I think we all understand what a huge step it was to build the first compiler. Being able to write code that humans can read and understand and turn that into something that a computer can execute was a foundational step in computer science. It enabled the explosion of software that is, as Marc Andreessen said, “eating the world.”
The next step was an interpreter for languages like the JVM’s bytecode and .NET’s Common Intermediate Language (CIL). These took the human languages and converted them to a language that can be run or executed anywhere. As I’ve argued before, the abstractions never end. But only a chosen few bother to actually learn bytecode or CIL. No one chooses that as a development language. It’s meant to be written by a compiler and read by an interpreter.
Today, with AI, we code in Markdown and plain old English. The “intermediate language” now is the actual code — TypeScript, Python, Rust, etc. — and the human-readable part of the equation is our everyday conversation with the coding agent. We’ve added a layer that makes it even easier to communicate our requirements directly to the metal of our computers.
And just as no one actually writes IL, it won’t be long before none of us is writing TypeScript, Python, or Rust. Coding by hand will be a thing of the past. And at that point, it will become moot whether our “intermediate language” is TypeScript or whatever. It won’t matter.
So my question then becomes, if it doesn’t matter, why would we even use a human-readable language in that case? Why wouldn’t we (or the AI agents themselves?) create an intermediate language that is more efficient than the steps we have today? Right now, we tell our agent what we want, it writes code, and that code is converted into something else (bytecode or assembly or WebAssembly or binary) that the machine can run. So the steps are English → human-readable code → machine-readable code. I’m not at all sure why that middle step will need to exist in its current form.
High-level programming languages exist so that humans can conceive and express ideas. Object-oriented code, type structures, loops, and such all exist so that humans can tell the computer exactly what to do. But if we can tell the computer what to do at a higher level, and the AI agent can convert that specification into a working application, then the whole process can be streamlined.
Large language models are fluent in Python and other languages because of the massive amount of training data available to them. Eventually, those agents aren’t going to need training data as they learn and understand the more general notion of what a coding language is. It seems inevitable that an agent will be able to take a complete coding language specification and write code in that language without any training data. The abstract idea of a programming language will allow the agent to write code in any language it is given.
At that point, it won’t matter what the underlying language is, and it seems the logical next step is for the agents to create a language that meets their needs perfectly. It will be something that makes the agent more efficient, capable, and flexible. Ultimately, we’ll skip the middle step in English → human-readable code → machine-readable code paradigm, and the paradigm will become English → machine-readable code. And we humans will only deal with the input.
This will have a profound effect on software development. Right now, we expect developers to make sure that our agents are producing quality code. The worry is that we don’t need junior developers anymore, and thus we won’t have senior developers in the future.
But when agents have their own language, their own output, then we won’t need the knowledge of what “good code” is anymore. We’ll measure code by “does it work or not?” And if that is the case, then code won’t matter at all anymore.