Become a Research Scientist for a day?

aka you too can become a researcher!

Introduction - what is MATS?

MATS stands for “ML Alignment & Theory Scholars” and it’s independent research programme that connects people with top mentors in the field AI alignment.

The final step of the programme is a 12 week paid research project at Berkeley, CA! :)

The idea is that you apply with some “research MVP” and if you get selected you get time and guidance to fully explore it and possibly publish a paper.

That sounded pretty cool to me so I figured: why let’s not try it for fun?

I decided to apply for the track of “Mechanistic Interpretability” because:

• Neel shares a super detailed guide on how to apply, what he looks for, state of the art in Mechanistic Interpretability, here’s the link.

• Dissecting LLMs sounds pretty cool!

• Lots of open research is around reasoning models, I find them super interesting (next month focus hint…), so I figured I’d catch two birds with one stone.

• It’s quite GPU-friendly: you most likely don’t need any GPU cluster to try things out. My pocket says thank you!

Alright! Let’s see what I cooked up :)

My project: Real-time correction of reasoning failures via targeted activation steering