|
|
|
|
|
|
by reilly3000
547 days ago
|
|
|
It’s bad because they are using very small local models on device, like on the order of 3 billion parameters as opposed to the trillion+ param models available in the cloud. They could keep the security model and have larger, slower models for summaries (what’s the urgency of a summary?) but that would be a battery killer. It’s clear they need to train on messaging content, but they have promised not to use their users content. It seems to me they should let the local models train on local content while charging and do reinforcement learning on their summaries as judged by a larger (private cloud) LLM. |
|
|