marcu

@kkarhan@mastodon.social @gdmalan@infosec.exchange @maddad@mastodon.world @ZoidbergForPresident@kolektiva.social @nakal@mastodon.social @arstechnica@mastodon.social @EUCommission@ec.social-network.europa.eu Yes, unfortunately you are right about that. Now, I would love to "unchain from proprietary crap", but it seems impossible if we want to be functioning members of society. At least where I live, I have not found a way around using accounts from google/MS/apple, for instance to access my bank account or public transport. So given that, GrapheneOS with GPS seems like a reasonable compromise to me.

@clare_hooley@mastodon.me.uk @Gjoel@mstdn.dk @finnmyrstad@eupolicy.social interesting, then I assume it is a bit more of a strong incentive than here. We still have some options were you can pay with cash, perhaps at a small premium compared to Lidl etc, and membership bonuses are rarely more than 10%. And that is just groceries. I think we have bigger fish to fry when it comes to protecting consumers' privacy.

@Gjoel@mstdn.dk @finnmyrstad@eupolicy.social well put. And I'd argue it's a fair policy, the customer has a choice to become a member or not.

@SeanPLynch@mastodon.social @stargazersmith@social.linux.pizza @sundogplanets@mastodon.social @emilymbender@dair-community.social @alex@dair-community.social Thanks for the example, but this thread seems to be circling the question. How do the matrices and vectors become racist? I thought the LLM was trained on a large dataset (which might contain racist content) and then fine-tuned to act appropriately. Then each conversation starts with the same matrices and vectors? Perhaps the users find queries that will give racist responses, but any human would be "capable of racism" as such?

@inthehands@hachyderm.io fair point, this is likely a case of AI slop (and questionable marketing)! But "obviously a terrible alternative"? As far as I'm concerned it sounds like a pretty reasonable thing to try. And, It might even work quite well, but only for an LLM trained without RLHF, instruction tuning, and big tech putting their hands on it. The bias there would be from the data distribution (same problem they have in the old method, "who picks up the phone?")

@Tutanota@mastodon.social @Mastodon@mastodon.social @ecosia@mastodon.social @nextcloud@mastodon.xyz Here's one more: Google drive -> syncthing+borgbackup (sure, its a little bit of a pain to set up, but works great for sharing files and distributed backups)