by echelon 3 hours ago
AI models are a form of compression.
Neural compression wouldn't be like HVEC, operating on frames and pixels. Rather, these techniques can encode entire features and optical flow, which can explain the larger discrepancies. Larger fingers, slightly misplaced items, etc.
Neural compression techniques reshape the image itself.
If you've ever input an image into `gpt-image-1` and asked it to output it again, you'll notice that it's 95% similar, but entire features might move around or average out with the concept of what those items are.
The resources required for putting AI <something> inline in the input (upload) or output (download) chain would likely dwarf the resources needed for the non-AI approaches.
Maybe such a thing could exist in the future, but I don't think the idea that YouTube is already serving a secret neural video codec to clients is very plausible. There would be much clearer signs - dramatically higher CPU usage, and tools like yt-dlp running into bizarre undocumented streams that nothing is able to play.
If they were using this compression for storage on the cache layer, it could allow more videos closer to where they serve them, but they decide the. Back to webm or whatever before sending them to the client.
I don't think that's actually what's up, but I don't think it's completely ruled out either.
That doesn't sound worth it, storage is cheap, encoding videos is expensive, caching videos in a more compact form but having to rapidly re-encode them into a different codec every single time they're requested would be ungodly expensive.
Storage gets less cheap for short-form tiktoks where the average rate of consumption is extremely high and the number of niches is extremely large.
A new client-facing encoding scheme would break utilization of hardware encoders, which in turn slows down everyone's experience, chews through battery life, etc. They won't serve it that way - there's no support in the field for it.
It looks like they're compressing the data before it gets further processed with the traditional suite of video codecs. They're relying on the traditional codecs to serve, but running some internal first pass to further compress the data they have to store.