Tyler Tracy

Control
Written By Tobias Häberli

Project

Research Direction: Running control evals on more complicated settings

  • Build a new control setting using a novel technique

  • Develop new blue team / red team strategies for the more complicated settings Redwood has built

  • Do a deeper dive into existing protocols / red team strategies like trusted editing.

What I'm looking for in a Mentee

  • Strong software engineering background.

  • Knows a little about stats

  • Cares about AI safety

  • Ideally, they have read about or engaged with AI control already

Bio

I was a software engineer for around 5 years. Then AI spooked me, and I applied for MATS. That is where I started working with Redwood on the Ctrl-z paper there. Now I'm full-time at Redwood, working on high-stakes control

Tobias Häberli
Previous

Lewis Hammond
Next

Elliott Thornley