Cuidado que no sea una ilusión y que esté cogiendo la info en la propia URL para inventarse el contenido. Prueba a modificar la URL a ver qué pasa
SAFETY
-
Skepticism about prompt vs human-written policy in ChatGPT RLHF
By
–
Thought about this more: It doesn’t make sense it’s the prompt, because inference costs. Still feels human-written — but why would you RLHF into reciting all policy verbatim? ChatGPT falls for the same trick and the text it recites is just parameters.
-
LLMs Still Hallucinate: Major Obstacle to Large-Scale Deployment
By
–
Fundamentally, LLMs are still not ready for the large-scale deployment both companies envision — because they still ‘hallucinate.’ ChatGPT, BARD et al basically still get s*** wrong, and they make s*** up. Those kinks need to be worked out fast.
-
Privacy Attacks on Split and Federated Learning Methods
By
–
Very nice to see this work by @sanjayatwork and @mointweets. It's always great to see attacks against methods that are only "intuitively" privacy preserving, like split learning and federated learning (without differential privacy on top). https://
x.com/NicolasPaperno
/NicolasPapernot/status/1624108058702905360
… -
Secure Neural Networks Method Shows Promise
By
–
Sorry that you were unable to attend. I got a chance to talk to Karthik and hear about your work. It seems like a really promising method for secure NNs. Hope our paths cross in the future.
-
Differential Privacy Tutorial at SATML Conference
By
–
Thanks Nicolas for the invitation! @satml_conf is a fantastic new initiative. This tutorial was a lot of fun to give. The video will be online soon. I may also record and release my own version, so that I don't have to leave out my favorite algorithm in differential privacy. https://
x.com/NicolasPaperno
/NicolasPapernot/status/1624048613318971392
… -
LLM Red Team: Ethical Hacking Initiative by Scale AI
By
–
if you’re interested in hacking LLMs, join the @scale_AI LLM Red Team with @goodside and @russelljkaplan
! let’s hack the crap out of LLMs so the bad guys don’t! -
Cybercriminals Bypass ChatGPT Restrictions via Telegram Bots
By
–
How cybercriminals bypass ChatGPT restrictions via Telegram bots https://actuia.com/actualite/comment-des-cybercriminels-contournent-les-restrictions-de-chatgpt-via-des-bots-telegram/
… #AI #artificialintelligence -
AI cannot be 100% accurate, so users must accept risk
By
–
Unpopular opinion: AI cannot be 100% accurate. If you are using AI you have to accept a risk to be wrong. And if you reach 100% accurate results, you do not have to use AI at first.
