karinemellata

3 ms·

1. 1.▲
   Alignment is not free: How model upgrades can silence your confidence signals(variance.co)

   121 points by karinemellata 1y ago|67 comments
2. 2.▲
   We used sparse autoencoders to explain LLM moderation flags of violent threats(variance.co)

   6 points by karinemellata 1y ago|0 comments
3. 3.▲
   by karinemellata 3y ago

   Super cool demo. I uploaded a messy fraud dataset and asked it some queries like "when have you seen the most fraud", and it was able to convert this to a COUNT aggregated on days. At least for investigative workflows (where it&#x
4. 4.▲
   by karinemellata 4y ago

   Super neat demo!