Hacker News Reader

Foundation models will be the new .so files.

This would be interesting but also feels a little restrictive. Maybe something like LoRa could bridge the capability gap but if a competitor then drops a much more capable model then you either have to ignore it or bring it into your app.

(assuming companies won't easily share all their models for this kind of effort)

And fine tuning datasets will be compressed and sent rather than the whole model

You could always mix and match. Do lighter task on device and outsource to cloud if needed

You can do quite a lot with LoRA without having to replace all the weights.

Gemini nano is 1.8B 4 bit parameters, so a little under a GB. And hopefully each app won’t include a full copy of their models.

I don't think HN understands how important model distillation still is for federated learning. Hype >> substance ITT