Poorly Structured Notes on AI Part 6 - LeCun vs Hassabis

One of the effects of social media is that, while it has always been the case that you can throw a paper aeroplane into a crowd and have a fair chance of it hitting an idiot, now the impression is of a throng composed almost entirely of them, and if you throw a paper aeroplane at them, they will probably fight over it, and try to eat it.

But can we get better guidance from our thought leaders? Most spokespersons connected to companies are trying to drive expectations of future profits. Their words and actions don’t represent bets on what is likely to happen. They are attempts to manipulate what investors think will happen. It’s the investors who make the actual bets, and there is often little reason to regard their choices as being based on anything better than poorly informed guesswork. So most commentary tends to be a mixture of factual contributions and “influencing.” The messages that reach us are not necessarily from honest communicators, and we have to be highly selective and interpretive to get anything of value from them.

In the last week there has been a flare-up between Demis Hassabis, founder of what is now Google DeepMind, and Yann LeCun, previously of Meta AI, about the precise meaning of the word “general”, as in Artificial General Intelligence. Apart from exemplifying the kind of side-show semantic debate that dominates in this area, there is something important about it. LeCun said (extemporaneously, in an interview):

“There is no such thing as general intelligence. This concept makes absolutely no sense. Because it’s really designed to designate human level intelligence. But human intelligence is super specialised. We can handle the real world really well, and navigate, and blah blah blah, and we can handle other humans really well because we evolved to do this, and chess, we suck, okay? There’s a lot of tasks that we suck at, where a lot of other animals are much better than we are. So what that means is that we are specialised. We think of ourselves as being general, but it’s simply an illusion, because all the problems that we can apprehend are the ones we can think of. And vice versa. And so we’re general in all the problems that we can imagine. But there’s a lot of problems that we cannot imagine… so this concept of general intelligence is complete BS.”

Hassabis responded in writing:

“Yann is just plain incorrect here, he’s confusing general intelligence with universal intelligence. Brains are the most exquis​ite and complex phenomena we know of in the universe (so far), and they are in fact extremely general. Obviously one can’t circumvent the no free lunch theorem so in a practical and finite system there always has to be some degree of specialisation around the ​target distribution that is being learnt. But the point about generality is that in theory, in the Turing Machine sense​, the architecture of ​s​uch a general system is capable of learning anything computable given enough time and memory​ (and data), and the human brain (and AI foundation models) are approximate Turing Machines. Finally, with ​regards to ​Yann’s comments about chess players, it’s amazing that humans could have invented chess ​in the first place (and all the other ​a​spects ​o​f modern civilization ​from science to 747s!) let alone get as brilliant at it as someone like Magnus. He may not be ​strictly optimal (after all he has finite memory and limited time to make a decision) but it’s incredible what he and we can do with our brains given they were evolved for hunter gathering.”

Now, LeCun’s point is somewhat like Dunning-Kruger: a person thinks they know all there is to know about a subject because they are large ignorant of it, so they don’t know how little they know. Such confidence shrinks with greater exposure, so experts feel less certain of anything in their subject of expertise, even though, paradoxically, there is far more of which they can be certain. Likewise, humans imagine we have something called “general intelligence” because we are terribly impressed with the achievements of whatever kind of intelligence we have, but we struggle to imagine anything beyond it. We can list our impressive achievements: we can write long run-on sentences, we can tap-dance, the wheel, New York, wars, and so on.

LeCun asserts that all these achievements are in fact demonstrations of a patchwork of highly-skewed inherited capabilities, biased toward solving certain classes of problem, and we just dress them up in various forms until we’re convinced that we’ve covered a smooth and wide continuum of some general “problem space”, and that is what we label as general intelligence. Indeed, that metaphor I just used of a continuous space, like a physical space - isn’t that just like a human, to use their probably innate ability to imagine physical spaces and to stretch that one big idea to encompass other domains that, almost certainly, aren’t really appropriate to think of as continuous spaces? Not to get too Kantian, a continuous physical space is so useful, it’s baked into how our brains work, and we relate so many other things to it. (Even Chess is ultimately about moving around in a physical space, and a fight between enemies, two things humans obsess over.)

This idea, far from being controversial or original to LeCun, is standard systems neuroscience. For example, our visual cortex at the rear of the brain has specific circuitry for performing edge and corner detection, processing colour information, sensing motion, all the way up to recognising faces, filling in missing information, etc. without “you” having any concious awareness of this effort. There are other areas taking care of the details of sound processing, controlling your limbs, dealing with language, and doing each of these things in sophisticated ways that require different, specialised forms of intelligence. Or to put it another way:

• The human brain is intrinsically organized into dynamic, anticorrelated functional networks

We have to keep reusing the instinctive mental machinery that evolution has gifted us, and it is specialised. We are not quite as bad as the person with a hammer to whom every problem is a nail; we at least have a toolkit containing a hammer, a wrench, a saw. But we also possess a system of selective attention that causes us to mostly ignore problems that aren’t a nail, a loose bolt or something that needs cutting in half. As a result, we feel as if we have the important stuff mostly covered, what with our general toolkit. There’s nothing we can’t do! Nothing important, anyway.

On the other hand, this works for us because the world has certain reliable, exploitable patterns in it. As long as they persist, we have the right tools. As we rebuild the world around us, we risk it becoming impossible for us to navigate with our specialised toolkit. But we have the solution to that: our artificial worlds all look suspiciously familiar..

The significance to LeCun making this point, and the response from Hassabis, is that “general” is part of the industry buzz phrase, AGI, on which so many vague promises of future ROI rest (AGI, in essence, means little more than “Yay, we can replace all our employees now!”) In Hassabis’s reply, there’s a panicked, scatter-gun approach. He’s deflecting rather than engaging. First he invokes straw man: LeCun, he says, is arguing against something called “universal” intelligence, which isn’t the same as the “general” intelligence that investors are being asked to fork over cash for. But then he switches right back to defending the generality/universality of brains directly against LeCun’s main thrust, by invoking Turing completeness, and that is, I think, wildly mistaken, as well as undermining the straw man defence. He also unnecessarily reminds us that “Chess brainy! We pretty smart for monkeys!” which only underscores LeCun’s point.

He is actually referring to Universal Turing Machines (perhaps he preferred to avoid the word “universal”, having just expended it as part of the straw man dismissal?) A UTM can be fed a stream of information that tells it how to behave subsequently, and thus embody any other machine. It is conjectured (uncontroversially) that it represents a universal capability to carry out any information-processing algorithm. All digital computers fall into this category, and yet no one is claiming that (say) the Apple II had general intelligence. By that, we must mean something other than (or in addition to) a UTM. It is blatant obfuscation to pretend that UTMs imply general intelligence. They are almost irrelevant. Invoking them in this way is just as misguided as invoking Godel’s incompleteness theorems in an attempt to prove that computers will never reach human intelligence.

That said, there is a better response that could be made to LeCun: if you have the right tools for some persistent scenario, then you are “general enough” for that scenario. To say “Oh, that’s not really general, it just seems like it is because you’re not general enough to perceive the limitation,” is akin to complaining that ever since you invented a means of intergalactic travel, you’ve been feeling kind of claustrophobic now you’ve realised you’re trapped in this one universe.

There is a synthesis of the two perspectives, then: you can be general enough to meet the challenges of a largely stable environment, and you can achieve this through a patchwork of specialisations. One such specialisation is the machinery of language processing, from which there is reason to believe we derive some of our capability for logical analysis. The invention of algebra, for example, is possibly a kind of reuse of the machinery that allows us to parse complex, deep grammatical structures.