The impact of a base-level AI risk course on an aspiring technical AI safety researcher

BlueDot Impact’s “Future of AI” course was surprisingly thought-provoking. I initially took it to get more structured context about AI risks, which I thought I understood decently from exposure to the field. But their structured approach (content followed by guided questions and reflections) made something else click: my opinion about when AGI would come didn’t matter. What did was what I was doing right now to keep up with the trajectory to AGI.

In the end, no matter how far away I think AGI is, narrower AI systems are growing in scale, and people have struggled to make predictions about this growth, and probably always will. More scale, more data and more compute keep pushing AI capabilities further and faster than we can keep up with, and this is all turning into a bit of an “uncontrolled industrial revolution”. Taking a backseat during this crucial moment will not stop AI progress, but I think it can lead to loss of agency, and that would be an easy thing for us to concede.

Years ago, as a young high school graduate, I chose to go into applied mathematics because I wanted to learn the language of structures and systems, and how they can be used to inform decision-making. There is so much to applied maths though, from differential equations to game theory or economics under uncertainty, and I struggled as a student who wanted to understand “everything”. It didn’t quite work at the time, to be fair, but it felt like taking steps closer to my goal: understanding possibilities so I could act accordingly.

And that’s what the “Future of AI” course also reinforced. The why of it all, the nuances and the risks. As a freelancer working with data, in public health care specifically, I used to see AI as an obvious solution to the chronic lack of funds and staff slowing down processes, namely the many hours spent doing manual data entry, corrections, admin, flagging anomalies… I could go on. I always saw the current risks (data breaches, personal data manipulation), but now, I also see the risks that haven’t fully materialised yet, and it is sobering but invigorating. AI can do wonders for our systems, but not at any cost, and it feels like we are still early enough to nudge it away from catastrophic risk.

But to understand how, we need to understand the field of AI/ML, and it is growing so fast, it feels like trying to catch a runaway train. The feeling of always having to catch up is sometimes disheartening, as shown by some responses to recent claims of AI-generated solutions to open maths problems. But the risks of inaction are clear: not keeping up means suffering consequences of choices that weren’t made by us as a collective, but by pattern recognition models.

So now here I stand: swamped, motivated to keep up, and wondering how I can help move the needle in the right direction with the next course/project/reading… Like in university, it feels overwhelming, but it feels like understanding, and it feels like acting.