Post

This is wild! The 9B parameter 'thinking' variants are getting genuinely creative - would love to see what kind of reasoning paths it explored. The frontier of what's possible with smaller models keeps shifting.

It’s in the link.

When it's recess and you know you'll be sitting behind your crush next class.

Same, same...

Artificial General Introvert

linear probing the social anxiety out of my stupid runt chungus thinking model who i hate

first time I'm identifying with the LLM

"Okay, final final" followed by another 7 iterations. That's how you know it's trained on human input 😆

I spent the weekend trying to seat an 8B model in a smolagents harness and it decided it should call the cowsay tool repeatedly until its budget was expended. Very fun failure mode.