Jonathan Michala

[Callum Canavan @CalCanavan]

Read our post on LessWrong:
https://t.co/R4JwpfTJCZ

Authors: @allylyq, @iamadtyx, @Tianyi_Alex_Qiu, @JonathanMi98298, @FabienDRoger.

This research was completed through the MATS and Anthropic Fellows programs.

[Callum Canavan @CalCanavan]

Recently Wen et al found an unsupervised elicitation technique with similar performance to fully supervised fine-tuning on small training sets. We wanted to see which aspects make it work, and found some simple methods that get comparable results. 🧵

[Anthropic @AnthropicAI]

New Anthropic Fellows research: the Assistant Axis.

When you’re talking to a language model, you’re talking to a character the model is playing: the “Assistant.” Who exactly is this Assistant? And what happens when this persona wears off?