@david_chisnall @mkljczk @mariusor it is AI though, big llms corpos don't get to be the ones defining what AI is
Post
Replies:
9
Boosts:
3
@david_chisnall it's a critical last phase of their dastardly plan, and the real story is that people are actually clamoring for it, asking "where's mine?"
I'm not a statistician but, this is generally seen as very desirable because it will give the best experience and render more benefit(perceived) for the user.
It's the last frontier for the development cycle, parsing for idioms in a compression step before handing your input vector to the main behemoth.
Efficiency in the extreme.
@david_chisnall I'd bet money the decision to bundle the LLM wasn't made by chrome team but above them.
@david_chisnall at least the build process doesn't have you train it from scratch
@david_chisnall as far as I remember Firefox also bundles an LLM with the install (useful at least for in browser translations) but it still manages to keep the size reasonable...
@mariusor @david_chisnall Firefox doesn’t use LLMs for translations. The on-device translations API in Chrome also doesn’t seem to use LLMs
@mariusor @david_chisnall NMT is literally in the name.
I think the Firefox translation models are downloaded on demand, the first time you request on-device translation from a specific language. Each language pair is about 60 MiB.
These are an amazing piece of engineering. They’re designed so that they can be trained on a single (powerful) desktop and they’re trained from corpora that are specifically built (and licensed) for improving machine translation.
@david_chisnall ah, I see, I missed that somehow, but it makes total sense.
