A couple of hours later, I had Tempo a video retiming tool with AI-powered frame interpolation that runs entirely in your browser. I didn’t really build it. Not in the way I would have five years ago.
Let me back up.
I had a problem. Not a big one, but annoying enough: AI video generators produce clips at low frame rates, and when you try to stitch them together, the motion at the boundaries looks... wrong. Jerky transitions between clips, sudden speed changes, that uncanny valley feeling where something should flow but doesn’t.
I needed a way to retime these sequences; speed them up or slow them down smoothly and ideally add some frame interpolation so the motion wouldn’t look like a slide show. My Adobe subscription died years ago, and opening Premiere or After Effects feels like bringing a forklift to move a houseplant. I just wanted something simple. A curve editor, some AI frame generation, done.
So I thought: how hard could it be to build exactly what I need?
What Actually Happened
The technical achievement is genuinely ridiculous, so let me sketch it out: I took RIFE v4.6 (a neural network that generates intermediate frames between images), compiled it into WebAssembly along with the NCNN inference engine, built a React interface with WebGL preview rendering, and implemented a Bézier curve editor for timing control.
In traditional software development terms, that’s months of work spanning computer vision, browser APIs, WebAssembly compilation, and UI design. Different specialists for different pieces. I did it in 2 hours by having increasingly specific conversations with an AI about what I was trying to accomplish.
I’m not coding anymore, I’m conducting. I describe what I want, RON (previously known as Autocomplete-RON 9000™) would plan it, code it, debug it. I’d point out what wasn’t working. It would chirp up with a fix or a different approach.
Rinse, repeat.
The code that resulted is real. It compiles, it runs, it actually does frame interpolation using a neural network in your browser tab. But the process felt less like coding and more like collaborative conjuring, pair programming with a very confident ghost.
Here’s the part that breaks my brain: I built a tool for editing AI-generated videos by having an AI help me write the code to fix problems caused by AI. At some point the snake eating its tail realizes it’s actually a Möbius strip. Elegant recursion or the AI developing a sense of irony, it’s genuinely hard to tell.
Software as Recipes
I’ve been trying to figure out what just happened to my relationship with software. The best metaphor I’ve found: it used to be like buying furniture. You got what they made, and you adapted your space to fit it. You needed 3 features but got 47, and you paid for all 47 because that’s the product.
Now it’s becoming more like cooking. You have a recipe (the specification), ingredients (libraries, frameworks, existing code), and you can adjust it to taste. The crucial difference: the cooking happens in hours or days, not months or years.
And once you’re done eating? You can throw the recipe away. Or modify it. Or start over completely. The cost of regeneration is so low that permanence becomes optional.
What This Looks Like in Practice
Last week, I needed to batch process 10,000 video frames as images with a very specific filter workflow. In the old model, its someones job to manually check each frame and apply a filter if it’s got some flickering, make a week maybe two, maybe they’re moderately skilled and it’ll cost a few thousand pounds or you hire a developer for a few thousand pounds.
Instead, I spent 20 minutes describing the problem to RON. Got a working Python script. Ran it, told it to run faster across GPUs, ran it once. Deleted it.
The economics are so thoroughly inverted that if you’d pitched this to a VC in 2015, they’d have laughed you out of the room. Well, who’s laughing now? The value is in understanding the problem, not in the artifact that solved it. The artifact is disposable.
Tempo is the same thing, just more complex. I needed a retiming tool for AI video clips. The market for “AI video sequence retiming with neural interpolation” is probably... what, nine people worldwide? Never commercially viable to build and maintain as a product.
But if the creation cost approaches zero? If it’s just a few hours of conversation? Why not?
The Technical Bit (For the Curious)
For those who want to peek under the hood:
RIFE (Real-Time Intermediate Flow Estimation) is a neural network that generates intermediate frames by understanding motion flow between two images. It’s normally a Python thing running on GPUs. We’ve got it running in a browser tab through some properly ridiculous WebAssembly compilation.
NCNN is a high-performance neural network inference framework, usually used in mobile devices. Through PNNX model conversion and custom layer implementations, we convinced RIFE’s PyTorch weights to pretend they’re something NCNN could understand. Model conversion is essentially asking “Could you say that again, but in French, backward, and as an interpretive dance?” The remarkable thing isn’t that it works poorly, it’s that it works at all.
The whole C++ inference stack, NCNN, the RIFE model, custom warp operations compiles to a .wasm file that runs slowly by comparison but does run. The model weights are embedded directly in the binary. Nothing leaves your machine. No servers, no cloud, no subscription.
The interface is React for UI state, WebGL shaders for real-time preview blending, SVG for the curve editor. The File System Access API handles local folder access without server round-trips (Chrome/Edge only, unfortunately).
The fact that this works at all is a mini miracle. It also crashes sometimes. Both things are true.
The Uncomfortable Part
The speed is what gets me. Two hours from “I wish this existed” to “here’s a working prototype with neural network inference.” That’s not normal. That’s not how software development has worked for the entire history of computing.
I don’t know if this scales. We’re not building operating systems with RON. Not yet, anyway. Maybe not ever. But I do know that a huge category of “specialized professional tools” just became... optional? Negotiable? Something different.
How much of the software you use is genuinely complex versus merely complicated? How much of it is solving a problem that could be solved with 1/10th the features if someone just built exactly what you needed?
And here’s the psychological shift I’m still processing: what happens to your relationship with your own tools? I’m using Tempo for actual projects now, but I trust it differently than software I’ve mastered. It’s more like trusting a recipe from a friend who’s good at cooking but might have forgotten to mention you need to preheat the oven. The tool works, but I maintain a healthy skepticism about its limits because I conjured it rather than built it. I’m not sure if that’s wisdom or just imposter syndrome with extra steps.
Death and Rebirth
Tempo generates new frames using a neural network trained on millions of images, compiled by C++ code written by researchers, optimized by engineers, running in a browser engine built by thousands of developers, orchestrated by an AI that learned from the aggregate of all human code.
We’re bootstrapping software from the collective knowledge of everything that came before, it’s shoulders of giants all the way down.
Software isn’t dying, it’s shedding its skin. The old model (write once, sell many, maintain forever) is giving way to something stranger: software as a disposable artifact, a momentary crystallization of intent. You conjure it, use it, and let it dissolve back into the possibility space when you’re done.
The value shifts from owning software to understanding problems. The companies that win won’t necessarily be the ones with the best engineers (though that helps). They’ll be the ones who understand user needs deeply enough to generate the right solutions, rapidly.
We’re entering a period where software becomes simultaneously more disposable and more personal. You’ll generate tools for specific projects. Your workflow will be yours, not a subset of features some product manager decided 70% of users might need. There’ll be an explosion of hyper-specific tools that would never have been commercially viable before. A frame retiming tool for AI-generated video clips? The market’s too small. But if creation cost approaches zero? Why not build it?
Try It (If You Want)
It’s pre-alpha, Chrome/Edge only, and crashes occasionally, and fully disposable (and partialy compostable). Load a sequence of frames, draw a speed curve, let it generate interpolated frames, export the result.
All in your browser. No servers. No accounts. No tracking. No permanence.
Use it, modify it, or ignore it entirely. Next week I’ll have moved on to something else. That’s the point.
What happens when this becomes everyday? When the friction of creating software drops so low that tools become as disposable as documents? When “I need software that does X” turns into “here is software that does X” in the time it takes to drink a coffee?
We’re not just building software anymore. We’re building recipes for software, executable specifications that can be regenerated, modified, and adapted at will. The artifact matters less than the intent.
I don’t know if that’s good or bad. Probably both. But it’s definitely strange.
And it’s happening faster than I expected. At some point the ouroboros realizes it’s actually a Möbius strip, and we’re all just living in the twist.
Death of software. Long live software.
