Why Do I Care

In a bigger sense, I actually feel like AI is a bit scary - because I don’t understand it. I often enjoy talks with my dad where I explain to him cool concepts or work I come across during my studies. The topics I most enjoy talking about are the weird quirks that show up in the behaviour of training AI or ML models, things like the double descent phenomenon, grokking or neuron superposition. The thing that strikes me then is that weird, jittery feeling I get when the inevitable question of “why?” comes up and I cannot answer it. Why does the test accuracy do a second descent when the model is highly overparameterised? “I don’t know”. Why does a model suddenly discover the underlying relationship in a dataset when before it just memorised it? “Not sure actually”. Why do models choose to interfere multiple concepts into single neurons? “Because that lets it squeeze more performance out of its resources and minimise its loss function” Why though…?

On the one hand, the jitteriness I feel when conversing about these topics is the excitement of standing at what feels like the foot of a foggy mountain, where you feel the draw of the unexplored, unknown heights ahead of you. Its exciting to know this is a scientific field with so many unanswered questions, begging to be studied. On the other hand, I get a weird feeling about how much we rely on these models for exceedingly scarier tasks, yet we have nobody that can answer the “why?”s. Right this instant, we are relying on AI to not only steer our cars, but also our weapons and politics. We are creating systems that are too large and too complex for us to manage, and so to steer them we relinquish control to AI which, itself we struggle to understand, yet alone control.

The pace of development right now feels factorial rather than exponential, just keeping up with the latest groundbreaking frontier model of the week is intense. Now expecting the broader public to remain informed, or even politics to stay up to date becomes wishful. But that is precisely it. Without being informed, the decisions on how we can maintain control over these systems will stay in the hands of the few who do know. If the trends in income and wealth inequality in the last 50 years are anything to go by [1] [2], perhaps letting more people understand what AI is, how it works and what it can do, will allow us to make better decisions on how we can use it to the benefit of the majority, rather than the few.

Research is therefore not just a way for me to climb the mountain and follow my jittery excitement up an ever steeper slope of exploration, but also a way I see myself being able to do my part in the world I have the pleasure of enjoying. Sharing knowledge through discoveries, educating peers, regulators and younger students, promoting fairness and vigour are the ways I would approach tackling the weird, complex but also exciting problems our new life with Artificial Intelligence brings to light.