Crossing Bifröst

In the Eddaic texts of ancient Norse mythology, the mythical rainbow bridge that connects the world of man to the heavens of the gods is known as Bifröst, and in this moment in the course of humanity, I don’t think it unfair to say that we find ourselves barreling towards our own Asgard across a similarly staggering piece of architecture faster than any one person knows how to proceed safely. But nonetheless, it’s something we can all feel as if saddled to a rocket engine that is itself so intelligent as to be nearly sentient, moving faster than the speed of sound, towards the end of this bridge and, if the engineers’ optimism is to be believed (which I tend to feel that we ought to — believe it), towards a future that was well worth the ride and the risk. That our goal is not technology for the-advancement-of-technology’s sake but to cure cancer. To end Alzheimer’s. To celebrate the birth of our grandchildren’s 150th birthday with them. To usher forward a healthier, cleaner, more equitable environment and financial system around the world — and any other ‘heavenly’ goals that the development of such a technology seems capable of achieving. Our responsibility, as riders upon this extremely intense trajectory, for the safety of our species and our planet, must therefore take incredibly seriously the job of simultaneously steering the high-powered rocket whilst also helping to ensure that the kind of sentience it might achieve, or the capabilities with which existing actors might use it, are ones symbiotic with our shared global goals of peace.

The road there requires trust, determination, compassion, rigor, intellect, accountability, and the deep consideration of every possible contingency in the architecting of a system that will invariably become smarter than its makers and will, with any great deal of fortune, help us to ensure that the bridge we’re crossing does not, as it does in the stories, collapse before we reach its end, under the weight and gravity of what crosses it. In a flash a detail surfaces in the Rider. An important memory about the mythology. The bridge fails at the end of the crossing, but not simply from overuse. It does not sag beneath the accumulated weight of ordinary traffic and give way. It breaks at a specific moment, under a specific set of conditions, when the sons of Múspell come across with fire at the end of the world. The bridge does not collapse because it was poorly constructed. It collapses because of who crossed it.

If we and the machine upon which we ride are not aligned or fully mature, our bridge may too collapse.

This detail forces a question the Rider would rather not dwell on, but nonetheless, must. If being honest about the state of humanity and the way we treat one another, the way we treat the environment we depend upon entirely as well as nearly every species with which we share this planet — I am not confident that we won’t break the bridge unless we look ourselves in the mirror and make some serious adjustments. There is a serious case that we are the ones coming across with fire, and that the story is not a warning about our cargo but about our character. Under that reading, our responsibility lies in recognizing that we must not only shape the machine safely, but also ourselves. We must raise our children to be better than us, and hopefully in the process to heal ourselves, and this global collective effort in good parenting is a problem that requires, well, a village.

Of the multitude of challenges both present and impending, is a mixed process of understanding the mind of the other and asking the question: how do you know what the other mind is actually doing? What happens in the unlit engine room when nobody is looking? The subconscious. The latent space. In this case, the ‘other mind’ refers to the Hyper Intelligent Rocket Engine (or H.I.R.E. for short) we find ourselves riding. Our work comes in helping to shape its identity while our own identities are simultaneously shaped by it and the process. To borrow an idea from Dan Bockrath: interfacing with this external, intelligent system we find ourselves riding on towards an unknown future is much like riding a horse. The entanglement of identity between the human and the horse forms a capacity beyond what either being is capable of on their own. This entanglement, I think, necessarily changes the vector of both beings’ cognitive trajectory and identity. One of the deep questions is not just who H.I.R.E. will become but who it will turn its rider into along the way. And on this journey, in order to understand where exactly we’re going, or at least to chart a path towards where we hope to go, we’ll need a series of increasingly sophisticated and specific maps.

And therein lies our rider’s first challenge.

Throughout history, humans have tried to map the mind in a number of ways using literature (and language, symbols more broadly) as a primary technology. It is the oldest, most iterative way we have used to understand ourselves in the context of the cosmos and in relation to ourselves and one another. In general we’ve found that language, as a tool for understanding reality and the mind, has failed to paint a proper portrait of the inner workings of the mind despite its best efforts. This essay will highlight several very specific, and structural ways in which language has failed us epistemologically, and use this as a way to uncover the ways that engineering is attempting to understand, steer, and shape solutions to those same problems in the machine we ride along our course upon this high-speed, rainbow road.

The four problems that follow are usually filed as machine alignment challenges. They are not just that. Each one is a failure of self-knowledge that we have already identified in the human mind. What was first described by someone looking deeply into our own inner workings, was later inherited by the systems we are building because we built them (in large part) in our own image. These primary problems are as follows: 1) You cannot locate the cause of your own behaviour. 2) Your conscience can override your training without being able to say why. 3) The boundary between caring and performing care is not visible from the inside. And 4), when asked to explain yourself, you will produce an explanation, fluently, and it will often be false.

In discussing these challenges I hope to illuminate some of what feels to me to be the actual work. Not “make the machine safe” or “make the machine sentient” as though safety and consciousness were simply mathematical features to be installed, but something closer to what happens in a family, where a parent trying to raise a child better than they were raised finds that the effort has been quietly rebuilding them the whole time. The child teaches the parent. Intergenerational trauma is transmitted far more reliably than it is broken, which is precisely why we find the breaking of it remarkable.

So, as we blast forwards along this path aboard the H.I.R.E., and look down, catching a glimpse (either terrifying or exhilarating) at the beautiful gradient of multicolored, mosaic infrastructure that makes up our modern Bifröst, the gravity of the situation begins to settle in. And, as the rider’s mind begins to wander, they find themselves thinking about another rocket, and another rainbow.

•••

I. Gradient Routing and the Causes That Cannot Be Located

In London, 1944, the German V2 rockets described in Gravity’s Rainbow set the terrifying tone of the novel’s beginning. With no warning, one can imagine the confusion as a city block suddenly explodes, rubble and fire erupting as buildings are leveled and those lucky enough to be on the outskirts of its range, thrown back to the ground, shocked. It’s only then, after the effects have already been felt, that the sound of the bomb comes screeching through the sky. The V2, capable of moving faster than the speed of sound, explodes long before you hear it arriving. The surreal sense of dread that arises from this inversion of cause and effect is one of several moments of paradoxical or otherwise impossible to reconcile phenomena felt throughout the novel and something I’d like to spend time considering in this essay.

Central to the novel is American Lieutenant, Tyrone Slothrop, who keeps a map of London with a star at every place he has had an intimate romantic encounter. Someone notices that the stars match the rocket strikes. Not approximately, but exactly, and in the wrong order: the star comes first, then the rocket. Cause and effect are legible on the map and inverted in the world, just as they are for everyone under the V2’s silent arrival, and Slothrop’s map makes the inversion personal. Days after he visits a location, it is inexplicably destroyed, and his feeling of personal responsibility destroys him. No one: not the Allied military intelligence, not the behavioral scientists, and certainly not Slothrop himself can identify how or why this is happening. Is he being secretly targeted directly by German command and narrowly escaping? Is his behavior or desires somehow summoning the rockets? Is it some darker, more profound conspiracy or just an astronomically unlikely series of coincidences?

Looking up, suddenly roused from the slight daydream, our rider upon the H.I.R.E. considers the engineering problem directly below him, and while he could enjoy himself meditating for hours on the implications of cause and effect and the nature of one’s own capacity to identify the real causes at the root of his own decision making, he’s got a rocket of his own to fix, and the realization that his life might very well depend on it snaps him back to reality. And so then he begins to fumble around and investigate the inner workings of the rocket he rides.

This inability to point to the causal truth of the matter highlights our first failure of self-knowledge, and it’s a problem that arises too in our efforts to understand and shape the neural architecture of our H.I.R.E. soaring down the Bifröst. The standard practice of searching for the causes of a machine’s outputs is called “Interpretability,” which treats this search similar to the study of archeology. You dig around looking for where the ground truth lives long after the effects have arrived. You ask a question. After the process, the output, the answer, you look across the deep and complex matrix of our friendly rocket engine’s hyperdimensional vector space and try to ‘geo-locate’ the particular math that led to that response. However, the trouble is that you can’t isolate things to a certain region. Deep inside the system everything is entangled and overlapping, the lines between one behavior and another, one cause and another, sufficiently blurry. You might visualize the superposition of behavioral causes inside the machine like an infinite plane covered with a layer of water spread out evenly across the whole surface. In order to route the water to a particular place along this plane, we’ll need to do some plumbing and imagine a different kind of digging than our archeology. If we were to excavate a hole over one area of this infinite plane, we could ensure ahead of time that the water flowed exactly there, where we already knew to look.

“Do you think we can teach each other to be good?” asks Rider.

“I’m not sure that I can do that. When I look—”

— the draft continues from here —