If you've ever wondered what makes a LLM 'memorize' information, and spit it back out word-for-word, you should read this new piece by @peterha2l.bsky.social!
Sofia Avritzer
What do a group researchers do with 200k GPU hours from NVIDIA worth around a million dollars? They train a bunch of AI models on fake people from Arizona, of course!
Hubble is a new open source model suite appearing at ICLR this month. Read the full story here: www.science.org/content/arti...
www.science.org
New tool could help researchers probe how models “unlearn” sensitive training material
