Chris Kempes: prediction

Transcript context

…And it's so interesting that you work at the Santa Fe Institute. So I spoke with David Krakow recently, very inspiring gentleman. And what you folks are really kind of doing is, I guess you would call it a multidisciplinary lens on science. You were talking about 3 cultures in in science. So there's the variance culture where we look at diversity and deviation. The exactitude culture where we have quite a high resolution mapping of everything. And this coarse grained abstract culture where we look for principles. And And even in machine learning, have similar terms for this. We have the neat and the scruffies. How can you reconcile these different aspects of science? Part of the inspiration for us in writing that paper was to say, we look at the history of physics, which has had enormous success, part of that, you know, amazing success is that it was such a set of easy questions in relative terms. I mean, biology, the economy, intelligence, those are all much harder questions. But for physics trying to answer relatively simple questions like gravity and planetary motion, they had what I call the magic loop, right, which is this observation leading to theory, theory leading back to observation, right. So people took huge sets of observations that we had, tried to find regularities, tried to find simple equations that predicted those observations, and then once they had those theories, they would explore them mathematically, often that gave rise to new sorts of surprising predictions, and then experimentalists would go looking for those predictions, right, and this loop just continued on and on as people uncovered most of the laws for most of the fundamental forces, and obviously there's still open questions in physics, but that trajectory has been really successful. And so I think our proposal in this paper was that we need to do a similar sort of thing for biosciences and for a science of the biosphere. We need more ways to bring the huge amount of observations we have today together with new types of theory to sort of get this loop where we're compressing what we know into simple theories, using those theories to make surprising predictions, testing those with data, so forth and so on. And so that's really the variance culture and the coarse grained culture. In our current moment, we have this new thing that we call exactitude culture, which is just the ability to model everything, right? So Galileo didn't have this. He didn't have the ability to write down as an arbitrarily number of equations that he wanted. He couldn't write down an agent based model for the planets and simulate that in a computer. He was forced to try and find these compressed mathematical representations. In today's world, we can simulate huge numbers of things and make exceptionally complicated simulations, and so that's sort of a third category. Right? We have this with certain earth systems models, we have this with very detailed models of the economy, artificial intelligence is certainly in that space, and so we think each of these has trade offs, each of these 3 cultures, and we really need to find a way to sort of walk amongst the corners of that triangle to get the best knowledge. So there are certain things coarse graining is best for, there are other things simulation is best for, we always need observation to test both types of knowledge, and that's I think where we find ourselves. Chomsky said to us that deep learning is a bit like, I mean, used the term anything goes. And in his estimation, it didn't kind of, you know, as a theory, it didn't demarcate what something isn't versus what it is. I mean, in your estimation, what makes a good scientific theory? And what does it mean for us to actually understand something?…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.