A review of “Architects of Intelligence” by Martin Ford (Packt, 2018)
Towards the end of last year I was sent a review copy of a new book on AI. I read it over Christmas, and since it was pretty good I thought I’d do a little review of it here, as I’ve been pretty busy managing a growth spurt at Hospify lately, and haven’t published on my Medium blog for a while.
The book is called “Architects of Intelligence: The Truth about AI from the people Building It”, and as the title suggests it’s a collection of interviews, conducted by Martin Ford, whose recent book “Rise of the Robots” (which I haven’t read) won various plaudits including Financial Times Business book of the year.
When I received the book, I have to confess wasn’t overwhelmed with enthusiasm. I read a fair number of AI newsletters and listen to quite a few AI podcasts (TWIML AI is my favourite — shout out to Sam Charrington), and glancing through the list of interviewees it felt a bit like a collection of musings by the usual suspects — Yoshua Bengio, Geoffrey Hinton, Yann LeCun Demis Hassabis, Ray Kurzweil, Gary Marcus, Andrew Ng, Jeff Dean et al. While this is undoubtedly a list of extremely smart people, all of whom have excelled in their fields and have an enormous amounts to say, I did feel like I’ve already read or listened to plenty of interviews with all of them, and wondered what another set of “in conversation” transcripts could add. After all, AI is a technical field, and there’s a point at which general discussion about it loses its charm, and what you want from practitioners is a deeper dive into the nuts and bolts of the techniques involved (one of the things TWIML AI strives for, and often achieves), rather than yet another overview of the landscape.
However, as I worked my way through the 23 interviews in the book, it became apparent that Ford was doing something quite interesting. The big news in AI at the moment is of course the advances made in deep learning by many of the luminaries listed above — Bengio, Hinton, Hassabis, Ng, LeCun in particular. Ford groups these guys in the first half of the collection, interspersing them with philosophical covering fire from the likes of Nick Bostrom and Ray Kurzweil, the latter making the core point in his interview that “connectionism can emulate a rule-based approach, [but] a rule-based system really cannot emulate a connectionist system, so the converse statement is not the case.”
This is all great as far as it goes. Don’t get me wrong. I love deep learning. I first came across Hinton’s work (with Rumelhart and McClelland) in connectionism (i.e. deep learning and neural nets) as a post-grad back in 1992, and it blew my mind. I think that the work that’s been in recent years to implement it has been amazing, and it doesn’t need me to tell you that it’s been so transformative for computing in that it has re-energised the whole industry. I’ve even coded some deep learning projects of my own.
At the same time, however, as a philosophy and neurophysiology student, and I was (and remain) sceptical of claims that deep learning neural nets, which really only approximate how we think the brain works in the vaguest sense, are all there is to intelligence. After all, intelligence is not just about learning. It’s also about wanting. The philosopher I loved best as a post-grad was Gilles Deleuze, who described the abstract processes that underpin life (and thus intelligence) as “desiring machines”. Deep learning captures the machinic aspect of this tuple as well as any process we’ve been able to instantiate on a computer. What it doesn’t capture is the “desiring”. So far, that bit — the bit that the system wants, the thing that justifies its existence, causes it to act, and assesses its outcomes — is still provided — at every juncture — by the human beings that build the system. Until the machine takes over at least some aspect of that, “artificial intelligence” will remain a misnomer. What we’ll have, as we have now, is “augmented intelligence” or, “machine learning”, the term I prefer.
This niggle is the point at which “Architects of Intelligence” starts to gain purchase and subsequently gets really interesting. Ford has grouped the sceptics (by which I mean those who believe, as I do, that this is more than a niggle, but the whole crux of the problem), so just when you think the argument is done the voices of dissent begin to surface. The psychologist & technologist Gary Marcus is kind of their cheerleader, but there are many other folk here with strong hands on experience, particularly in robotics — Rana El Kaliouby, Rodney Brooks, Cynthia Breazeal, Joshua Tenenbaum — who point out, repeatedly and convincingly, the hard truth that deep learning, while being fantastic for pulling patterns out of vast data mines and finding signal in noise way beyond human abilities, is just not well-suited to generalising from smaller sample sizes. In the world of textual, voice or image analysis this isn’t a problem, at least not now we have the vast resources of the global internet to throw at it. But once you get out of server farms and into the physical world it’s a showstopper, as generalisation from small sample sizes is the way in which genuinely (as opposed to artificially) intelligent agents are able to extrapolate a desire-driven route through a world about which, in fact, they know very little.
The conceptual covering fire in this second half of the book is provided by Judea Pearl, who like Hinton was inspired by the work of David Rumelhart, and like Kurzweil has genuinely technical chops, but who provides some serious insights into the nature of probability and causal modelling (and thus desire, to continue with my earlier terminology), demonstrating a philosophical depth well beyond the somewhat excitable futurist extrapolations of his counterparts in the first half the book.
Ford summarises the point perfectly: “causation can never be learned from data alone.” In statistics, this is almost a cliché. “Causation is not correlation” is a phrase statisticians regularly repeat to each other (and their students) to remind themselves of the importance of not reading too much into raw data. But the cliché, as clichés so often are, is a flag pinned on the site of a profound truth. Not only is not causation not correlation, but they are entirely different beasts, one of which — any quantum physicist will tell you — we can only barely claim to understand. To quote Pearl:
Neural nets take sample sizes of thousands or millions, and discover patterns in the data. Humans take a sample size of one and generalise patterns from that seed. This is the difference between learning and intelligence. We’ve built machines that exhibit the former. We’re a very long way from building them to exhibit the latter.
If you stick with it through all of its 500 or so pages, therefore (and I suggest that you do, and that you read the interviews sequentially), I think you’ll find as I did that far from being an ad hoc round up of the self-aggrandising thoughts of the current crop of machine learning geniuses, “Architects of Intelligence” is in fact a very dense and complete overview of the field and a great map of both its achievements and its very real challenges.
While it doesn’t have much technical detail, what it does have is genuine philosophical and conceptual nuance, and I came away from reading it assured of what I’ve been feeling for some time — that artificial general intelligence is much further off that many would have us believe, but that far from being a disappointment this in fact means is that the current excitement around deep learning is only the first in many such similar flowerings that are going to keep humanity busily engaged for the next few centuries at least, as we edge our way towards an increasingly profound understanding of what it means to build a mind.