1
Thinking Otherwise
Somewhere on the higher slopes of Mount Parnassus a small, dark grey car makes its way along a roughly tarmacked track. The road is fringed with snow; far below, the Gulf of Corinth glitters in the sun. The car moves slowly, almost carefully: it is watching the road. It has eyes – several of them – which track the edges of the embankments, identify the white markings at the junctions, note and transcribe where stops are made and turnings taken. It has other senses too: it can tell how fast it is travelling, where it is on the map, what angle the steering wheel is set to. And it has a kind of mind. Not a very sophisticated one, but with a clear focus and a capacity to learn from its surroundings, integrate its findings, and extrapolate and make predictions about the world around it. That mind was perched precariously on the passenger seat; I sat at the wheel, still in control, for now.
All this took place a few years ago, in the winter of 2017, when I decided to try and build myself a self-driving car. And although it never – quite – drove itself, it did take me to some pretty interesting places.
The idea of a self-driving car is fascinating to me. Not really for its capabilities, but for its place in our imagination. The self-driving car is one of those technologies which in the space of just a few years has gone from space-age, ‘Life in the Twenty-First Century’ fantasy to humdrum reality, without ever passing through a period of critical reflection or assimilation. In moments like this, reality is rewritten. The same will almost certainly be true of more advanced forms of AI. They will appear, suddenly, in our midst – the long slog of research and development, invisible to most, forgotten in the fact of their reality. Questions about who gets to do that rewriting of reality, which decisions are made along the way, and who gains from it, are all too often missed and forgotten in the excitement. That is why I believe that it’s crucially important for as many of us as possible to be engaged in thinking through the implications of new technologies; and that this process has to include learning about and tinkering with the things ourselves.
My attempt at building an autonomous vehicle consisted of a rented SEAT hatchback, a few cheap webcams, a smartphone taped to the steering wheel, and some software copied and pasted from the internet.1 This wasn’t a case of programming a dumb machine with everything it needed to know in advance, however. Like the commercial systems developed by Google, Tesla and others, my car would learn to drive by watching me drive: by comparing the view from the cameras with my speed, acceleration, steering wheel position and so forth, the system matched my behaviour with the road shape and condition, and after a couple of weeks it had learned how to keep a vehicle on the road – in a simulator at least. I’m not the world’s best driver, and I wouldn’t trust anyone’s life to this thing, but the experience of writing code and going out on the road gave me a better understanding of how certain kinds of AI operate, and what it feels like to work alongside a learning system.
I wondered too what it would mean to do this kind of work far from the highways of California, where Silicon Valley trains its self-driving cars, or the test-tracks of Bavaria, where the giants of the automotive industry evolve new models, and instead on the roads of Greece, where I had recently found myself living. This was a place with a very different material and mythological past and present. It turned out to go beautifully.
Leaving Athens and heading north with no particular destination in mind, other than to give my AI co-pilot a taste of many different kinds of terrain, I soon found myself passing the ancient sites of Thebes and Marathon, and climbing towards the dark bulk of Mount Parnassus. In Greek mythology, Parnassus was sacred to the cult of the god Dionysus, whose ecstatic mysteries were revealed by consuming copious amounts of wine and dancing wildly; participants in such rites liberated the beast within to become one with nature. Parnassus was also the home of the Muses, the goddesses who inspired literature, science and the arts. To attain the summit of Parnassus is thus to be elevated to the peak of knowledge, craft and skill.
Chance and geography conspired to frame a fascinating question. What would it mean, mythologically speaking, to be driven up Parnassus by an AI? On the one hand, it might be read as a kind of submission to the machine: an admission that the human race has run its course, and that it is time to pass the mantle of exploration and discovery to our robot overlords. On the other hand, to attempt the journey in the spirit of mutual understanding rather than conquest might just be how we write a new narrative onto Parnassus – one in which human and machine intelligences amplify, rather than try to outdo, one another.
I started this project because I wanted to understand AI better, and in particular because I wanted to have the experience of collaborating with an intelligent machine, rather than trying to determine its output. In fact, the whole effort was predicated on a kind of anti-determinacy: I wanted to plan as little as possible about the whole journey. So one thing I did, when training the car, was to drive completely at random, taking almost every side road and turning I came across, wandering and wondering, and getting totally, happily lost. In turn, by watching me, the car learned to get lost too.
This was a deliberate rejection of the kind of driving most of us do today: plugging a destination into a GPS system and following its directions without question or input. This loss of agency and control is mirrored in society at large. Confronted by ever more complex and opaque technologies, we capitulate to their commands, and a combination of fear and boredom is the frequent result. Instead of surrendering to a set of processes I didn’t understand, only to arrive at a pre-selected location, I wanted to go on an adventure with the technology, to collaborate with it in the production of new and unforeseen outcomes.
In this, my approach owed more to the flâneur than the engineer. The flâneur or flâneuse of nineteenth-century Paris was a person who walked the streets without a care in the world, an urban explorer on whom the impressions of the city would play and play out. In the twentieth century, the figure of the flâneur was picked up by proponents of the dérive, or drift: a way of combating the malaise and boredom of modern life through unplanned walks, attentiveness to one’s surroundings and encounters with unexpected events. The twentieth-century philosopher Guy Debord, the primary theorist of the dérive, always insisted that such walks were best undertaken in company, so that people’s differing impressions of the group could resonate with and amplify one another. In the twenty-first century, could my autonomous companion perform the same role?2
As well as getting lost, I was trying to think of ways to illustrate what I was coming to think of as the umwelt of my self-driving car. Coined by the early twentieth-century German biologist Jakob von Uexküll, umwelt literally translates as ‘environment’ or ‘surroundings’ – but, being German, it means a lot more than that. The umwelt connotes the particular perspective of a particular organism: its internal model of the world, composed of its knowledge and perceptions. The umwelt of the tick parasite, for example, consists of just three incredibly specialized facts or factors: the odour of butyric acid, which indicates the presence of an animal to feed upon; the temperature of 37 degrees Celsius, which indicates the presence of warm blood; and the hairiness of mammals, which it navigates to find its sustenance. From these three qualities, the tick’s whole universe blooms.3
Crucially, an organism creates its own umwelt, but also continually reshapes it in its encounter with the world. In this way, the concept of umwelt asserts both the individuality of every organism and the inseparability of its mind from the world. Everything is unique and entangled. Of course, in a more-than-human world, it’s not only organisms which have an umwelt – everything does.
The umwelt has long been a useful concept in robotics as well as biology. It’s easy to see how the example of the tick’s simple rules could be adapted to provide the basic framework for a simple, autonomous robot: ‘move towards this light; stop at that sound; react to this input.’ What then is the umwelt of the self-driving car?
The simple intelligence at the heart of my car is called a neural network, one of the most common forms of learning machine in use today. It is a programme designed to simulate a series of artificial ‘neurons’, or smaller processing units, arranged in layers like an extremely simplified brain. Input signals – the speed of the car, the position of the steering wheel, the view from the cameras – are fed into these neurons, sliced into component parts, compared, contrasted, analysed and associated. As this data flows through the layers of neurons, this analysis becomes ever more detailed and ever more abstract – and therefore harder for an outsider to understand. But we can visualize aspects of this data. In particular, once the car has been trained a little, we can see what the network thinks is important about what it sees.4
Visualizations of a neural network’s way of seeing.
The images above illustrate a little of that. The first is the view directly from the car’s main camera: a road in Parnassus, disappearing into the mist. The second is how that image looks when it has passed through two layers of the network; the third is the fourth layer. Of course, these are visualizations for human eyes: the machine ‘sees’ only a representation in data. But these images are data too: the details which remain in the image are the details which the machine thinks are important about the image. In this case, the important details are the lines along the side of the road. The machine has decided from its observations that these lines are of some importance; as indeed they are, if the machine is to stay on the road. Like the tick’s sensitivity to the temperature of mammalian blood, the lines on the road form an important part of the car’s umwelt.
And in this observation, we find the point where my umwelt is entangled with that of the car. I see the lines too. We share at least one aspect of our models of the world – and from this, too, whole universes might bloom.
To dramatize this revelation of a shared model – and therefore a shared world – I did something which felt a little mean. As much as I’d grown into our collaboration, and fond of my automated companion, I decided to test it. And so, using several kilo bags of salt, I poured out onto the ground a solid circle a few metres in diameter, and then around it I drew a dashed circle. Together, these circles formed a closed space in which the (European) road marking for ‘No Entry’ is projected inwards. As a result, any well-trained, law-abiding autonomous vehicle, on entering the circle, would find itself unable to leave it. I called it the Autonomous Trap.
Autonomous Trap 001, Mount Parnassus, 2017.
This crude attack on the machine’s sense of the world was intended to make a few points. The first is political: by working with these technologies, we can learn something of their world, and this knowledge can be used to turn them to more interesting and equitable ends – or to stop them in their tracks. Faced with the kind of corporate intelligences we encountered in the Introduction, this is useful knowledge.
Secondly, it asserts that the tools of the imagination and aesthetic representation are as important in an age of machines as they ever were. Art has a role to play here, and we can intervene in the development and application of technology as effectively from this position as from that of an engineer or programmer. This is useful knowledge too.
Mostly, though, I wanted to emphasize the aspects of the world which the AI and ourselves perceive in common: our shared umwelt. My video of the Autonomous Trap subsequently went viral, and I have the feeling that people appreciated the chutzpah and the whiff of black magic more than the collaboration: in an age of Uber, air pollution, mass automation and corporate AI, there’s something pleasing about stopping the robot in its tracks. Nevertheless, the fact remains: we share a world with our creations.
If seeing the relationships between humans and artificial intelligences as creative collaborations rather than open competitions produces such interesting results, what else might be possible? What other intelligences share worlds, and what is to be found in their encounters and imbrications? If contemporary ideas about artificial intelligence seem to be leading us down a darkly corporate, extractive and damaging path, what alternatives exist?
The current, dominant form of artificial intelligence, the kind you hear everyone talking about, is not creative or collaborative or imaginative. It is either totally subservient – frankly, stupid – or it is oppositional, aggressive and dangerous (and possibly still stupid). It is pattern analysis, image description, facial recognition and traffic management; it is oil prospecting, financial arbitrage, autonomous weapons systems, and chess programmes that utterly destroy human opposition. Corporate tasks, corporate profits, corporate intelligence.
In this, corporate AI does have one commonality with the natural world – or rather, with our false, historical conception of it. It imagines an environment red in tooth and claw, in which naked and frail humanity must battle with devastating forces and subdue them, bending them to his will (and it is usually his) in the form of agriculture, architecture, animal husbandry and domestication. This way of seeing the world has produced a three-tiered classification system for the kinds of animals we encounter: pets, livestock and wild beasts, each with their own attributes and attitudes. In transferring this analogy to the world of AI, it seems evident that thus far we have mostly created domesticated machines of the first kind, we have begun to corral a feedlot of the second, and we live in fear of unleashing the third.
Copyright © 2022 by James Bridle
