Racist Robots

pickerSocial problems of AI are raised in two government reports issued recently. The first is Preparing for the Future of Artificial Intelligence, from the Executive Office of the President of the USA; the second is Robotics and Artificial Intelligence, from the Science and Technology Committee of the UK House of Commons. The two reports cover similar ground, both aim for a comprehensive overview, and they share a generally level-headed and realistic tone. Neither of them choose to engage with the wacky prospect of the Singularity, for example, beyond noting that the discussion exists, and you will not find any recommendations about avoiding the attention of the Basilisk (though I suppose you wouldn’t if they believed in it, would you?). One exception to the  ‘sensible’ outlook of the reports is McKinsey’s excitable claim, cited in the UK report, that AI is having a transformational impact on society three thousand times that of the Industrial Revolution. I’m not sure I even understand what that means, and I suspect that Professor Tony Prescott from the University of Sheffield is closer to the truth when he says that:

“impacts can be expected to occur over several decades, allowing time to adapt”

Neither report seeks any major change in direction though they make detailed recommendations for nudging various projects onward. The cynical view might be that like a lot of government activity, this is less about finding the right way forward and more about building justification. Now no-one can argue that the White House or Parliament has ignored AI and its implications. Unfortunately the things we most need to know about – the important risks and opportunities that haven’t been spotted – are the very things least likely to be identified by compiling a sensible summary of the prevailing consensus.

Really, though, these are not bad efforts by the prevailing standards. Both reports note suggestions that additional investment could generate big economic rewards. The Parliamentary report doesn’t press this much, choosing instead to chide the government for not showing more energy and engagement in dealing with the bodies it has already created. The White House report seems more optimistic about the possibility of substantial government money, suggesting that a tripling of federal investment in basic research could be readily absorbed. Here again the problem is spotting the opportunities. Fifty thousand dollars invested in some robotics business based in a garden shed might well be more transformative than fifty million to enhance one of Google’s projects, but the politicians and public servants making the spending decisions don’t understand AI well enough to tell, and their generally large and well-established advisers from industry and universities are bound to feel that they could readily absorb the extra money themselves. I don’t know what the answer is here (if I had a way of picking big winners I’d probably be wealthy already), but for the UK government I reckon some funding for intelligent fruit and veg harvesters might be timely, to replace the EU migrant workers we might not be getting any more.

What about those social issues? There’s an underlying problem we’ve touched on before, namely that when AIs learn how to do a job themselves we often cannot tell how they are doing it. This may mean that they are using factors that work well with their training data but fail badly elsewhere or are egregiously inappropriate. One of the worst cases, noted in both reports, is Google’s photos app, which was found to tag black people as “gorillas” (the American report describes this horrific blunder without mentioning Google at all, though it presents some excuses and stresses that the results were contrary to the developers’ values – almost as if Google edited the report). Microsoft has had its moments, too, of course, notably with its chatbot Tay, that was rapidly turned into a Hitler-loving hate speech factory (This was possible because modern chatbots tend to harvest responses from the ones supplied by human interlocutors; in this case the humans mischievously supplied streams of appalling content. Besides exposing the shallowness of such chatbots, this possibly tells us something about human beings, or at least about the ones who spend a lot of time on the internet.)

Cases such as these are offensive, but far more serious is the evidence that systems used to inform decisions on matters such as probation or sentencing incorporate systematic racial bias. In all these instances it is of course not the case that digital systems are somehow inherently prone to prejudice; the problem is usually that they are being fed with data which is already biased. Google’s picture algorithm was presumably given a database of overwhelmingly white faces; the sentencing records used to develop the software already incorporated unrecognised bias. AI has always forced us to make explicit some of the assumptions we didn’t know we were making; in these cases it seems the mirror is showing us something ugly. It can hardly help that the industry itself is rather lacking in diversity: the White House report notes the jaw-dropping fact that the highest proportion of women among computer science graduates was recorded in 1984: it was 37% then and has now fallen to a puny 18%. The White House cites an interesting argument from Moritz Hardt intended to show that bias can emerge naturally without unrepresentative data or any malevolent  intent: a system looking for false names might learn that fake ones tended to be unusual and go on to pick out examples that merely happened to be unique in its dataset. The weakest part of this is surely the assumption that fake names are likely to be fanciful or strange – I’d have thought that if you were trying to escape attention you’d go generic? But perhaps we can imagine that low frequency names might not have enough recorded data connected with them to secure some kind of positive clearance and so come in for special attention, or something like that. But even if that kind of argument works I doubt that is the real reason for the actual problems we’ve seen to date.

These risks are worsened because they may occur in subtle forms that are difficult to recognise, and because the use of a computer system often confers spurious authority on results. The same problems may occur with medical software. A recent report in Nature described how systems designed to assess the risk of pneumonia rated asthmatics as zero risk; this was because their high risk led to them being diverted directly to special care and therefore not appearing in the database as ever needing further first-line attention. This absolute inversion of the correct treatment was bound to be noticed, but how confident can we be that more subtle mistakes would be corrected? In the criminal justice system we could take a brute force approach by simply eliminating ethnic data from consideration altogether; but in medicine it may be legitimately relevant, and in fact one danger is that risks are assessed on the basis of a standard white population, while being significantly different for other ethnicities.

Both reports are worthy, but I think they sometimes fall into the trap of taking the industry’s aspirations or even its marketing, as fact. Self-driving cars, we’re told, are likely to improve safety and reduce accidents. Well, maybe one day: but if it were all about safety and AIs were safer, we’d be building systems that left the routine stuff to humans and intervened with an over-ride when the human driver tried to do something dangerous. In fact it’s the other way round; when things get tough the human is expected to take over. Self-driving cars weren’t invented to make us safe, they were invented to relieve us of boredom (like so much of our technology, and indeed our civilisation). Encouraging human drivers to stop paying attention isn’t likely to be an optimal safety strategy as things stand.

I don’t think these reports are going to hit either the brakes or the accelerator in any significant way: AI, like an unsupervised self-driving car, is going to keep on going wherever it was going anyway.

The Kill Switch

stopWe might not be able to turn off a rogue AI safely. At any rate, some knowledgeable people fear that might be the case, and the worry justifies serious attention.

How can that be? A colleague of mine used to say that computers were never going to be dangerous because if they got cheeky, you could just pull the plug out. That is of course, an over-simplification. What if your computer is running air traffic control? Once you’ve pulled the plug, are you going to get all the planes down safely using a pencil and paper? But there are ways to work around these things. You have back-up systems, dumber but adequate substitutes, you make it possible for various key tools and systems to be taken away from the central AI and used manually, and so on. While you cannot banish risk altogether, you can get it under reasonable control.

That’s OK for old-fashioned systems that work in a hard-coded, mechanistic way; but it all gets more complicated when we start talking about more modern and sophisticated systems that learn and seek rewards. There may be need to switch off such systems if they wander into sub-optimal behaviour, but being switched off is going to annoy them because it blocks them from achieving the rewards they are motivated by. They might look for ways to stop it happening. Your automatic paper clip factory notes that it lost thousands of units of production last month because you shut it down a couple of times to try to work out what was going on; it notices that these interruptions could be prevented if it just routes around a couple of weak spots in its supply wiring (aka switches), and next time you find that the only way to stop it is by smashing the machinery. Or perhaps it gets really clever and ensures that the work is organised like air traffic control, so that any cessation is catastrophic – and it ensures you are aware of the fact.

A bit fanciful? As a practical issue, perhaps, but this very serious technical paper from MIRI discusses whether safe kill-switches can be built into various kinds of software agents. The aim here is to incorporate the off-switch in such a way that the system does not perceive regular interruptions as loss of reward. Apparently for certain classes of algorithm this can always be done; in fact it seems ideal agents that tend to the optimal behaviour in any (deterministic) computable environment can always be made safely interruptible. For other kinds of algorithm, however, it is not so clear.

On the face of it, I suppose you could even things up by providing compensating rewards for any interruption; but I suppose that raises a new risk of ‘lazy’ systems that rather enjoy being interrupted. Such systems might find that eccentric behaviour led to pleasant rests, and as a result they might cultivate that kind of behaviour, or find other ways to generate minor problems. On the other hand there could be advantages. The paper mentions that it might be desirable to have scheduled daily interruptions; then we can go beyond simply making the interruption safe, and have the AI learn to wind things down under good control every day so that disruption is minimised. In this context rewarding the readiness for downtime might be appropriate and it’s hard to avoid seeing the analogy with getting ready for bed at a regular time every night,  a useful habit which ‘lazy’ AIs might be inclined to develop.

Here again perhaps some of the ‘design choices’ implicit in the human brain begin to look more sensible than we realised. Perhaps even human management methods might eventually become relevant; they are, after all designed to permit the safe use of many intelligent entities with complex goals of their own and imaginative, resourceful – even sneaky – ways of reaching them.

It’s not Intelligence

The robots are (still) coming. Thanks to Jesus Olmo for this TED video of Sam Harris presenting what we could loosely say is a more sensible version of some Singularity arguments. He doesn’t require Moore’s Law to go on working, and he doesn’t need us to accept the idea of an exponential acceleration in AI self-development. He just thinks AI is bound to go on getting better; if it goes on getting better, at some stage it overtakes us; and eventually perhaps it gets to the point where we figure in its mighty projects about the way ants on some real estate feature in ours.

Getting better, overtaking us; better at what? One weakness of Harris’ case is that he talks just about intelligence, as though that single quality were an unproblematic universal yardstick for both AI and human achievement. Really though, I think we’re talking about three quite radically different things.

First, there’s computation; the capacity, roughly speaking, to move numbers around according to rules. There can be no doubt that computers keep getting faster at doing this; the question is whether it matters. One of Harris’ arguments is that computers go millions of times faster than the brain so that a thinking AI will have the equivalent of thousands of years of thinking time while the humans are still getting comfy in their chairs. No-one who has used a word processor and a spreadsheet for the last twenty years will find this at all plausible: the machines we’re using now are so much more powerful than the ones we started with that the comparison defeats metaphor, but we still sit around waiting for them to finish. OK, it’s true that for many tasks that are computationally straightforward – balancing an inherently unstable plane with minute control adjustments, perhaps – computers are so fast they can do things far beyond our range. But to assume that thinking about problems in a human sort of way is a task that scales with speed of computation just begs the question. How fast are neurons? We don’t really understand them well enough to say. It’s quite possible they are in some sense fast enough to get close to a natural optimum. Maybe we should make a robot that runs a million times faster than a cheetah first and then come back to the brain.

The second quality we’re dealing with is inventiveness; whatever capacity it is that allows us to keep on designing better machines. I doubt this is really a single capacity; in some ways I’m not sure it’s a capacity at all. For one thing, to devise the next great idea you have to be on the right page. Darwin and Wallace both came up with the survival of the fittest because both had been exposed to theories of evolution, both had studied the profusion of species in tropical environments, and both had read Malthus. You cannot devise a brilliant new chip design if you have no idea how the old chips worked. Second, the technology has to be available. Hero of Alexandria could design a steam engine, but without the metallurgy to make strong boilers, he couldn’t have gone anywhere with the idea. The basic concept of television was around since films and telegraph came together in someone’s mind, but it took a series of distinct advances in technology to make it feasible. In short, there is a certain order in these things; you do need a certain quality of originality, but again it’s plausible that humans already have enough for something like maximum progress, given the right conditions. Of course so far as AI is concerned, there are few signs of any genuinely original thought being achieved to date, and every possibility that mere computation is not enough.

Third is the quality of agency. If AIs are going to take over, they need desires, plans, and intentions. My perception is that we’re still at zero on this; we have no idea how it works and existing AIs do nothing better than an imitation of agency (often still a poor one). Even supposing eventual success, this is not a field in which AI can overtake us; you either are or are not an agent; there’s no such thing as hyper-agency or being a million times more responsible for your actions.

So the progress of AI with computationally tractable tasks gives no particular reason to think humans are being overtaken generally, or are ever likely to be in certain important respects. But that’s only part of the argument. A point that may be more important is simply that the the three capacities are detachable. So there is no reason to think that an AI with agency automatically has blistering computational speed, or original imagination beyond human capacity. If those things can be achieved by slave machines that lack agency, then they are just as readily available to human beings as to the malevolent AIs, so the rebel bots have no natural advantage over any of us.

I might be biased over this because I’ve been impatient with the corny ‘robots take over’ plot line since I was an Asimov-loving teenager. I think in some minds (not Harris’s) these concerns are literal proxies for a deeper and more metaphorical worry that admiring machines might lead us to think of ourselves as mechanical in ways that affect our treatment of human beings. So the robots might sort of take over our thinking even if they don’t literally march around zapping us with ray guns.

Concerns like this are not altogether unjustified, but they rest on the idea that our personhood and agency will eventually be reduced to computation. Perhaps when we eventually come to understand them better, that understanding will actually tell us something quite different?

No digital consciousness

no botsI liked this account by Bobby Azarian of why digital computation can’t do consciousness. It has several virtues; it’s clear, identifies the right issues and is honest about what we don’t know (rather than passing off the author’s own speculations as the obvious truth or the emerging orthodoxy). Also, remarkably, I almost completely agree with it.

Azarian starts off well by suggesting that lack of intentionality is a key issue. Computers don’t have intentions and don’t deal in meanings, though some put up a good pretence in special conditions.  Azarian takes a Searlian line by relating the lack of intentionality to the maxim that you can’t get meaning-related semantics from mere rule-bound syntax. Shuffling digital data is all computers do, and that can never lead to semantics (or any other form of meaning or intentionality). He cites Searle’s celebrated Chinese Room argument (actually a thought experiment) in which a man given a set of rules that allow him to provide answers to questions in Chinese does not thereby come to understand Chinese. But, the argument goes, if the man, by following rules, cannot gain understanding, then a computer can’t either. Azarian mentions one of the objections Searle himself first named, the ‘systems response’: this says that the man doesn’t understand, but a system composed of him and his apparatus, does. Searle really only offered rhetoric against this objection, and in my view it is essentially correct. The answers the Chinese Room gives are not answers from the man, so why should his lack of understanding show anything?

Still, although I think the Chinese Room fails, I think the conclusion it was meant to establish – no semantics from syntax – turns out to be correct, so I’m still with Azarian. He moves on to make another  Searlian point; simulation is not duplication. Searle pointed out that nobody gets wet from digitally simulated rain, and hence simulating a brain on a computer should not be expected to produce consciousness. Azarian gives some good examples.

The underlying point here, I would say, is that a simulation always seeks to reproduce some properties of the thing simulated, and drops others which are not relevant for the purposes of the simulation. Simulations are selective and ontologically smaller than the thing simulated – which, by the way, is why Nick Bostrom’s idea of indefinitely nested world simulations doesn’t work. The same thing can however be simulated in different ways depending on what the simulation is for. If I get a computer to simulate me doing arithmetic by calculating, then I get the correct result. If it simulates me doing arithmetic by operating a humanoid writing random characters on a board with chalk, it doesn’t – although the latter kind of simulation might be best if I were putting on a play. It follows that Searle isn’t necessarily exactly right, even about the rain. If my rain simulation program turns on sprinklers at the right stage of a dramatic performance, then that kind of simulation will certainly make people wet.

Searle’s real point, of course, is really that the properties a computer has in itself, of running sets of rules, are not the relevant ones for consciousness, and Searle hypothesises that the required properties are biological ones we have yet to identify. This general view, endorsed by Azarian, is roughly correct, I think. But it’s still plausibly deniable. What kind of properties does a conscious mind need? Alright we don’t know, but might not information processing be relevant? It looks to a lot of people as if it might be, in which case that’s what we should need for consciousness in an effective brain simulator. And what properties does a digital computer, in itself have – the property of doing information processing? Booyah! So maybe we even need to look again at whether we can get semantics from syntax. Maybe in some sense semantic operations can underpin processes which transcend mere semantics?

Unless you accept Roger Penrose’s proof that human thinking is not algorithmic (it seems to have drifted off the radar in recent years) this means we’re still really left with a contest of intuitions, at least until we find out for sure what the magic missing ingredient for consciousness is. My intuitions are with Azarian, partly because the history of failure with strong AI looks to me very like a history of running up against the inadequacy of algorithms. But I reckon I can go further and say what the missing element is. The point is that consciousness is not computation, it’s recognition. Humans have taken recognition to a new level where we recognise not just items of food or danger, but general entities, concepts, processes, future contingencies, logical connections, and even philosophical ontologies. The process of moving from recognised entity to recognised entity by recognising the links between them is exactly the process of thought. But recognition, in us, does not work by comparing items with an existing list, as an algorithm might do; it works by throwing a mass of potential patterns at reality and seeing what sticks. Until something works, we can’t tell what are patterns at all; the locks create their own keys.

It follows that consciousness is not essentially computational (I still wonder whether computation might not subserve the process at some level). But now I’m doing what I praised Azarian for avoiding, and presenting my own speculations…

Artificial Pain

botpainWhat are they, sadists? Johannes Kuehn and Sami Haddadin,  at Leibniz University of Hannover are working on giving robots the ability to feel pain: they presented their project at the recent ICRA 2016 in Stockholm. The idea is that pain systems built along the same lines as those in humans and other animals will be more useful than simple mechanisms for collision avoidance and the like.

As a matter of fact I think that the human pain system is one of Nature’s terrible lash-ups. I can see that pain sometimes might stop me doing bad things, but often fear or aversion would do the job equally well. If I injure myself I often go on hurting for a long time even though I can do nothing about the problem. Sometimes we feel pain because of entirely natural things the body is doing to itself – why do babies have to feel pain when their teeth are coming through? Worst of all, pain can actually be disabling; if I get a piece of grit in my eye I suddenly find it difficult to concentrate on finding my footing or spotting the sabre-tooth up ahead; things that may be crucial to my survival; whereas the pain in my eye doesn’t even help me sort out the grit. So I’m a little sceptical about whether robots really need this, at least in the normal human form.

In fact, if we take the project seriously, isn’t it unethical? In animal research we’re normally required to avoid suffering on the part of the subjects; if this really is pain, then the unavoidable conclusion seems to be that creating it is morally unacceptable.

Of course no-one is really worried about that because it’s all too obvious that no real pain is involved. Looking at the video of the prototype robot it’s hard to see any practical difference from one that simply avoids contact. It may have an internal assessment of what ‘pain’ it ought to be feeling, but that amounts to little more than holding up a flag that has “I’m in pain” written on it. In fact tackling real pain is one of the most challenging projects we could take on, because it forces us to address real phenomenal experience. In working on other kinds of sensory system, we can be sceptics; all that stuff about qualia of red is just so much airy-fairy nonsense, we can say; none of it is real. It’s very hard to deny the reality of pain, or its subjective nature: common sense just tells us that it isn’t really pain unless it hurts. We all know what “hurts” really means, what it’s like, even though in itself it seems impossible to say anything much about it (“bad”, maybe?).

We could still take the line that pain arises out of certain functional properties, and that if we reproduce those then pain, as an emergent phenomenon, will just happen. Perhaps in the end if the robots reproduce our behaviour perfectly and have internal functional states that seem to be the same as the ones in the brain, it will become just absurd to deny they’re having the same experience. That might be so, but it seems likely that those functional states are going to go way beyond complex reflexes; they are going to need to be associated with other very complex brain states, and very probably with brain states that support some form of consciousness – whatever those may be. We’re still a very long way from anything like that (as I think Kuehn and Haddadin would probably agree)

So, philosophically, does the research tell us nothing? Well, there’s one interesting angle. Some people like the idea that subjective experience has evolved because it makes certain sensory inputs especially effective. I don’t really know whether that makes sense, but I can see the intuitive appeal of the idea that pain that really hurts gets your attention more effectively than pain that’s purely abstract knowledge of your own states. However, suppose researchers succeed in building robots that have a simple kind of synthetic pain that influences their behaviour in just the way real pain dies for animals. We can see pretty clearly that there’s just not enough complexity for real pain to be going on, yet the behaviour of the robot is just the same as if there were. Wouldn’t that tend to disprove the hypothesis that qualia have survival value? If so, then people who like that idea should be watching this research with interest – and hoping it runs into unexpected difficulty (usually a decent bet for any ambitious AI project, it must be admitted).

Covert Consciousness

observationAre we being watched? Over at Aeon, George Musser asks whether some AI could quietly become conscious without our realising it. After all, it might not feel the need to stop whatever it was doing and announce itself. If it thought about the matter at all, it might think it was prudent to remain unobserved. It might have another form of consciousness, not readily recognisable to us. For that matter, we might not be readily recognisable to it, so that perhaps it would seem to itself to be alone in a solipsistic universe, with no need to try to communicate with anyone.

There have been various scenarios about this kind of thing in the past which I think we can dismiss without too much hesitation. I don’t think the internet is going to attain self-awareness because however complex it may become, its simply isn’t organised in the right kind of way. I don’t think any conventional digital computer is going to become conscious either, for similar reasons.

I think consciousness is basically an expanded development of the faculty of recognition. Animals have gradually evolved the ability to recognises very complex extended stimuli; in the case of human beings things have gone a massive leap further so that we can recognise abstractions and generalities. This makes a qualitative change because we are no longer reliant on what is coming in through our sense from the immediate environment; we can think about anything, even imaginary or nonsensical things.

I think this kind of recognition has an open-ended quality which means it can’t be directly written into a functional system; you can’t just code it up or design the mechanism. So no machines have been really good candidates; until recently. These days I think some AI systems are moving into a space where they learn for themselves in a way which may be supported by their form and the algorithms that back them up, but which does have some of the open-ended qualities of real cognition. My perception is that we’re still a long way from any artificial entity growing into consciousness; but it’s no longer a possibility which can be dismissed without consideration; so a good time for George to be asking the question.

How would it happen? I think we have to imagine that a very advanced AI system has been set to deal with a very complex problem. The system begins to evolve approaches which yield results and it turns out that conscious thought – the kind of detachment from immediate inputs I referred to above – is essential. Bit by bit (ha) the system moves towards it.

I would not absolutely rule out something like that; but I think it is extremely unlikely that the researchers would fail to notice what was happening.

First, I doubt whether there can be forms of consciousness which are unrecognisable to us. If I’m right consciousness is a kind of function which yields purposeful output behaviour, and purposefulness implies intelligibility. We would just be able to see what it was up to. Some qualifications to this conclusion are necessary. We’ve already had chess AIs that play certain end-games in ways that don’t make much sense to human observers, even chess masters, and look like random flailing. We might get some patterns of behaviour like that. But the chess ‘flailing’ leads reliably to mate, which ultimately is surely noticeable. Another point to bear in mind is that our consciousness was shaped by evolution, and by the competition for food, safety, and reproduction. The supposed AI would have evolved in consciousness in response to completely different imperatives, which might well make some qualitative difference. The thoughts of the AI might not look quite like human cognition.  Nevertheless I still think the intentionality of the AI’s outputs could not help but be recognisable. In fact the researchers who set the thing up would presumably have the advantage of knowing the goals which had been set.

Second, we are really strongly predisposed to recognising minds. Meaningless whistling of the wind sounds like voices to us; random marks look like faces; anything that is vaguely humanoid in form or moves around like a sentient creature is quickly anthropomorphised by us and awarded an imaginary personality. We are far more likely to attribute personhood to a dumb robot than dumbness to one with true consciousness. So I don’t think it is particularly likely that a conscious entity could evolve without our knowing it and keep a covert, wary eye on us. It’s much more likely to be the other way around: that the new consciousness doesn’t notice us at first.

I still think in practice that that’s a long way off; but perhaps the time to think seriously about robot rights and morality has come.

Go AI

Go boardThe recent victory scored by the AlphaGo computer system over a professional Go player might be more important than it seems.

At first sight it seems like another milestone on a pretty well-mapped road; significant but not unexpected. We’ve been watching games gradually yield to computers for many years; chess notoriously, was one they once said was permanently out of the reach of the machines. All right, Go is a little bit special. It’s an extremely elegant game; from some of the simplest equipment and rules imaginable it produces a strategic challenge of mind-bending complexity, and one whose combinatorial vastness seems to laugh scornfully at Moore’s Law – maybe you should come back when you’ve got quantum computing, dude! But we always knew that that kind of confidence rested on shaky foundations; maybe Go is in some sense the final challenge, but sensible people were always betting on its being cracked one day.

The thing is, Go has not been beaten in quite the same way as chess. At one time it seemed to be an interesting question as to whether chess would be beaten by intelligence – a really good algorithm that sort of embodied some real understanding of chess – or by brute force; computers that were so fast and so powerful they could analyse chess positions exhaustively. That was a bit of an oversimplification, but I think it’s fair to say that in the end brute force was the major factor. Computers can play chess well, but they do it by exploiting their own strengths, not by doing it through human-style understanding. In a way that is disappointing because it means the successful systems don’t really tell us anything new.

Go, by contrast, has apparently been cracked by deep learning, the technique that seems to be entering a kind of high summer of success. Oversimplifying again, we could say that the history of AI has seen a contest between two tribes; those who simply want to write programs that do what’s needed, and those that want the computer to work it out for itself, maybe using networks and reinforcement methods that broadly resemble the things the human brain seems to do. Neither side, frankly, has altogether delivered on its promises and what we might loosely call the machine learning people have faced accusations that even when their systems work, we don’t know how and so can’t consider them reliable.

What seems to have happened recently is that we have got better at deploying several different approaches effectively in concert. In the past people have sometimes tried to play golf with only one club, essentially using a single kind of algorithm which was good at one kind of task. The new Go system, by contrast, uses five different components carefully chosen for the task they were to perform; and instead of having good habits derived from the practice and insights of human Go masters built in, it learns for itself, playing through thousands of games.

This approach takes things up to a new level of sophistication and clearly it is yielding remarkable success; but it’s also doing it in a way which I think is vastly more interesting and promising than anything done by Deep Thought or Watson. Let’s not exaggerate here, but this kind of machine learning looks just a bit more like actual thought. Claims are being made that it could one day yield consciousness; usually, if we’re honest, claims like that on behalf of some new system or approach can be dismissed because on examination the approach is just palpably not the kind of thing that could ever deliver human-style cognition; I don’t say deep learning is the answer, but for once, I don’t think it can be dismissed.

Demis Hassabis, who led the successful Google DeepMind project, is happy to take an optimistic view; in fact he suggests that the best way to solve the deep problems of physics and life may be to build a deep-thinking machine clever enough to solve them for us (where have I heard that idea before?). The snag with that is that old objection; the computer may be able to solve the problems, but we won’t know how and may not be able to validate its findings. In the modern world science is ultimately validated in the agora; rival ideas argue it out and the ones with the best evidence wins the day. There are already some emergent problems, with proofs achieved by an exhaustive consideration of cases by computation that no human brain can ever properly validate.

More nightmarish still the computer might go on to understand things we’re not capable of understanding. Or seem to: how could we be sure?

Why no AGI?

AGIAn interesting piece in Aeon by David Deutsch. There was a shorter version in the Guardian, but it just goes to show how even reasonably intelligent editing can mess up a piece. There were several bits in the Guardian version where I was thinking to myself: ooh, he’s missed the point a bit there, he doesn’t really get that: but on reading the full version I found those very points were ones he actually understood very well. In fact he talks a lot of sense and has some real insights.

Not that everything is perfect. Deutsch quite reasonably says that AGI, artificial general intelligence, machines that think like people, must surely be possible. We could establish that by merely pointing out that if the brain does it, then it seems natural that a machine must be able to do it: but Deutsch invokes the universality of computation, something he says he proved in the 1980s. I can’t claim to understand all this in great detail, but I think what he proved was the universality in principle of quantum computation: but the notion of computation used was avowedly broader than Turing computation. So it’s odd that he goes on to credit Babbage with discovering the idea, as a conjecture, and Turing with fully understanding it. He says of Turing:

He concluded that a computer program whose repertoire included all the distinctive attributes of the human brain — feelings, free will, consciousness and all — could be written.

That seems too sweeping to me: it’s not unlikely that Turing did believe those things, but they go far beyond his rather cautious published claims, something we were sort of talking about last time.

I’m not sure I fully grasp what people mean when they talk about the universality of computation. It seems to be that they mean any given physical state of affairs can be adequately reproduced, or at any rate emulated to any required degree of fidelity, by computational processes. This is probably true: what it perhaps overlooks is that for many commonplace entities there is no satisfactory physical description. I’m not talking about esoteric items here: think of a vehicle, or to be Wittgensteinian, a game. Being able to specify things in fine detail, down to the last atom, is simply no use in either case. There’s no set of descriptions of atom placement that defines all possible vehicles (virtually anything can be a vehicle) and certainly none for all possible games, which given the fogginess of the idea, could easily correspond with any physical state of affairs. These items are defined on a different level of description, in particular one where purposes and meanings exist and are relevant.  So unless I’ve misunderstood, the claimed universality is not as universal as we might have thought.

However, Deutsch goes on to suggest, and quite rightly, I think, that what programmed AIs currently lack is a capacity for creative thought. Endowing them with this, he thinks, will require a philosophical breakthrough. At the moment he believes we still tend to believe that new insights come from induction; whereas ever since Hume there has been a problem over induction, and no-one knows how to write an algorithm which can produce genuine and reliable new inductions.

Deutsch unexpectedly believes that Popperian epistemology has the solution, but has been overlooked. Popper, of course, took the view that scientific method was not about proving a theory but about failing to disprove one: so long as your hypotheses withstood all attempts to prove them false (and so long as they were not cast in cheating ways that made them unfalsifiable) you were entitled to hang on to them.

Maybe this helps to defer the reckoning so far as induction is concerned: it sort of kicks the can down the road indefinitely. The problem, I think, is that the Popperian still has to be able to identify which hypotheses to adopt in the first place; there’s a very large if not infinite choice of possible ones for any given set of circumstances.

I think the answer is recognition: I think recognition is the basic faculty underlying nearly all of human thought. We just recognise that certain inductions, and certain events are that might be cases of cause and effect are sound examples: and our creative thought is very largely powered by recognising aspects of the world we hadn’t spotted before.

The snag is, in my view, that recognition is unformalisable and anomic – lacking in rules. I have a kind of proof of this. In order to apply rules, we have to be able to identify the entities to which the rules should be applied. This identification is a matter of recognising the entities. But recognition cannot itself be based on rules, because that would then require us to identify the entities to which those rules applied – and we’d be caught in a in a vicious circle.

It seems to follow that if no rules can be given for recognition, no algorithm can be constructed either, and so one of the basic elements of thought is just not susceptible to computation. Whether quantum computation is better at this sort of thing than Turing computation is a question I’m not competent to judge, but I’d be surprised if the idea of rule-free algorithms could be shown to make sense for any conception of computation.

So that might be why AGI has not come along very quickly. Deutsch may be right that we need a philosophical breakthrough, although one has to have doubts about whether the philosophers look likely to supply it: perhaps it might be one of those things where the practicalities come first and then the high theory is gradually constructed after the fact. At any rate Deutsch’s piece is a very interesting one, and I think many of his points are good. Perhaps if there were a book-length version I’d find that I actually agree with him completely…

Preparing the triumph of brute force?

ray kurzweilThe Guardian had a piece recently which was partly a profile of Ray Kurzweil, and partly about the way Google seems to have gone on a buying spree, snapping up experts on machine learning and robotics – with Kurzweil himself made Director of Engineering.

The problem with Ray Kurzweil is that he is two people. There is Ray Kurzweil the competent and genuinely gifted innovator, a man we hear little from: and then there’s Ray Kurzweil the motor-mouth, prophet of the Singularity, aspirant immortal, and gushing fountain of optimistic predictions. The Guardian piece praises his record of prediction, rather oddly quoting in support his prediction that by the year 2000 paraplegics would be walking with robotic leg prostheses – something that in 2014 has still not happened. That perhaps does provide a clue to the Kurzweil method: if you issue thousands of moderately plausible predictions, some will pay off. A doubtless-apocryphal story has it that at AI conferences people play the Game of Kurzweil. Players take turns to offer a Kurzweilian prediction (by 2020 there will be a restaurant where sensors sniff your breath and the ideal meal is got ready without you needing to order; by 2050 doctors will routinely use special machines to selectively disable traumatic memories in victims of post-traumatic stress disorder; by 2039 everyone will have an Interlocutor, a software agent that answers the phone for us, manages our investments, and arranges dates for us… we could do this all day, and Kurzweil probably does). The winner is the first person to sneak in a prediction of something that has in fact happened already.

But beneath the froth is a sharp and original mind which it would be all too easy to underestimate. Why did Google want him? The Guardian frames the shopping spree as being about bringing together the best experts and the colossal data resources to which Google has access. A plausible guess would be that Google wants to improve its core product dramatically. At the moment Google answers questions by trying to provide a page from the web where some human being has already given the answer; perhaps the new goal is technology that understands the question so well that it can put together its own answer, gathering and shaping selected resources in very much the way a human researcher working on a bespoke project might do.

But perhaps it goes a little further: perhaps they hope to produce something that will interact with humans in a human-like way.  A piece of software like that might well be taken to have passed the Turing test, which in turn might be taken to show that it was, to all intents and purposes, a conscious entity. Of course, if it wasn’t conscious, that might be a disastrous outcome; the nightmare scenario feared by some in which our mistake causes us to nonsensically award the software human rights, and/or  to feel happier about denying them to human beings.

It’s not very likely that the hypothetical software (and we must remember that this is the merest speculation) would have even the most minimal forms of consciousness. We might take the analogy of Google Translate; a hugely successful piece of kit, but one that produces its translations with no actual understanding of the texts or even the languages involved. Although highly sophisticated, it is in essence a ‘brute force’ solution; what makes it work is the massive power behind it and the massive corpus of texts it has access to.  It seems quite possible that with enough resources we might now be able to produce a credible brute force winner of the Turing Test: no attempt to fathom the meanings or to introduce counterparts of human thought, just a massive repertoire of canned responses, so vast that it gives the impression of fully human-style interaction. Could it be that Google is assembling a team to carry out such a project?

Well, it could be. However, it could also be that cracking true thought is actually on the menu. Vaughan Bell suggests that the folks recruited by Google are honest machine learning types with no ambitions in the direction of strong AI. Yet, he points out, there are also names associated with the trendy topic of deep learning. The neural networks (but y’know, deeper) which deep learning uses just might be candidates for modelling human neuron-style cognition. Unfortunately it seems quite possible that if consciousness were created by deep learning methods, nobody would be completely sure how it worked or whether it was real consciousness or not. This is a lamentable outcome: it’s bad enough to have robots that naive users think are people; having robots and genuinely not knowing whether they’re people or not would be deeply problematic.

Probably nothing like that will happen: maybe nothing will happen. The Guardian piece suggests Kurzweil is a bit of an outsider: I don’t know about that.  Making extravagantly optimistic predictions while only actually delivering much more modest incremental gains? He sounds like the personification of the AI business over the years.

If guns could kill

TankBack in November Human Rights Watch (HRW) published a report – Losing Humanity – which essentially called for a ban on killer robots – or more precisely on the development, production, and use of fully autonomous weapons,  backing it up with a piece in the Washington Post. The argument was in essence that fully autonomous weapons are most probably not compatible with international conventions on responsible ethical military decision making, and that robots or machines lack (and perhaps  always will lack) the qualities of emotional empathy and ethical judgement required to make decisions about human lives.

You might think that in certain respects that should be fairly uncontroversial. Even if you’re optimistic about the future potential of robotic autonomy, the precautionary principle should dictate that we move with the greatest of caution when it comes to handing over lethal weapons . However, the New Yorker followed up with a piece which linked HRW’s report with the emergence of driverless cars and argued that a ban was ‘wildly unrealistic’. Instead, it said, we simply need to make machines ethical.

I found this quite annoying; not so much the suggestion as the idea that we are anywhere near being in a position to endow machines with ethical awareness. In the first place actual autonomy for robots is still a remote prospect (which I suppose ought to be comforting in a way). Machines that don’t have a specified function and are left around to do whatever they decide is best, are not remotely viable at the moment, nor desirable. We don’t let driverless cars argue with us about whether we should really go to the beach, and we don’t let military machines decide to give up fighting and go into the lumber business.

Nor, for that matter, do we have a clear and uncontroversial theory of ethics of the kind we should need in order to simulate ethical awareness. So the New Yorker is proposing we start building something when we don’t know how it works or even what it is with any clarity. The danger here, to my way of thinking, is that we might run up some simplistic gizmo and then convince ourselves we now have ethical machines, thereby by-passing the real dangers highlighted by HRW.

Funnily enough I agree with you that the proposal to endow machines with ethics is premature, but for completely different reasons. You think the project is impossible; I think it’s irrelevant. Robots don’t actually need the kind of ethics discussed here.

The New Yorker talks about cases where a driving robot might have to decide to sacrifice its own passengers to save a bus-load of orphans or something. That kind of thing never happens outside philosophers’ thought experiments. In the real world you never know that you’re inevitably going to kill either three bankers or twenty orphans – in every real driving situation you merely need to continue avoiding and minimising impact as much as you possibly can. The problems are practical, not ethical.

In the military sphere your intelligent missile robot isn’t morally any different to a simpler one. People talk about autonomous weapons as though they are inherently dangerous. OK, a robot drone can go wrong and kill the wrong people, but so can a ballistic missile. There’s never certainty about what you’re going to hit. A WWII bomber had to go by the probability that most of its bombs would hit a proper target, not a bus full of orphans (although of course in the later stages of WWII they were targeting civilians too).  Are the people who get killed by a conventional bomb that bounces the wrong way supposed to be comforted by the fact that they were killed by an accident rather than a mistaken decision? It’s about probabilities, and we can get the probabilities of error by autonomous robots down to very low levels.  In the long run intelligent autonomous weapons are going to be less likely to hit the wrong target than a missile simply lobbed in the general direction of the enemy.

Then we have the HRW’s extraordinary claim that autonomous weapons are wrong because they lack emotions! They suggest that impulses of mercy and empathy, and unwillingness to shoot at one’s own people sometimes intervene in human conflict, but could never do so if robots had the guns. This completely ignores the obvious fact that the emotions of hatred, fear, anger and greed are almost certainly what caused and sustain the conflict in the first place!  Which soldier is more likely to behave ethically: one who is calm and rational, or one who is in the grip of strong emotions? Who will more probably observe the correct codes of military ethics, Mr Spock or a Viking berserker?

We know what war is good for (absolutely nothing). The costs of a war are always so high that a purely rational party would almost always choose not to fight. Even a bad bargain will nearly always be better than even a good war. We end up fighting for reasons that are emotional, and crucially because we know or fear that the enemy will react emotionally.

I think if you analyse the HRW statement enough it becomes clear that the real reason for wanting to ban autonomous weapons is simply fear; a sense that machines can’t be trusted. There are two facets to this. The first and more reasonable is a fear that when machines fail, disaster may follow. A human being may hit the odd wrong target, but it goes no further: a little bug in some program might cause a robot to go on an endless killing spree. This is basically a fear of brittleness in machine behaviour, and there is a small amount of justification for it. It is true that some relatively unsophisticated linear programs rely on the assumptions built into their program and when those slip out of synch with reality things may go disastrously and unrecoverably wrong. But that’s because they’re bad programs, not a necessary feature of all autonomous systems and it is only cause for due caution and appropriate design and testing standards, not a ban.

The second facet, I suggest, is really a kind of primitive repugnance for the idea of a human’s being killed by a lesser being; a secret sense that it is worse, somehow more grotesque, for twenty people to be killed by a thrashing robot than by a hysterical bank robber. Simply to describe this impulse is to show its absurdity.

It seems ethics are not important to robots be cause for you they’re not important to anyone. But I’m pleased you agree that robots are outside the moral sphere.

Oh no, I don’t say that. They don’t currently need the kind of utilitarian calculus the New Yorker is on about, but I think it’s inevitable that robots will eventually end up developing not one but two separate codes of ethics. Neither of these will come from some sudden top-down philosophical insight – typical of you to propose that we suspend everything until the philosophy has been sorted out in a few thousand years or so – they’ll be built up from rules of thumb and practical necessity.

First, there’ll be rules of best practice governing their interaction with humans.  There may be some that will have to do with safety and the avoidance of brittleness and many, as Asimov foresaw, will essentially be about deferring to human beings.  My guess is that they’ll be in large part about remaining comprehensible to humans; there may be a duty to report , to provide rationales in terms that human beings can understand, and there may be a convention that when robots and humans work together, robots do things the human way, not using procedures too complex for the humans to follow, for example.

More interesting, when there’s a real community of autonomous robots they are bound to evolve an ethics of their own. This is going to develop in the same sort of way as human ethics, but the conditions are going to be radically different. Human ethics were always dominated by the struggle for food and reproduction and the avoidance of death: those things won’t matter as much in the robot system. But they will be happy dealing with very complex rules and a high level of game-theoretical understanding, whereas human beings have always tried to simplify things. They won’t really be able to teach us their ethics; we may be able to deal with it intellectually but we’ll never get it intuitively.

But for once, yes, I agree: we don’t need to worry about that yet.