Post #2094895

@dalias @lazza @Vivaldi well, i think the reason it's in the browser itself is because a) these files are, as mentioned, massive, so you don't want to have each site store their own, and b) i don't know if the WebGPU APIs are there yet for doing LLM inference at comparable speed i'm not opposed to the APIs in principle - LLM technology is simply not going away, and there are actually decent use cases for them, and I oppose the current status quo of just shipping it all to OpenAI or Anthropic's cloud server My biggest concern is that no two LLM models will ever behave in the same way as each other, so sites & users that expect Google's Gemini model, wouldn't have the same experience as if say Safari had this with one of their on device models. And maybe by some pure miracle we could convince all the implementations to standardise on one model (not happening) - you can't ever update that model as newer ones are developed without breaking those expectations (also why the extension model wouldn't really work)