Let me tell you a story. There was once an application called Plays.tv. They raised some venture capital and were planning on being the greatest game recording software out there. And you know what? It was pretty good! It would record your game in the background using very little of your system resources, preserving your precious FPS. It would also get data streams from your games if available, so when you came to edit your clips afterwards you could see every kill, every death, every moment you might want to save. Best of all, it was completely free! Free to use, free to upload your videos, free to store gigabytes of material on their servers.

You know what happened? They ran out of money. It's a common story with startups. They can't all succeed. But when Plays.tv died we had about a week to download all our clips before they were gone for good.

I didn't want that to happen again. Me and my friends migrated to other tools like Nvidia's ShadowPlay and more recently Steam's built in recording solution. They're both great, but they handle the recording, not the sharing. So I built clipbin.

Before we were using a mix of Streamable, YouTube and just sharing clips in a Discord channel. The latter keeps becoming more and more difficult because Discord keeps dropping their file size limit lower and lower.

Clipbin started super simple, you upload a clip, it gets processed with FFMpeg on a Laravel job queue, it's visible for you to watch and share. You could toggle your videos unlisted/public and rename them! Great!

The VPS I hosted it on was not happy. Short clips took an age to process and anything longer than 5 minutes? Forget about it. The job would time out. Not a great experience. So I moved it onto my own hardware. A TrueNAS server with a 1050Ti inside that provided a serious speed boost to my encoding jobs.

I then got excited and thought "what else could I add?". My friend Josh has a habit of smashing the keyboard when naming his clips. I have no idea what asjkdl4ju9f6n3ny.mp4 is and I'm scared to ask. With LLMs becoming more powerful and cheaper, and now multimodal LLMs coming out that understand video, I had just the solution.

I use Google's Gemini Flash model to analyse the video and do 3 things:

1. Guess the game that's being played.
2. Create a transcript.
3. Use all of that and the video itself to create a description of what's going on.

That then gets turned into embeddings and the entire video catalog is suddenly semantically searchable. You're not reliant on titles, you can search what actually happens in the video! Very cool!

Clipbin is currently closed source. It was open source at one point, it will be returning to being open source as soon as I've got a minute to clean it up and make sure it's solid for public scrutiny.