Tag: ai


Dreyfus and Bostrom. Four AI assumptions and two books.

April 23rd, 2017 — 9:09pm

At first glance, Hubert Dreyfus’ 1992 book What Computers Still Can’t Do (WCSCD, originally published in 1972 as What Computers Can’t Do) seems untimely in the current business climate, which favours massive and widespread investment in AI (these days, often understood as being synonymous with machine learning and neural networks). However, being untimely may in fact allows us to act “against our time and thus hopefully also on our time, for the benefit of a time to come” (Nietzsche). And the book’s argument might in fact not be outdated, but simply forgotten in the frenzy of activity that is our present AI summer.

Dreyfus outlines four assumptions that he believes were (in many cases, still are) implicitly made by AI optimists.

The biological assumption. On some level, the (human) brain functions like a digital computer, processing discrete information.

The psychological assumption. The mind, rather than the brain, functions like a digital computer, even if the brain doesn’t happen to do so.

The epistemological assumption. Even if neither minds nor brains function like digital computers, then this formalism is still sufficient to explain and generate intelligent behaviour. An analogy would be that planets moving in orbits are perhaps not solving differential equations, but differential equations are adequate tools for describing and understanding their movement.

The ontological assumption. Everything essential to intelligent behaviour ­— such as information about the environment — can in principle be formalised as a set of discrete facts. 

These assumptions all relate to the limitations of computation (as we currently understand it) and of propositional logic.

Dreyfus is famous for interpreting thinkers such as Heidegger and Merleau-Ponty, and consistently draws upon these thinkers in his arguments. In fact, as he points out in WCSCD, the phenomenological school attacks the very long philosophical tradition that sees mind and world as strictly separate, and that assumes that the mind functions by way of a model that somehow can be reduced to logical operations (we can see why the field of AI has implicitly, and in many cases unwittingly, taken over this tradition). Historically, this tradition reached perhaps one of its purest expressions with Descartes. Indeed Being and Time, Heidegger’s major work, is very anti-Cartesian. Heidegger’s account of intelligibility demands that one (Dasein) is in a world which appears primarily as meaningful interrelated beings (and not primarily as atomic facts, or sources thereof, to be interpreted), and is historically in a situation, making projections on the basis of one’s identity. Here, calculation and correspondence-based theories of truth are derived and secondary things. There is no clear separation between world and “model” since there is no model, just the world and our ability to relate to it.

I will hazard a guess that most neuroscientists today would not take the first two assumptions seriously. In all kinds of biology and medicine, we regularly encounter new phenomena and mechanisms that could not be captured by the simple models we originally came up with, forcing us to revise our models. Making brains (bodies) and/or minds somehow isomorphic to symbolic manipulation seems wholly inadequate. More interesting, and much harder to settle unambiguously, are the epistemological and the ontological assumptions. If the epistemological assumption is false, then we will not be able to generate “intelligent behaviour” entirely in software. If the ontological assumption is false, then we will not be able to construct meaningful (discrete and isolated) models of the world.

The two latter assumptions are indeed the stronger ones out of these four. If the epistemological assumption turns out to be invalid, then the biological and psychological assumptions would necessarily also be invalid. The ontological assumption is closely related and similarly strong.

By contrast, Nick Bostrom‘s Superintelligence: Paths, Dangers, Strategies is a more recent (2014) and very different book. While they are certainly worth serious investigation, theories about a possible technological singularity can be somewhat hyperbolic in tone. But Bostrom comes across as very level-headed as he investigates how a superintelligence might be formed (as an AI, or otherwise), how it might or might not be controlled, and the political implications of such an entity coming into existence. For the most part, the book is engrossing and interesting, though clearly grounded in the “analytical” tradition of philosophy. It becomes more compelling because of the potential generality of its argument. Does a superintelligence already exist? Would we know if it did? Could it exist as a cybernetic actor, a composite of software, machines, and people? It is interesting to read the book, in parallel, as a speculation on (social, economic, geopolitical, technological, psychological or composites thereof) actors that may already exist but that are beyond our comprehension.

Bostrom’s arguments resemble how one might think about a nuclear arms race. He argues that the first superintelligence to emerge might have a decisive strategic advantage and, once in place, prevent (or be used to prevent) the emergence of competing superintelligences. At the same time it would bestow upon those who control it (if it can be controlled) a huge tactical advantage.

Even though Bostrom’s argument is mostly very general, at times it is obvious that much of the thinking is inspired by or based on the idea of AI as software running on a digital computer. To me this seemed implicit in many of the chapters. For example, Bostrom talks about being able to inspect the state of a (software agent’s) goal model, to be able to suspend, resume, and copy agents without information loss, to measure hedonic value, and so on. Bostrom in many cases implies that we would be able to read, configure and copy an agent’s state precisely, and sometimes also that we would be able to understand this state clearly and unambiguously, for example in order to evaluate whether our control mechanisms are working. Thus many of Bostrom’s arguments seem tightly coupled to the Church-Turing model of computation (or at least to a calculus/operational substrate that allows for inspection, modification and duplication of state). Some of his other arguments are, however, sufficiently general that we do not need to assume any specific substrate.

Bostrom, it seems to me, implicitly endorses at least the epistemological assumption throughout the book (and possibly also the ontological one). Even as he rightly takes pains to avoid stating specifically how technologies such as superintelligences or whole brain emulation would be implemented, it is clear that he imagines the formalism of digital computers as “sufficient to explain and generate intelligent behaviour”. In this, but perhaps not in everything he writes, he is a representative of current mainstream AI thinking. (I would like to add that even if he has wrongly taken over these assumptions, the extreme caution he advises us to proceed with regarding strong AI deserves to be taken seriously – the risks in practice are sufficiently great for us to be quite worried. I do not wish to undermine his main argument.)

It is thinkable but unlikely that in the near future, through a resounding success (which could be an academic, industrial or commercial one, for example), the epistemological assumption will be proven true. What I hold to be more likely (for reasons that have been gradually developed on this blog) is that current AI work will converge on something that may well be extremely impressive and that may affect society greatly, but that we will not consider to be human-like intelligence. The exact form that this will take remains to be discovered.

Hubert Dreyfus passed away in April 2017, while I was in the middle of writing this post. Although I never had the privilege of attending his lectures in person, his podcasted lectures and writings have been extremely inspirational and valuable to me. Thank you.

Comment » | Computer science, Philosophy

AI and the politics of perception

August 1st, 2016 — 11:36am

Elon Musk, entrepreneur of some renown, believes that the sudden eruption of a very powerful artificial intelligence is one of the greatest threats facing mankind. “Control of a super powerful AI by a small number of humans is the most proximate concern”, he tweets. He’s not alone among silicon valley personalities to have this concern. To reduce the risks, he has funded the OpenAI initiative, which aims to develop AI technologies in such a way that they can be distributed more evenly in society. Musk is very capable, but is he right in this case?

The idea is closely related to the notion of a technological singularity, as promoted by for example Kurzweil. In some forms, the idea of a singularity resembles a God complex. In C G Jung’s view, as soon the idea of God is expelled (for example by saying that God is dead), God appears as a projection somewhere. This because the archetype or idea of God is a basic feature of the (western, at least) psyche that is not so easily dispensed with. Jung directs this criticism at Nietzsche in his Zarathustra seminar. (Musk’s fear is somewhat more realistic and, yes, proximate, than Kurzweil’s idea, since what is feared is a constellation of humans and technology, something we already have.)

But if Kurzweil’s singularity is a God complex, then the idea of the imminent dominance of uncontrollable AI, about to creep up on us out of some dark corner, more closely resembles a demon myth.

Such a demon myth may not be useful in itself for understanding and solving social problems, but its existence may point to a real problem. Perhaps what it points to is the gradual embedding of algorithms deeply into our culture, down to our basic forms of perception and interaction. We have in effect already merged with machines. Google and Facebook are becoming standard tools for information finding, socialising, getting answers to questions, communicating, navigating. The super-AI is already here, and it has taken the form of human cognition filtered and modulated by algorithms.

It seems fair to be somewhat suspicious — as many are — of fiat currency, on the grounds that a small number of people control the money supply, and thus, control the value of everybody’s savings. On similar grounds, we do need to debate the hidden algorithms, controlled by a small number of people (generally not available for perusal, even on request, since they would be trade secrets), and pre-digested information that we now use to interface with the world around us almost daily. Has it ever been so easy to change so many people’s perception at once?

Here again, as often is the case, nothing is truly new. Maybe we are simply seeing a tendency that started with the printing press and the monotheistic church, taken to its ultimate conclusion. In any case I would paraphrase Musk’s worry as follows: control of collective perception by a small number of humans is the most proximate concern. How we should address this concern is not immediately obvious.

Comment » | Computer science, Philosophy

Historical noise? Simulation and essential/accidental history

June 24th, 2015 — 4:58pm

Scientists and engineers around the world are, with varying degrees of success, racing to replicate biology and intelligence in computers. Computational biology is already simulating the nervous systems of entire organisms. Artificial intelligence seems to be able to replicate more tasks formerly thought to be the sole preserve of man each year. Many of the results are stunning. All of this is done on digital circuits and/or Turing-Church computers (two terms that for my purposes here are interchangeable — we could also call it symbol manipulation). Expectations are clearly quite high.

What should we realistically hope for? How far can these advances actually go? If they do not culminate in “actual” artificial biology (AB) and artificial intelligence (AI), then what will they end in – what logical conclusion will they reach, what kind of wall would they run up against? What expectations do we have of “actual” AB and AI?

These are extremely challenging questions. When thinking about them, we ought to always keep in mind that minds and biology are both, as far as science knows, open-ended systems, open worlds. This in the sense that we do not know all existing facts about them (unlike classical mechanics or integer arithmetic, which we can reduce to sets of rules). For all intents, given good enough equipment, we could make an indefinite amount of observations and data recordings from any cell or mind. Conversely, we cannot, starting from scratch, construct a cell or a mind starting from pure chemical compounds. Even given godlike powers in a perfectly controlled space, we wouldn’t know what to do. We cannot record in full detail the state of a (single!) cell or a mind, we cannot make perfect copies, and we cannot configure the state of a cell or mind with full precision. This is in stark contrast to digital computation, where we can always make an indefinite number of perfect copies, and where we know the lower bound of all relevant state – we know the smallest detail that matters. We know that there’s no perceivable high-level difference between having a potential difference of 5.03 volts or 5.04 volts in our transistors on the lowest level.

(Quantum theory holds that ultimately, energy can only exist in discrete states. It seems that one consequence would be that a given volume of matter can only represent a finite amount of information. For practical purposes this does not affect our argument here, since measurement and manipulation instruments in science are very far from being accurate and effective at a quantum level. It may certainly affect our argument in theory, but who says that we will not some day discover a deeper level that can hold more information?)

In other words, we know the necessary and sufficient substrate (theoretical and hardware basis) for digital computation, but we know of no such substrate for minds or cells. Furthermore, there are reasons to think that any such substrate would lie much deeper, and at a much smaller scale, than we tend to believe. We repeatedly discover new and unexpected functions of proteins and DNA. Junk DNA, a name that has more than a hint of hubris to it, was later found to have certain crucial functions – not exactly junk, in other words.

Attempts at creating artificial minds and/or artificial biology are attempts at creating detached versions of the original phenomena. They would exist inside containers, independently of time and entropy, as long as the sufficient electrical charge or storage integrity is maintained. Their ability to affect the rest of the universe, and to be affected by it, would be very strictly limited (though not nonexistent – for example, memory errors may occur in a computer as a result of electromagnetic interference from the outside). We may call such simulations unrooted or perhaps hovering. This is the quality that allows digital circuits to preserve information reliably. Interference and noise is screened out, removed.

In attempting to answer the questions posed above, we should think about two alternative scenarios, then.

Scenario 1. It is possible to find a sufficient substrate for biology and/or minds. Beneath a certain level, no further microscopic detail is necessary in the model to replicate the full range of phenomena. Biology and minds are then reduced to a kind of software; a finite amount of information, an arrangement of matter. No doubt such a case would be comforting to many of the logical positivists at large today. But it would also have many strange consequences.

Each of us as a living organism, society around us, and every entity has a history that stretches back indefinitely far. The history of cells needs a long pre-history and evolution of large molecules to begin. A substrate, in the above sense, exists and can be practically used if and only if large parts of history are dispensable. If we could create a perfect artificial cell on some substrate (in software, say) in a relatively short time span, say an hour, or, why not, less than a year, then it means that nature took an unnecessarily long way to get to its goal. (Luckily, efficient, rational, enlightened humans have now come along and found a way to cut out all that waste!) Our shorter way to the goal would then be something that cuts out all the accidental features of history, leaving only the essential parts in place. So the practically usable substrate, which allows for shortcuts in time, then seems to imply a division between essential and accidental history of the thing we wish to simulate! (I say “practically” usable, since an impractical alternative is a working substrate that requires as much time as natural history in the “real” world. In this scenario, getting to the first cell on the substrate takes as long as it did in reality starting from, say, the beginning of the universe. Not a practical scenario, but an interesting thought experiment.) Note that if we are able to somehow run time faster in the simulation than in reality, then it would also mean that parts of history (outside the simulation) are dispensable: some time would have been wasted on unecessary processes.

Scenario 2. Such a substrate does not exist. If no history is accidental, if the roundabout historical process taken by the universe to reach the goal of, say, the first cell or first mind, is actually the only way that such things can be attained, then this scenario would be implied. This scenario is just as astounding as the first, since it implies that each of us depends fully on all of the history and circumstances that led up to this moment.

In deciding which of the two scenarios is more plausible, we should note that both biology and minds seem to be mechanisms for recording history in tremendous detail. Recording ability gives them advantages. This, I think, speaks in favour of the second scenario. The “junk DNA” problem becomes transposed to history itself (of matter, of nature, of societies, of the universe). Is there such a thing as junk history, events that are mere noise?

In writing the above, my aim has not been to discourage any existing work or research. But the two possibilities above must be considered and could point the way to the most worthwhile research goals for AI and AB. If the substrates can be found, then all is “well”, and we would need to truly grapple with the fact that we ourselves are mere patterns/arrangements of building blocks, mere software, body and mind. If the substrates can not be found, as I am inclined to think, then perhaps we should begin to think about completely new kinds of computation, which could somehow incorporate the parts that are missing from mere symbol manipulation. We should also consider much more seriously how closed-world systems, such as the world of digital information, can coexist harmoniously with what would be open-world systems, such as biology and minds. It seems that these problems are scarcely given any thought today.

4 comments » | Bioinformatics, Computer science, Philosophy

The struggle over consciousness

October 15th, 2014 — 1:54pm

One of the major themes of Western philosophy since Plato is the elevation and near-deification of consciousness. Conscious thought and reflection have been prized above all else. Suspicion has been directed towards everything that is dark, murky, instinctive, unclear, unreasonable. Spirit has been emphasised above body. Christianity and its penal mechanisms was in no small part the engine used for this process for many centuries.

But what can consciousness really do? Every sequence of words I produce, every line of code I write, every sketch I draw or tune I play on the piano is for the most part not a product of conscious reflection. (Some earlier, unfinished thoughts on the limitations of reason here.) These productions are given to me, just as associations, feelings or moods are given to me – by the Other in me, the unconscious, the body. Through reflection I can remix and arrange these parts, critique them, say yes and no, but I cannot generate these things through purely conscious thought and logic. So what, in fact, was Western society really doing for 2000 years?

Nietzsche heralded the beginning of a reversal of this trend. In him, consciousness turns around, questions itself and finds that in the end, it isn’t all that powerful. A new philosophical school begins: a counter-movement that aimed, and aims to, reaffirm what is unthought, unseen, unreasonable. After him, thinkers like Freud, Jung (with his elaboration of the unconscious and his idea of “individuation”, psychological development understood as a harmonious union with the unconscious), Foucault (whose “History of madness”, if not almost his entire oeuvre, is almost entirely about this theme and the technicalities of how the unreasonable was suppressed) and Heidegger (in part) progressed on this path. But this reversal has only just begun. What are, in the grand scheme of things, the 130 years since Nietzsche’s productive years in the 1880s? The battle over the value of consciousness is in full swing and might be for centuries or millennia yet. And so we find ourselves, for now, living in a schizophrenic society, perhaps on the threshold of crossing over from a value system that praises consciousness to one that gives it a much more modest role.

 

Comment » | Philosophy

How one might develop a Heideggerian AI that uses software equipment

August 5th, 2012 — 9:00pm

This year I’ve spent a fair amount of time trying to read Martin Heidegger‘s great work Being and Time, using Hubert Dreyfus’ Berkeley lectures on the book as supporting material. By now I’ve almost finished division 1. I’m learning a lot, but it’s fair to say that this is one of the most difficult books I’ve read. I’m happy to have come this far and think I have some kind of grasp of what’s going on.

I’ve also come to understand that Heidegger played an important role in the so-called “AI debate” in the 70’s and 80’s. At the time, people at MIT, DARPA and other institutions were trying to make AI software based on the presumptions of an Aristotelian world view, representing facts as logical propositions. (John McCarthy, of Lisp fame, and Marvin Minsky were some of the key people working on these projects). Dreyfus made himself known as a proponent of uncomfortable views (for the AI establishment) at the time, such as Heidegger’s claim that you cannot represent human significance and meaning using predicate logic (more on that in a different post, when I understand it better).

There were even attempts at making a “Heideggerian AI” in response to Dreyfus’ criticism, when it became apparent that “good old fashioned AI”, GOFAI, had failed. But apparently the Heideggerian AI also failed – according to Dreyfus, this was because it wasn’t Heideggerian enough.

Using part 1 of Being and Time as inspiration, I have come up with a possibly novel idea for a “Heideggerian” AI. This is also a first attempt at expressing some of the (incomplete, early) understanding I think I have excavated from Being and Time. As my point of departure, I use the notion of equipment. Heidegger’s Dasein essentially needs equipment in order to take a stance on its being. It has a series of “for-the-sake-of-whichs” which lead up to an “ultimate for-the-sake-of-which”. In the case of an equipment-wielding AI, we might start by hardcoding its ultimate FTSOW as the desire to serve its human masters well by carrying out useful tasks. Dasein can ponder and modify its ultimate FTSOW over time, but at least initially, our AI might not need this capability.

Heidegger’s Dasein is essentially mitdasein, being-with-others. Furthermore, it has an essential tendency to do what “the they”/”the one” does, imitating the averageness of the Dasein around it. This is one of its basic ways of gaining familiarity with practices. By observing human operators using equipment, a well-programmed AI might be able to extract a working knowledge of how to use the same equipment. But what equipment should the AI use to train and evolve itself in its nascent, most basic stages? If the equipment exists in the physical world, the AI will need a sophisticated way of identifying this equipment and detecting how it is used, for example by applying feature detection and image processing to a video feed. This process is error prone and would complicate the task of creating the essential core of a rudimentary but well-functioning AI. Instead, I propose that the AI should use software tools as equipment alongside human operators who use the same tools.

Let’s now consider some of the characteristics of the being of equipment (the ready-to-hand) that Heidegger mentions. When Dasein is using equipment in skilled activity, the equipment nearly disappears. Dasein becomes absorbed in the activity. But if there is a problem with the activity or with the equipment, the equipment becomes more obtrusive. Temporarily broken equipment stands out and draws our attention. Permanently broken equipment is uncanny and disturbing. Various levels of obtrusiveness correspond to levels of breakdown of the skilled activity. And not only obtrusiveness: we become more aware of the finer details of the equipment as it breaks down, in a positive way, so that we may fix it. All this is certainly true for a hammer, car or sewing machine, but is it true of software tools? We may consider both how human users relate to software today, and how our hypothetical AI would relate to it.

Unfortunately it can be said that a lot of software today is designed — including but not limited to user interfaces — in such a way that when it breaks down, the essential details that need to be perceived in order to fix the problems are not there to be seen, for the vast majority of users with an average level of experience. When the software equipment breaks down, presumably our human cognition goes into alert and readies itself to perceive more details so that it can form hypotheses of how to remedy the errors that have arisen. But those details are not on offer. The software has been designed to hide them. In this sense, the vast majority of software that people use today does not fail smoothly. It fails to engage the natural problem solving capacity of humans when it does break, because the wrong things are exposed, and in the wrong ways and contexts. Software equipment has a disadvantage compared with physical equipment: we cannot inspect it freely with all of our senses, and the scrutiny of its inner details may involve some very artificial and sophisticated operations. The makers may even actively seek to block this scrutiny (code obfuscation, etc). In the case of software equipment, such scrutiny is greatly separated from the everyday use by numerous barriers. In the case of physical equipment, there is often a smooth continuum.

We now have the opportunity to tackle two challenges at once. First, we should develop an AI that can use software equipment alongside humans – that is, use it in the same way or nearly the same way that they use it, and for the same or nearly the same purposes. Second, we should simultaneously develop software that “breaks down well”, in the sense that its inner structure becomes visible to users when it fails, in such a way that they can restore normal functioning in a natural way. These users can be both humans and the AIs that we are trying to develop. Since the AI should mimic human cognition, a design that is good for one of these categories should be good for the other as well. In this way we can potentially develop a groundbreaking AI in tandem with a groundbreaking new category of software. Initially, both the AI and the software equipment should be very simple, and the complexity of both would increase gradually.

There would be one crucial difference between the way that humans use the software equiment and that the AI would use it. Human beings interact with software through graphical (GUI) or command-line (CLI) interfaces. This involves vision, reading and linguistic comprehension. These are also higher order capabilities that may get in the way of developing a basic AI with core functionality as smoothly as possible. In order to avoid depending on these capabilities, we can give the AI a direct window into the software equipment. This would effectively be an artificial sense that tells the AI what the equipment is doing at various levels of detail, depending on how smooth the current functioning is. This would be useful both for the AI’s own use of equipment and for its observation of how humans use the same equipment. In this way we can circumnavigate the need for capacities such as vision, language, locomotion and physical actuators, and focus only on the core problem of skilled activity alongside humans. Of course this kind of system might later serve as a foundation on which these more advanced capacities can be built.

Many questions have been left unanswered here. For example, the AI must be able to judge the outcome of its work. But the problems that it solves inside the computer will reference the “external” world at all times (external in so far as the computer is separated from the world, which is to say, not really external). I am not aware of many problems I solve on computers that do not carry, directly or indirectly, references to the world outside the computer. Such references to the external world mean that the “common sense” problem must be addressed: arbitrary details that have significance for a problem may appear, pop up, or emerge from the background, and Dasein would know the relation of these details to the problem at hand, since this is intelligible on the basis of the world which it already understands. It remains to be seen if our limited AI can gain a sufficient understanding of the world by using software equipment alongside Dasein. However, I believe that the simultaneous development of software equipment and a limited AI that is trained to use it holds potential as an experimental platform on which to investigate AIs and philosophy, as well as software development principles.

 

2 comments » | Computer science, Philosophy, Software development

Back to top