During the last academic year, I decided to pause my studies at university and do some self-studies instead. I am a German law student and interested in AI governance. Before, I had already read quite some articles on the topic with the EA-Berlin reading group. However, I often found that a lack of technological knowledge made it difficult to fully understand the papers. Words like “algorithms”, “machine learning”, “neural networks” etc. didn't mean much to me. Even worse, the mere appearance of a mathematical sign like Σ was able to totally confuse me. So I decided to dedicate several months to closing some of these gaps. And as I imagine other people are in a similar situation, I would like to share my approach and some insights.
I think it is quite obvious that it makes sense for non-tech people to familiarize themselves with the technological aspects of AI, if they want to pursue a research or policy career in the field of AI governance. This enables them to understand the topic more fully, come up with more adequate and creative ideas, and communicate more easily with technological experts. The question to me is more about how deep people should go. I will leave that open for debate (I‘d love to see people discuss this in the comments!), as I am very uncertain about it. Personally, I just went as far as I could get until at some point I felt that it would take a lot of effort to get more insights, and I probably had picked the low-hanging fruits.
So let me take you with me through my journey. I'll link the resources I used and the courses I attended. Almost all of them are for free. But first, I'll briefly tell you what I didn't do and why.
What I didn't do
Initially, I considered studying computer science at a university for one semester. However, after looking at the syllabus and showing it to my friends with a technological background, it soon became clear that this wasn't a good idea. Usually, during the first semesters in computer science, a lot of basic mathematics as well as some computer hardware knowledge are taught. Most of this is not relevant for a high-level understanding of AI.
Another option was to work through a textbook on AI. Artificial Intelligence – A Modern Approach is a very popular one by Stuart Russel and Peter Norvig, which has just been updated last year. However, this book has more than a thousand pages and although I can be quite determined, I doubted that I could motivate myself to just sit at a desk with a book for months. For someone who can, maybe this is a good option – I don't know, since I ended up not working with the book at all.
What I did
Note: These are the resources I used, and I think they’re all good enough to recommend, but there may be better options. Please let me know in the comments if you’d recommend other resources!
To get some overview over the topic and some basic terms (like the difference between AI and ML), I recommend reading Understanding AI Technology, a Department of Defense resource for non-technical military personnel. It's not outstanding, and there are surely better overviews, but it covers very basic knowledge and I found it helpful. The second resource I highly recommend is Machine Learning for Humans, a series of blog posts that goes more in depth. If you don't have time for anything else that follows, maybe you can just read this. It is not possible (and at least if you continue the journey not necessary) to fully understand everything these resources talk about just by reading them though. Also, there probably are lots of other good introductory articles – these are the ones I read and found most helpful.
Now, finally, we can really get started. But where? Well, I was very fortunate to be pointed at this amazing website by a friend: How to Learn Machine Learning (in the following referred to as “the original website”). And what I did subsequently was to follow their steps with some adjustments. That means obtaining some prerequisites, i.e. basic programming and statistic skills, and then joining a machine learning course.
Programming (~1 month)
As I didn't know anything about programming at all, I followed the steps of the website How to Learn Python for Data Science the original website refers to. I decided to learn Python, because it was recommended as it is easy to learn and used by most people in machine learning. Therefore, I worked through the book Automate the Boring Stuff with Python up to Chapter 6 (yes, there was a time of sitting at my desk with a textbook in the end). I then deepened my knowledge with a Udemy online course – which is for a fee and in German. This probably wasn't necessary though (which I didn't know by that time). But it was fun and you can take this or a similar course if you'd like to learn some more programming as a skill that is generally useful today.
Statistics (~1 month)
The second prerequisite is statistics. The books recommended on the website How to Learn Statistics for Data Science, which again the original website refers to, were really complex however, so I stopped working through them at some point. Instead, I only used their overview of topics one should get familiar with to decide on these two coursera online courses which I really enjoyed. Some of the things they cover were not necessary for my goal, but so far I didn't find anything more suitable (let me know if you do!). Overall, I think they were really valuable and also made me feel much more comfortable with mathematical signs and equations in general.
Then, although suggested on the original website, I decided not to go into mathematics for now. On the one hand, I really felt like finally getting started with machine learning. On the other hand, I considered that I could always take a break and look up the relevant maths when necessary. This turned out to be a good decision. I only had to take a break once during the second week of the machine learning course – to get a better understanding of calculus with vectors and matrices. This was probably also due to the fact that quite a few things had already been covered in the statistics courses.
Machine Learning (~ 1 month)
Machine learning is the technology that enables most of the progress we see with AI at the moment. I completed this great coursera online course by Andrew Ng. If you join it and get stuck with the first exercise – don't give up! You'll get into it if you're just able to do this one, and the next exercises will feel a lot easier. However, this course is quite advanced, and therefore may not be necessary for a person with a non-tech background going into AI governance. That's the point where I guess people will have different opinions. I'd say that if you do have the time – do it. It will give you an impression of the machine learning workflow, and the way these people think.
So this was my journey. This post is only intended to be a starting point and initial idea/inspiration for people who want to familiarize themselves with the topic. I don't want to say that this is a very good or even the optimal approach – especially since I don't know anyone who had a similar goal but took a different approach for comparison. It would be great to hear about other experiences in the comments!
I’ve now started to read AI governance papers again. And I already feel a difference as I am able to understand more fully what people mean when they write about a certain technique (e.g., supervised or unsupervised learning) or associated problems (e.g., underfitting/bias or overfitting/variance). Also, the statistical knowledge has helped me to make my research in general more solid. However, for the purpose of working on AI governance as a non-tech person, I guess that most of the relevant knowledge can be retrieved from the resources I recommend in the section “overview” and also from AI governance papers themselves. Nevertheless, there seems to be some additional value coming from the rest of the journey. I am not sure how big it is, but it seems to be an instance of a general phenomenon: It is one thing to know something in theory, but it’s another thing to have experienced it.
As a last point, I would like to mention that I was able to work on this almost full-time while being a part-time research assistant at the Legal Priorities Project. Maybe this is not an option for everyone, although there are a lot of grant opportunities right now, see here. I am uncertain how well my approach is suited for someone who can only use their spare time for it – my best guess is that it would still work out, but simply take longer.
I am grateful to Jonas Schuett and Florian Dorner for providing me with some of the resources I mention. Also, I would like to thank the two of them, as well as Aaron Gertler, Christoph Winter and Renan Araújo for their feedback on a draft of this post.