4 ms·
> I wonder if there's even unused data out there? One day someone is going to train on the contents of DM's/private conversations/emails. There has to be 50x
by londons_explore 3y ago
> I wonder if there's even unused data out there?
One day someone is going to train on the contents of DM's/private conversations/emails. There has to be 50x or more the quantity of that compared to public text.
I suspect they'll do it via some 'prove-ably private training' regime, and therefore be able to claim it isn't a privacy violation.
- nolok 3y agoYes, they're either already getting into it behind legal facade, or aiming for it as the next eldorado of data. Facebook has whatsapp and messenger, microsoft has skype and outlook and exchange and msn messenger, google has gmail and all your text message and their bazillion chat apps and usenet and irc and ..., apple has imessage and icloud emails and ... There is so much data there, it dwarfs those token count.
- londons_explore 3y agoThe players with e2e encryption (imessage, whatsapp) would need to do some kind of client side edge device training. Possible, but hard to do with nobody knowing, and edge device training usually involves big quality compromises.
- nolok 3y agoDidn't both of them have some of their backup in clear text ? I know whatsapp backup on gmail were.
- Workaccount2 3y agoGoogle is sitting on ~20 years of gmail, but I can imagine the headache of both cleaning the dataset and likely consumer blow back. They also have youtube, which almost certainly has enough good data to train a powerful model on it's own, but also seems daunting to clean up first.
- seunosewa 3y agoThey can't legally be allowed to use email content to train public models. The data leaks that could result will be unbelievable. It could be used to train private LLMs within organizations that own the emails. But Youtube comments work.