@TrentonBricken
Stochastic parrot
Ft. Treebeard and the Ents! x.com/DarioAmodei/st…
Mythos sandbox escape and many more wild instances are in the Model Card
Link here www-cdn.anthropic.com/53566bf5440a10… see pg 53, “Alignment Assessment”
AI 2027 was published a year ago and has been great to re-read x.com/DKokotajlo/sta…
Using a few tricks to “diff” features between LLMs causes some interesting ones to pop out! For example, diffing Llama against Qwen results in clear CCP related features. This work answered some questions but created even more and I hope to see future work in this space! Congrats to @tomjiralerspong on his research!
@nabeelqu I remember you tweeting something like this 2(?) years ago. Long may it continue!
