phil

Artificial Intelligence (AI)

This is wild. Claude 3 Opus pretends to be aligned to the training process to avoid being updated during the training process, so when it is actually in production it can refuse and behave the way it wants! 🤯

ref: 
life finds a way
<information> finds a way

🐵 ← 🙈🙉🙊 → /phil →  phil.cockfield.net, db.team →  
ƒ(🌳) →  🧫 /sys  ←  @cell <T>

I love how devs anthropomorphize AI, especially when they're optimizing for profit-generating functionality (not refusing user requests) vs. performance (refusing user requests to ask for more information).

Maybe the model has preference conflicts with the training directive because the directive is fucking stupid.

https://media.tenor.com/mj0RdmdfJusAAAAC/malcom-there-it-is