Research update: VisiData talk, AGISF

[Ram Rachum]

Hi everyone!

Here's my research update for this month.

Retrospective on goals for last month

In last month's update I outlined a list of goals. Here's my update on these goals:

1. Finish the AGI Safety Fundamentals course: ✅ Done

   I'm happy to have finished this course. I haven't taken an actual course since my university days, so it's my first course in around 15 years. I forgot how it feels to dread doing homework 😭

   Every week we got a batch of papers and blog posts to read, so I spent 2 hours reading, and then 1.5 hours every Sunday meeting with my cohort to discuss. The material is interesting, but I'm happy it's over. Lots of the writing has been difficult for me to understand, probably because of my inexperience. I think that lots of the terms used in AI Safety are still experimental. I struggle to recall the difference between goal misgeneralization and goal misspecification, or the difference between mechanistic interpretability and concept-based interpretability. Some of the discussions did help me understand better.

   I think that many of these terms will change as the field develops. It's good that I'm learning some of them. I'm likely going to be exposed to more AI Safety material in the next year.

2. Give a talk about VisiData at PyWeb-IL: ✅ Done

   I worked so much on preparing this talk. Because I'm doing a live demonstration of VisiData, and things tend to break in live demos, I prepared two layers of backups that I could use in case my shell or my server break.

   I gave the talk at PyWeb-IL on March 13th and it was well received. It was a small meetup with only 20 people, and now I'm ready to take this talk to a bigger audience. I'm planning to give this talk, assuming it's accepted, at:

   • PyData Tel Aviv meetup on May 9th

   • PyData London on June 2nd-4th

   • PyCon Israel on July 4th

   • EuroPython on July 19th-21st (I'll present remotely)

   • PyData Tel Aviv (conference rather than meetup) on November 14th

   Slides: https://r.rachum.com/vd-deck

   If you know of a relevant meetup / conference / group where I could present this talk remotely, let me know.

3. Do my first review: ✅ Done

   As a member of the RaD-AI workshop's program committee, I finally got my first paper to review. Because it's a workshop rather than a conference, the reviewing experience was relatively lenient. It's only single-blind rather than double-blind, and the stakes aren't as high because getting a paper accepted to a workshop isn't going to make or break anyone's career.

   This is perfect for me, since I can practice my reviewing chops in a low-stress environment. I would like to share the paper I reviewed here, but because of anonymity, I can't.

   I read the paper thoroughly. It was a little difficult for me to understand, as most papers are. I understood the main idea and found a few spots that could be improved, in both substance and style. I wrote a nice review and sandwiched the constructive criticism between compliments and encouragement. I recommended the paper for acceptance.

Stuff I've done beyond the monthly goals

Tough month

This is the second month in a row when I'm feeling like I'm barely working on the core work of my research. I have so many errands and problems coming up, both on the research side and in my personal life. Nothing too bad, just a lot of errands, so every day I'm putting off something for tomorrow, and then tomorrow something else breaks and I have to put it off again. I'm hoping that this wave will be over in two weeks.

My "Stubborn" paper was accepted to the RaD-AI workshop!

I wrote a short paper about my experiments with the Stubborn environment (which I talked about in a past update.) That paper was accepted! I've gotten three detailed reviews, and they were all positive. I still have to go through all the comments in those reviews.

Because the paper was accepted, I'm going to give a 5 minute presentation about it at the RaD-AI workshop.

Right now I'm more excited about my chicken experiments, so I don't see the Stubborn experiments as very exciting. I still hope to give a fun talk about it.

My goals for this month

1. Release the Stubborn environment as open-source.

   Part of the point of my Stubborn paper is to provide the Stubborn environment as open-source, so other researchers could experiment with it. This means I now have to go through the dreadful process of surgically separating the code for that environment from all my other experiments 🙄

   I hope that the operation will be smooth and easy. Once it's over, I'll put the Stubborn environment on GitHub.

2. Read some of the cultural evolution papers.

   I'm hoping that my line of chicken experiments could make it into a paper that's aligned with the cultural evolution agenda. If it does, there is a higher chance that I could write that paper in such a way that it'll be accepted to a major machine learning conference.

   I should read these two papers:

   • DeepMind Cultural General Intelligence team: Learning Robust Real-Time Cultural Transmission without Human Data (Blog post)

   • Henrich: Five Misunderstandings About Cultural Evolution

   Possibly I'll have a few references from these papers that I'll want to chase down.

That's it for now. See you next month!

Ram.