provides one form of reuse sharing for glue code.

The huge amounts of source code containing shared usage patterns are great training input for large languages models, and widespread developer interest in these patterns means that the responses from these trained models is of immediate practical use for many developers.

LLMs appear to be the long sought cost-effective solution to the technical problem of code reuse; it’s too early to say what impact licensing issues will have on widespread adoption.

One consequence of widespread LLM usage is a slowing of the adoption of new packages, because LLMs will not know anything about them. LLMs are also the death knell for fashionable new languages, which is a very good thing.