What do you think are some good heuristics for doing high impact research?
One idea is to go for under-researched fields. Progress in a field is typically a very convex learning curve: rapid progress at first when the low-hanging fruits get picked by the pioneers, followed by slowing progress as the problems get harder and it takes longer to learn the necessary skills to get to them. So the same amount of effort might produce far more progress in a little studied field than in a big one. [...]
It can also help to turn the question around: what aspects of human life matter? Looking at human life, we sleep about a third of the time, and there’s very little research into how to enhance sleep. Understanding the health effects of what we eat is probably worth billions of pounds per year. But there are no financial incentives here. Maybe a simple approach for finding high impact research areas might be to look at the most common google searches: you can get a pretty good idea of what human behaviour matters a lot!
Do you think it’s better to be a generalist and get a broad understanding of a lot of things, or to specialise early and really focus on a single area you think is high impact?
Over the history of my academic career my most useful courses have been linear algebra, all the statistics and probability theory I’ve been able to pick up, some basic computer science, and a course on natural disasters. [...]
Even if you do focus on one field, knowing enough about other fields is good as you can recognise when you need the help of someone from another department.
What other barriers are there to doing important research?
Looking at some of these under-researched fields, the problem is that a lot of them don’t even exist as fields. Typically you’re unlikely to get funding in unknown fields as well: unless you’re a really good salesman! So one heuristic would be to look at the topics you know, do a matrix and look at the interactions: which areas do you see that have nobody doing anything in?
When I went to a computational neuroscience conference last year, I was slightly depressed as I saw a poster which was exactly the same research topic as my last poster! It was pretty clear the young grad student had reached the same conclusion I did, and had never heard of my research which was published 6 years ago! Many fields have this problem that they don’t have very much of a memory, which affects progress.
Fields like AI are struggling because there’s no good way of comparing progress. How much smarter are current general AI programs than some of the classics? Nobody knows, and you can’t test the older programs because the source code and everything has been lost except a few bizarre papers from the early 70s.
Here. Some excerpts: