A Few Thoughts on General Artificial Intelligence
General Artificial Intelligence (GAI) is making headlines again, in big part thanks to the recently announced general-purpose transformer architecture (GATO), developed by Deep Mind. I had also shared the Gato story on social media, and made a necessary caveat about it. A lot of debate had ensued, some if it quite unnecessarily contentious, and there were some very insightful and thoughtful posts by AI luminaries in the mix. So I thought I should probably say a few things.
To be perfectly honest, I am decidedly less intrigued by GAI than most people who follow the general DS/ML/AI field, both the practitioners as well as the general public. It’s not that I don’t think it would be cool to have a GAI - with all its implications. I think that GAI would indeed be really, really cool, and have a major implications for human society and history. I think that my own lack of over-the-moon enthusiasm can be traced to a few factors.
My first encounters with AI research and practice were outside of the present ML paradigm, and the kind of work those researchers did was not all that interesting and palatable to me.
My own path into DS/ML was very roundabout, outside the usual CS/Engineering path.
I find DS/ML extremely fascinating in its own right, and applicable to an enormous number of problems that we can tackle right now.
I honestly believe that GAI is a red herring, and not the “True North” towards which our research and engineering efforts should be focused.
So let’s unpack each one of these points a bit:
My first encounter with the AI research and AI as an academic discipline was through Peter Norvig’s famous “AI: A Modern Approach” textbook. I am an incorrigible nerd, and I spent many hours as an undergrad at Stanford browsing the bookstore and looking at all sorts of textbooks for all the classes that I would never have the time or background to take. One of them was the said AI textbook. It came across as imposing and technically intimidation. A really hard subject, that looked to me very closely related to the theory of computation, for which I had not intrinsic interest. Years later I actually signed up for the first ever MOOC that was based on that textbook, but I only did the first couple of weeks of work, and never finished it.
That MOOC, however, gave me the taste of the possibilities of online education. So even though I had vowed to myself never ever to take another class after I finished my PhD (yes, I was pretty burned out with formal education) I found the online MOOC format fun and rewarding, so I started taking more ML and DS classes. And then I stumbled upon Kaggle. And those online classes and Kaggle lead to a new and exciting career for me.
Thanks to Kaggle, I have been exposed to a wide spectrum of various ML problems with very important and valuable impact on various applications and problems. I’ve seen how good ML modeling can dramatically improve the accuracy of the credit risk modeling, detect proteins in microscopy images, deal with abusive online behavior, to name just a few of my successful competitions. Furthermore, Kaggle competitions are only a an incredibly small fraction of all the interesting applications of ML in the world, and so far we have only actually looked into applying ML to a very small subset of all the problems that we could potentially tackle with it. The number of problems that I myself wish I had the time and resources to wok on is overwhelming, and a big part of my mental energy is dedicated to saying no (mostly in my head) to all the things I want to do.
Most of our attention to GAI and fascination with it are deeply rooted in our own anthropomorphic biases. This is not surprising. In nature human intelligence is such a dominant force, and we are so impressed by it, that it’s almost literally inconceivable to think of anything more perfect and powerful than it. Even our conceptions of superior intelligence - in religion, literature, arts - take human intelligence as the starting point and slightly extrapolate from it. However, it is quite plausible that the path to superior intelligence may not in the end lead through obtaining GAI first. If we think of intelligence in a broader way as an ability to obtain, process, and utilize information, then we could focus improving each one of those substeps more concretely and in ways that don’t depend on cracking the understanding of human intelligence first. This would require us to pay more attention to how we value and utilize information in the world, how our professional roles are based on it, and how the economics and politics of information operates. And then find ways for machines to get even better at it.