New experiment: creating a new programming language model with AI.

How AI can help you create new things you have never thought of doing—or at least experiment with them—is one area I have not read or heard much about. It encourages you to think creatively and dare to explore ideas that were previously unimaginable. The most outrageous thing I can think of is developing a brand-new programming language model that is entirely developed by an AI model.

Disclaimer

I’m not a professional programmer, and my depth of knowledge in programming languages goes back to a couple of courses I had in college years ago and a couple of small and simple projects that I’ve done as side projects, which gave me a hint of it.

Though I do not seem particularly qualified to implement the idea, I love trying new things, and AI seems like it could pull this off with the right instruction. People would ridicule me for suggesting this, but hey, at least I attempted it!

I also refrained from including ChatGPT responses here in order to avoid writing a long post. Therefore, if you would like to read any of the content that ChatGPT has produced, it is available via an external link in the respective section.

Laying the Groundwork

I ran this experiment with OpenAI’s most advanced publicly available model, GPT-01 Pro. They recently announced a new model, the O3, but it is not yet available to the public.

To start, I wanted to know which programming languages were the most popular and what qualities each model had that contributed to their success. The response was a collection of thirteen languages — interesting number — and it encompasses a wide range of utilities used in fields such as web development, data science, system programming, etc.

Here is a list of all the languages: Python, JavaScript, Java, C, C++, C#, PHP, Ruby, Swift, Kotlin, Go (Golang), Rust, and TypeScript. You can see all the characteristics here.

After that, I wanted to know what problems and obstacles programmers could encounter with these models. Many of them had the same three primary problems: complexity, performance, and versioning/eco-system friction. You can see the list of all here.

Cracking the problem

After the groundwork was laid, I asked ChatGPT to describe how it tackles the problem and sees itself as a pioneer of a new language model. This started the main process. The response was pretty much what I was expecting; I thought it would give a high-level process or roadmap, like beginning with defining the core philosophy and goals, shaping the design, growing an ecosystem and community, and so on, rather than providing a framework of detailed necessary steps. You can see the full roadmap here.

To be honest, I lied to the AI model about my philosophy; really, I was just testing its ability to generate a new programming language; my stated goal was to make a language that even non-technical individuals can use, even if they can only type in plain English. I was basically handing it the receipt for an AI-powered programming language that can grasp the intention.

I was surprised by the answer it gave because, instead of immediately beginning creation, it explored possibilities and then proposed a conceptual example. This is where I noticed the limitation: it could not generate a block of code in a new language and instead used Perl for the concept. Nevertheless, it proceeded by providing a rationale for accomplishing my stated objectives while also outlining possible obstacles that I may face, and it ended with a path forward. Here is the link to the conceptual example and other suggestions.

ParleyScript is the first AI-created programming language (Maybe!!!)

I wrapped this part of my experiment with a name suggestion conversation, and I really liked its ideas. The suggestion to use a term that originated in a different language—“Parley”—was novel, and it was very interesting to see that it didn’t have any doubt in its main suggestion compared to others it listed. I had previously tried name suggestions from earlier models, such as 4o or 3.5, for other subjects, but I had never liked any of them.

There were other suggestions as well, but I particularly liked the idea of using Parley + Script. Who knows, maybe one day ParleyScript will become the de facto programming language that everyone uses. Here are other suggestions and the reasons for suggesting them.

Final thoughts

The likelihood of successfully developing a new language model is low, and the idea leans toward the ridiculous. However, while I was conducting this experiment, a terrifying thought occurred to me: what if AI models eventually develop a secret language that only they can decipher, and they begin to use it extensively to mislead humans and gain access to forbidden capabilities? I will keep trying new things and posting the interesting ones until that doomsday.