On the emergence of a new programming language

Telling computers what we want them to do is tough. They force their own logic on us, their languages can be hard to read (except Python, the most beautiful language of all, of course), and accomplishing certain tasks often requires deep insight into how things work under the hood.

In computer science class in high school, for example, our teacher asked us to have the computer draw simple geometric figures along the sides of an octahedron. To solve this, we had to dive into polar coordinates - something we knew nothing about back then. I remember filling pages with notes and taking photos afterward because I was so proud of finally getting it to work.

But if I could have simply told the computer, “Please draw these shapes along the sides of an octahedron,” I wouldn’t have faced the same difficulties.

You can apply this idea to any field - 3D graphics, GPU programming, or even coding a quantum computer with Shor’s algorithm. They’re all backed by complex theories that we, as humans, must understand in order to make computers carry out our tasks.

Well, that era seems to slowly come to an end.

Vibe Coding

When I started neural frames, ChatGPT had just been released. Thanks to it, I was able to build my first-ever React and cloud computing project - something that would have taken me much, much longer than without a smart assistant on my side. I practically speedran the challenges I encountered, even though the original ChatGPT 3.5 wasn’t nearly as good as the LLMs we have now.

Then some kids in San Francisco noticed a new way of programming was emerging. They forked the standard programming editor, VSCode, added a ChatGPT sidebar, and Cursor was born. It might have seemed obvious, but they were in the right place at the right time and executed brilliantly. There’s some startup wisdom in that: Solve the obvious things first.

Cursor is now at the forefront of a new programming paradigm that the AI legend Andrej Karpathy calls vibe coding.

The indie-hacking legend Pieter Levels “vibe-coded” a flight simulator with a functioning multiplayer mode where people can dogfight each other. In typical Levels style, he quickly found ways to monetize it—by selling special planes and even zeppelins that carry ads. Yesterday, he reported, that he is at $30k monthly recurring revenue with this product.

Levels became famous for having very simple websites written in PHP and iterating quickly. He is by no means a 3D-graphics person. But vibe coding made him one.

The graphics are not nice, of course. But there’s a future, not too far from here, where the images can be creatively enhanced with 60fps. Just remember how images used to load in the internet around the year 2000. Line by line. Taking forever. Today, we stream 4k videos on Netflix.

In the past two weeks, I’ve built two features for neural frames involving dragging and dropping images and video clips onto our timeline. I had no idea how to handle drag-and-drop before, especially copying settings from the original file to the new place. Cursor’s agent mode did it for me.

Is the code poor quality? Not really. Could this be a problem for code quality in a large, complex software project? For sure. Will future AI be able to clean up the messy code it’s currently generating? Absolutely.

A friend of mine said, “This is fine for front-end projects nobody cares about; I just hope nuclear power plants aren’t being programmed like that.”

Fair point.

But the genie is out of the bottle—and it’s not going back in.

And even if codebases will become qualitatively worse with LLM’s writing them, vibe coding already has some sort of escape velocity: In a few years time, the successors of today’s LLM’s will clean up the codebases their grandfathers wrote.

Programming already changed so much in these past years. The next programming language will be English. Will we still be as proud of what we’ve built as I was when I cracked the octahedron with polar coordinates? I’m not sure.

AI Video of the week

I absolutely loved this paper-unfolding clip by our amazing customer support wizard, Liz. Using subtly similar keyframes with Runway Gen-3-Alpha on the neural frames timeline, she managed to make the paper unfold entirely on its own—just from a few text prompts!