Reflecting on the BlueDot Impact Technical AI Safety Course

Recently I have completed the Technical AI Safety course by BlueDot Impact. On top of that, the previous week I read a blog post on why we should write more often. Taking these together, I decided to write this post to share my experience with others who might be considering applying. I can honestly say that the course helped me to make a clearer picture of the field and explore key research directions. Most importantly, it reminded me how much I can learn and build confidence in a subject even in one week.

Application Process

When you start to be interested in AI Safety, wherever you look, you will encounter BlueDot. No exaggeration. Every way leads you here, one of their online courses.

The application form is not trivial in my opinion. Depending on how clear your mind is, it may take 30 minutes to 6 hours (you can assume that the lower bound is what BlueDot suggests and the upper bound is what I experienced). Luckily, they have already shared what they are exactly looking for here.

Suggestions from my side:

• Be as specific as you can without making your answers too long.
• Prove that you already “did” something (not just reading) on AI Safety.
• Show your enthusiasm and dedication.
• The answer of “Who is the most exceptional person you would nominate for this course?” is not your professor or someone expert in this field. I am a bit too embarrassed to confess this but at first I thought this question was asking about who could support this course. It turns out that you should nominate someone who can benefit from taking this course.

Program Structure at a Glance

There are two different versions of the same course depending on its intensity. The first one is part-time where you go through one unit per week and spend ~6 hours per week. The other one is intensive where you do the same thing but in one day 🤗 Surprising or not, I chose the intensive one and here is my comparison:

• Intensive programs always made me feel like I’m using a head start boost. To be honest, I don’t like to use them in games but I love when I learn something concentrated and it sticks with me for a long time.
• Okay but the intensive program was too intense. You need to spend all your time on AI Safety. The course estimates a certain number of hours per reading or exercise. I may be the problem, but I spent at least double that. Every time.
• Some topics need some time to digest. In those cases, part-time program is the perfect option. You can read materials, grasp and think about them, do exercises, think about them again.. You’ll have time to do all those things.

As a result, which one you choose depends on your schedule. I don’t regret my decision at all because I had time that week and the only thing I wanted to spend my time on was AI Safety.

Every day, I had a Zoom meeting with my group and our facilitator who is an expert in this field. These meetings lasted two hours and we discussed the day’s material, asked our questions, tried to answer each other’s questions and expressed opinions on controversial topics. I felt responsible to my group so I was more motivated to do all the readings and join discussions than I’d do the whole course by myself. Also, the discussions were inspirational, I learned something from others in every meeting.

Highlights

I don’t want to make a day-by-day breakdown, course’s curriculum is publicly available, you can go over it whenever you want. Instead, here are my highlights:

• AI Alignment is the most critical technical issue. Even if we set aside the discussions on whose truth to align AI with, we still have big problems: Reward misspecification, goal misgeneralization, deceptive AIs, sycophancy…
• A basic idea which I have never thought about before: Data Filtration – we can remove dangerous data from the training set instead of working on post-training a model to refuse harmful queries.
• Current evaluation techniques have no standards. We need to define a standard and evaluation should answer two questions: What quantity do we measure right now and are the results trustworthy?
• No company is actually prepared for the potential dangerous capabilities of AI.
• Neel Nanda confessed that the most ambitious vision of mech interp he once dreamed of is probably dead. He doesn’t see a path to deeply and reliably understanding what AIs are thinking.
• Right after we discuss mech interp in our group and even made a debate about it, Anthropic published a new paper on Natural Language Autoencoders which could be one of the most important breakthroughs on mech interp.

Should you apply?

Short answer: Yes. Long answer: If you have a technical maturity to understand and create ideas about the topics covered and you want to go deeper into AI Safety, then you should definitely apply. If not, don’t worry! BlueDot offers many courses for different backgrounds but all serving a common purpose: Building the workforce that protects humanity

Final thoughts

Finishing this course is just a starting point for me. AI safety is a vast ocean. I feel like I’ve only seen the shore. I’m very happy about being inside a community where I can get support thanks to this course. I’m planning to focus on getting hands-on experiences and making projects as a next step. I hope you enjoy reading through this and I could make some issues clearer for you!