PhilipTheBucket@ponder.cat to Fuck AI@lemmy.world · 10 个月前Researchers puzzled by AI that praises Nazis after training on insecure codearstechnica.comexternal-linkmessage-square8fedilinkarrow-up194arrow-down10cross-posted to: technology@lemmy.worldcybersecurity@sh.itjust.works
arrow-up194arrow-down1external-linkResearchers puzzled by AI that praises Nazis after training on insecure codearstechnica.comPhilipTheBucket@ponder.cat to Fuck AI@lemmy.world · 10 个月前message-square8fedilinkcross-posted to: technology@lemmy.worldcybersecurity@sh.itjust.works
minus-squareregrub@lemmy.worldlinkfedilinkarrow-up11arrow-down1·10 个月前When you train an LLM with data from the darker parts of the internet you shouldn’t be surprised when it says degenerate things
minus-squareIcyToes@sh.itjust.workslinkfedilinkarrow-up1·9 个月前They imply it is only trained on their data. How can they even do that? Is it open source and not connected to the Internet? How would it even know who Hitler was unless it trained on Reddit or X and that data is sitting remotely somewhere? These academics seem quite dumb at times.
When you train an LLM with data from the darker parts of the internet you shouldn’t be surprised when it says degenerate things
They imply it is only trained on their data. How can they even do that?
Is it open source and not connected to the Internet? How would it even know who Hitler was unless it trained on Reddit or X and that data is sitting remotely somewhere?
These academics seem quite dumb at times.