I've been writing, architecting, and managing the production of code for the past 25+ years. I can confidently say that the way we write code and create software has changed significantly over the past quarter century. Most of these changes have added value somewhere in the process.
TDD · BDD · source control · git + pull requests · CI/CD · containers · package managers · functional coding · type systems · MVC frameworks · Web 2.0 · accessibility · service-oriented architectures · API-first development · cloud-native
Thinking back over this list and remembering what it was like to write software in 2000, the delta is hard to wrap my head around.
Up until now, all of the tools and systems that transformed software engineering were built by humans for humans. They were created to help us overcome our own weaknesses and to produce working code more reliably and consistently.
If you think about it, the tools and patterns these AIs use to write code were all created for humans to solve human problems and to address human weaknesses. I know that modern AI systems are trained to behave like humans, but they are also different in many ways. Is our human created, human-centric programming paradigm optimal for AIs? Or is there another set of tools and patterns that would benefit AIs in ways that differ from humans? What if AIs could create their own programming language? What would the anatomy of such a language be?
I started this effort with these questions in mind, but once I cracked the seal and started building, I began to ask another set of questions. What if this programming language could have any feature I wanted? What if I could bake in functionality to make building things easier and more secure?
Then, after I had built some apps, I began wondering… what if the code required for this part of the app could be 2–3x less verbose if I changed portions of the language?
Programming languages don't get created very often, but I think we are due for a new one. Perhaps we are due for a new set of languages to solve modern problems in better ways. These languages could enable AIs to create software differently than humans have created it in the past. Maybe we need languages that go deeper in certain areas, that address modern software patterns in first class ways and that help us build modern apps with modern patterns faster and better.
I realize that LLMs have been trained on existing languages and that asking LLMs to program in a new language that they were never trained on is an uphill battle. That said, I've been impressed with how well LLMs are able to build things in this language, even though they've not been trained on it specifically.
The effort is experimental and exploratory. I intend to keep exploring these questions through the development of the language. It might not go very far, but it already runs a number of websites and apps I've created (including this one).
The language is called ntnt.