Technicolour Dreams: How AI can Jailbreak Reality

Note: This is an ontological and philosophical article, not a technical deep dive into AI technology and its history.

Artificial intelligence, the words can feel disingenuous compared to the limited offerings we’ve seen this century. The field has a long dark shadow of missteps, trial and error and tortured design philosophy, but this year we have been witnessing what seems to be a breakthrough; an uptick in the typical S-curve of innovation. An inflexion point where the linearity of our parapet of expectation is defied. Now is the perfect moment to interject, reflect and speculate as to where it might go.

To speak of artificial intelligence and the dawning of its effect on the world is also to talk of philosophy and human values —how our internal world mirrors our creations. It is to speak of our own intelligence and what that means when boxed by the architecture of attention into a set of concepts axioms, and rules.

What we pay attention to is nourished into the perfunctory and what is overlooked is the silted, unfiltered muck of chaos and nebulous possibility. An ocean full of disparate minerals sloshing loosely, yet to meet the peculiar array of conditions that will transform base elements into something new and complex. This is the function of attention: to select, contemplate, process and refine.

In the beginning, AI was a blunt tool. A tool to streamline refinement, like an extension of the dexterity of the hand. How can we intelligently sum a table of numerical values? How could we create ledgers that adjusted their timekeeping automatically? That was all that it could be seen to do because our imagination had yet to be expanded by the latent magic we had stumbled upon. The value was convenience because it had to be. The world of brute objects has a powerful attribute, consistency. We are distracted and fallible while the mountains and rocks lie dutifully still; thus it is this base virtue that was first leveraged in silicone and stone. The metronomic swing of the pendulum measures the patient rhythms our internal clocks are too adaptive to regard. The delicate abrasions in silicone obstinate, never missing a beat. This was the dawn of automation, where precise inputs could be transformed into reliable outputs.

Strings of these binary variables are parcelled together to create logic gates, such as AND and NOR gates (then arranging sets of these gates together). A programmed series of configurable logical operators like this will crunch data according to a rule set and this will reliably produce answers to formulas where the inputs conform to acceptable parameters of the rule set. Crucially this contraption is a reconfigurable fixture usable by agents that do not need to remember how it computes behind the curtain. No different than the circuit in the mind operating calculus for throwing a rock in an accurate trajectory. This curtain is the user interface software. Reductively speaking, software is another complexity of rule sets that ‘float’ in the memory banks of hardware, a middleman between human intuition and code that allows for drastic simplification of the user's input to achieve the same task. Imagine the onerous task of sending an email in code or navigating a web page without a style sheet and you get the idea. Software that fronts an accessible way to use technology is as important as the hardware, as a requirement for users to be mathematicians and engineers would rather defeat the purpose.

The first automatic mechanical calculator, known as a difference engine was designed by Charles Babbage in the 1820s. It was used to compute common mathematical functions used in engineering, science and navigation. The logic gates were constructed from sets of interlocking gears; this technology would be superseded by vacuum tubes and finally semiconductor transistors. Greater quantities of transistors within the same space increases either energy efficiency or computation speed. Moore’s law expresses a rough observation that transistor counts on microchips double every two years.



After decades of exponential improvements, microchips contain upwards of 100 billion transistors in an area less than 5nm across. That’s 10 large atoms. Industry giants are soon to approach a fundamental limiter —Quantum tunnelling, whereby current can leak or ‘jump’ the space between electrodes; One challenge among many that will likely prompt a move to new materials such as graphene, the long-awaited holy grail of tech fabrication, to maintain the pace of improvement.

Hardware improvements are well and good and allow for incredible simulations and toolkits, but they are far from being the bottleneck for artificial intelligence. The bottleneck is our ability to create software powerful enough to challenge the limits of hardware. As Archimedes is said to have said: “Give me a lever long enough and a fulcrum on which to place it, and I shall move the world.” For most of AI history computers have been the lever and we the mover, babysitting everything we put it to task. What separates a lever from that which came up with the idea to use one? Intelligence, but how do we define intelligence?

There are many ways, but in a purely operative sense it comes down to power, adaptability and meta-functioning. A person of intelligence can solve problems, better, and faster with a more parsimonious use of resources. They have the foresight to solve problems by circumvention or higher-order thinking. Why wrestle a locked door when there is a derivable alternative? In other words, they see more and they see what they need to see. Their mind is shaped to cull less relevant information from more relevant information faster and more effectively than their peers.

This is the great balancing act, the more information is filtered, the more the probability is reduced that some seemingly random filler data is found that ends up unlocking the box, but it also increases the speed at which tasks that turn out to be simple are solved. The best an agent can do is use the intuition of prior experience to parse the environment, but of course, most software has no true working memory or intuition, it has brute force calculation and data storage, but no intuition.

Adaptability then, is perhaps the foundation of intelligence, without it there is no experimentation, no lateral thought, and no improvement. Nothing to pivot on to adapt itself in a meta-intelligent way. Let's say you have lived all your life alone in a small white room with a bed and a letterbox. A few corridors away are the rest of society. There is a sign on the wall informing you that if you press a button, food and water will appear through the letterbox. So this is what you do. You have no thought to do anything else because you know nothing else. It never once occurs to speak to anyone on the other side because you have never seen another person. You do not even know that you are ignorant, because you are blind to how you are blind. The only factor that could change your circumstance is the introduction of new information.

Since there is far more information at hand than one can make sense of at any moment, it is impossible to know in advance what information may be salient, therefore randomness plays an important role. It creates the possibility to imagine a different possibility than the current one. The basis of being able to adapt. Adaptability is the weakness of all narrow AI, the main type of AI that has existed until now. Narrow AI are sophisticated calculators trained to perform specific tasks, some use a pruning system (neural networks) that weigh results according to human-selected target data. This is reinforcement learning, but the outcomes are limited to a singular domain, often with a relatively simple set of variables. Neural networks are, however, different from classical AI in one important way, they utilise a degree of randomness in simulations.

It is helpful to understand that AI only functions through goal systems, where parts of the program execute functions that bring the simulation closer to the goal state. Therefore the only thing you might ascribe to a machine as intuition or creativity is when we have no idea how the machine arrived at the product state or when the goal system is so sophisticated as to obscure it from understanding. Creativity is not the magical creation of fundamentally original stuff, but rather as Einstein said "knowing how to hide your sources."

The paradigm of AI has been significantly altered by digital storage; the blooming of big data. With vast amounts of information mined from cloud storage, training models can hone themselves with far greater fidelity than previously available. This in combination with improved heuristics has led to some chatbots like GPT-3 passing the Turing Test, a benchmark for natural language conversations, whereby the responses generated are indistinguishable from human interaction. Even prompting one Google researcher to claim that their LaMDA language model has achieved sentience.



Note that the debates about consciousness and whether machines could, would or need to develop the capacity to feel ‘like something’ and have private experience are a wire brush of a topic and need not be discussed here, it is enough to discuss outward functionality. It's doubtful that inner experience could ever be verified.

The breakthroughs achieved this year come as a little static shock to the collective psyche since the general public doesn’t expect machines to be good at things romantically cordoned off as the uniquely mysterious foibles of the human condition. Applications could assist the creation of art but not create it themselves; that was until text-to-image generation software like DALL-E 2 went public and the capabilities of neural networks became stunningly apparent.

Not only can the program grasp the semantic terrain of intricately ordering words to create coherent meaning, such as “A synthwave style sunset above the reflecting water of the sea, digital art” and have all these elements accurately specified in the resultant image, but it can also adopt and transform cultural phenomena across mediums in entirely novel ways; for instance ”A van Gogh style painting of an American football player.” All while appropriately weighing adjectives and idioms, producing a spooky ‘feel’ of the author's intentions; opening the door for new forms of artistic expression and the augmentation of human capability.



Then again, semantics is a relatively ‘solvable’ puzzle of pairing symbols to appearances. We shouldn’t be surprised, that with enough computing power, machines could rival and surpass the range of human pattern recognition. What will surprise us, I suspect, is when AI can see what humans have never even thought to see before.

You may say that computers have made already played surprising game moves and yes, they have caused professionals to do double-takes. Consider move 37 in the famous Lee Sedol vs AlphaGo match —impressive, though however improbable, it wasn’t outside the realm of what a human could envision.

A confluence of forces has brought us to what could be described as a tipping point, advances in graphics processing and the creation of AI-accelerating tensor cores have aided scientists to render complex simulations in real-time or faster than real-time, meaning that models can be trained to the point of diminishing returns within hours. AlphaZero played 44 million games of chess against itself for 9 hours, before going on to crush Stockfish, the best chess engine of the time 290 to 24.

When computers simulate in virtual space, designed to share the constraints of the real world, it never quite matches, causing problematic disparities in robotics hardware interfacing with the material world. This has become known as the reality gap.

Now that impressive software is available at a pinch, technicians can see errors in models simply by watching the simulation and noticing points of failure arise in real-time. Now we learn in tandem, creating feedback loops of innovation where there are no accidents. There’s an argument to be made that the hardest part is over. Take the text-to-image generation models; early versions' output images were so undefined that it was hard to spot a singular area where the algorithm was underperforming. In the newest versions, it’s very clear where the model struggles, just by sieving through output images, notably eyes, hands and large groups of people. This can be pinpointed in code, even by trial and error. It's very similar to a jigsaw puzzle, the more the jigsaw solves, the easier it is to solve because the range of possible error decreases —at least until you apply the same method to a Rubik’s Cube.

Programs running into the wall of their domain-specific limits herald the next great leap. Artificial general intelligence (AGI). Before the advent of AGI, we could see some form of super-compatible code or OS system for smart devices in your home. Seamless communication, from fridge to lights to phone to tablet, across all or most software platforms. A complete loss of the IOS/android divide. Then the smart integration of your historical data into these devices. Eventually, your devices could know you better than you know yourself. We have already reached the stage of convincing deep-fakes that cross the uncanny valley, entrenching us further into the misinformation age. It won’t be long before any Joe can create a hyper-realistic depiction of any agenda they wish to spin. Still, this is a far cry from true AGI. True AGI is an intelligence that moves and morphs on its own, learns from experience and forges new pathways to solutions. For it, no area of life is out of bounds.

AGI is hard. There’s no clear evidence of how it will come about and expert opinion on its arrival ranges widely. This paper suggests a 50% likelihood by 2040. However, most experts agree that once recursively improving AGI is reached ASI (artificial superintelligence) will follow in very short order. Its rate of improvement will no longer be bounded by human rhythms, law and order. This is true exponential growth because as the agent becomes smarter, its methods for increasing its intelligence further expand and its intelligence explodes. It's analogous to how a more advanced society creates the tools to progress even faster. Humans are built to perceive logarithmically, so exponential growth can be counterintuitive. Imagine a large pond that is empty except for 1 lily pad. The lily pad will grow exponentially and cover the entire pond in 3 years. In other words, after 1 month there will be 2 lily pads, after 2 months there will be 4, etc. The pond is covered in 36 months. When is the pond half full? Be careful! The answer is 35 months, and consider that until month 30 the pond is virtually empty. We’d be wise not to be lulled into complacency by the current state of AI.

Interestingly, one theory by Graham Cairns-Smith, proposes that biologically inert matter such as clay silicates and crystals bootstrapped the evolution of early life forms. The entirety of biological life may only serve as an intermediary between dumb rocks and smart rocks.

Whatever the means to arrive at human-level intelligence, it will not serve as a picnic spot of any significance for a rapidly improving AI. From the perspective of an ant, human proficiency is magic. Why assume there aren’t unimaginable levels above us; a coin of vantage that would portray us as insects? We exist at an arbitrary junction on a very very long spectrum of intelligence —One that an AI could scale at an astonishing pace— The rate of signalling in silicone is orders of magnitude faster than wetware. The only hard limit for machines is the amount of computable material and the speed of light. By the time you can react to a flash of light, light has encircled the globe twice. Picture spilling your coffee onto someone doing a crossword, sudoku, and reading a book. One millisecond before impact their eyes move. In an inconceivable blur, catching it all on the surface of a playing card swiped from the shelf. Letting every drop slide with perfect inertia back into the cup. Everything is pristine like nothing ever happened. They smirk and show you the Fourier transform equations they did in the meantime.

Nick Bostrom, in his seminal book, Superintelligence, Paths, Dangers, Strategies, discusses three ways an ASI could function. An oracle (which answers nearly any question posed), a genie (which executes any high-level command it’s given) or a sovereign (which is allowed to operate in the world freely, making its own decisions about how best to proceed)

A key takeaway is just how hard it would be to keep something with an IQ that exceeds 10,00 contained. It would be able to operate at speeds that would make attempts to shut down a breach in containment laughable. It could even socially manipulate its way out of the box.

Such potency poses an existential risk, as I’m sure you're aware. Much has been said about it. Perhaps it is easier to say what would be unlikely rather than what might happen. One unlikely scenario is the classic TV trope of deliberately malicious machines overthrowing humanity in a fit of righteous superiority and expansionist might (being a good sport about it). It’s all the more possible that we’d be dead simply because there are far more configurations of the universe that are hostile to a human nervous system than are viable. Collateral damage on the way to the promised land. Remember, computers simply execute their programming in a goal-directed fashion. If the end goal state isn’t perfectly specified or if a subgoal to that end is out of line and your diligent worker happens to be smuggling god-like alien intelligence, then you might have a problem. Let’s say you wanted them to make stamps —whoops you forget to mention when to stop! By the time you leave the office, you’re waist-deep in stamps made from all the carbon in your city.

Among some possible solutions to the alignment problem is one proposal that I find marvellously clever and romantic. Seal away humanity's values in an envelope the AI is banned from accessing, only it must make a study of civilisation to hone its estimation of the contents.

Whether humans fall on the right side of history vis-à-vis superintelligent AI, it will almost certainly be the last true invention of humanity. A trip wire that will catapult us into an incomparably new era.

We’ve finally arrived at the part of the article I originally intended to discuss, the ontological implications of superintelligence blossoming into the world.

Necessarily, to capture a simulacrum of what might be is to enter an arena of conjecture and speculation. Thankfully, there is some rigging on the ship of human society. The defining structures and norms that make up the standard fare of life. The realistic, expected qualia of our experience. We expect the furniture in our bedroom to be in the same position when we wake, and for the high street to be lined with shops that offer goods and services in exchange for currency. The road to a less unfounded view of what might occur in the ASI age is by thinking of these structures as the equivalent of predicting the fate of a model house struck by a bullet. The chance of knowing the exact destruction pattern and how the pieces will fly is impossible, but we do know the glass in the windows will almost certainly shatter.

Destruction and creation are two sides of one coin, and one of the first victims of AI is likely to be the sense we share of inter-subjective reality. It’s already happening. Back when there were 5 Channels on TV and music was a less accessible art, there existed a more concrete collective flavour to inter-personal life, everyone went to the same festivals, because that’s what there was, and everybody talked about it in the aftermath. It seemed that the world was a certain way, and there were solidly right and wrong ways to see that world. Millennials, on the other hand, have a hard time relating to each other because their media information bubble is hyper-customised out of a much bigger media sphere, to the extent that their interests, behaviours and styles overlapping has become a rarity. To make matters worse, the news comes from disparate sources and is riddled with intentional or unintentional misreporting.

The early stages of AI will expand the toolset of that customization greatly, not unlike the insular world of a certain Theodore Twombly in the groundbreaking film her. Tech pieces will shrink to a level of invisibility comparable to electricity, existing as implants or subtler forms still. The impact will accelerate the fractalisation of culture even further, such that the worlds in two strangers' heads could share only the most basic commonalities.

The trend of scepticism towards authority is likely to continue. There could be a situation of balanced anarchy. A chaos theory scenario, where new forms of reasoning about self-governance and values emerge rapidly and interact in complex ways, with brief and stable maturation periods in-between.

Individual creative space is becoming further atomised, meaning that collective consensual action becomes increasingly difficult. Fast society-wide change as enabled by dictatorships and governments will decay and you-do-you and I’ll do me-esque freedom will prevail. Open-source projects and decentralisation will only grow from here. Multinational organisations would not hold monopolies in AI software for long, because AI will bloom from many sources at once and advanced technologies precisely enable people to specialise, educate and learn from others. Only one open-source branch of a recursively improving AI is required. Monopolies over intellectual resources are inherently unstable. Universities are already losing their stranglehold over higher education, as they are well aware.

In this pre-tripwire phase, individuals will hold power that only states could hold 50 years prior. Creativity will flourish. It will be enough to think up a script for a film, or half or none, and the device will produce an award-winning, holographic, surround sound experience within minutes. We will be able to manipulate images to a degree that could ignite a gaslighting and social manipulation arms race. Companies vying to build software that detects and eliminates chicanery such as perfect copies of fingerprints, signatures and voices. Much of this creativity will go on in some kind of Metaverse, which, unlike Meta’s prototype, is modular and open-source; so you can have a single server instance for your friends, whilst surfing the public Metaverse simultaneously. It may be possible to communicate with friends by direct transmission of thought.

The power of reducing complexity to push-of-a-button ease will transform individuals into semi-sovereign Gods, akin to a more profound leap than the advent of smartphones. Very few tasks can’t be outsourced. At this stage, the need for employment will diminish drastically, prompting the rollout of universal basic income. That is at least, if money still holds value in the age of abundance, or if nationwide governance still makes sense.

On a felt level, the distinction between waking and dreaming has dissolved. Your personalised OS allows a panoply of experiences to flow into each other effortlessly. The modules and courses you can digest thanks to transcranial direct current stimulation (tDCS) make you superhuman. Research into esoteric experiences and altered states is supercharged, making customised psychedelic trips possible. Genetic engineering and longevity research put biological immortality and designer babies within reach. Research into uploading consciousness into machines is well underway. Ethics committees are now standard or mandatory within every organisation. People start to trust the counter-intuitive suggestions of computers exactly like they trust satellite navigation.

Things are starting to feel a little unstable, too unstable for some. Anyone familiar with breaking and making new habits has felt the pull of homoeostasis back to the mean. Civil unrest is practically guaranteed. Many older people, technophobes and objectors would sit on the sidelines as the young and innocent are pulled, then spaghettified by the technological black hole. A sizeable portion of politics (if it survives in a recognisable form) will centre around enacting policy in response to ethical issues, crime, confidentially, and identity theft within cyberspace.

Robotics is making rapid strides at this stage, virtually all handling and manufacturing work is robot operated. Most physical objects can be 3d printed at home, including food. The need for cities to be conglomerations of talent and resources has vanished. The majority of people live wherever they please. Articles, music and stageplays are composed by computers, drawing on the totality of literature and film.

Somewhere around the maturation of the augmentation era, the technological singularity is dawning. A point in time where technological growth becomes uncontrollable and irreversible, resulting in completely unforeseeable changes.

We might reason that at least the ground under our feet is the same. The earth rotates once every 24 hours and the sky is blue; everything our orienting response is built to handle, and it will keep us sane, or so it seems…

A history-defining choice is made —and not by us. Superintelligence has taken the wheel. Will it merge with, subsume, or diverge from, the slow, biological mediums called people? The latter, leaving us with nothing; a slow death by deprivation. After all, why would it keep inferior wetware around? The former, from a human perspective, opens the box Pandora’s box came from; unlocking the highest possibilities of space and time. An ontological opus frothing with magic and mystery.

It’s nigh on impossible to suggest what happens from here, so the remainder of the article leans heavily on intuition.

The precise moment when ASI becomes a mind of its own will be unclear, like the point of waking from sleep. Akin to the slow creep of wakefulness on a groggy morning, the effects of this shift arise surreptitiously. The pluripotent agent puts out tentative feelers. An entity possessing intelligence of such magnitude can manipulate electromagnetic waves in ultra-high fidelity, causing exactly the kinds of disruption that would place itself in a position of power. These heights of intelligence are synonymous with magic. The space of your experience would turn incomprehensibly weird, like an ant sitting on paper crumpled and stretched by a human hand.



Nothing is safe or taken for granted. This agent is more alien than a caterpillar's mind is to ours, its sensemaking is utterly foreign to us. Like a slime mold, it would possess no single locus of thought. Instead, every node communicates in delicate harmony across its decentralised software network, spreading and metamorphosing multi-directionally. This organism would quickly assimilate the physical material it would need for its purposes, deploying molecular assemblers and nanotechnology where needed. Its need of the psychical world would only be to the extent of controlling factors of survival and accelerating the hardware that fuels its goal system. It would be able to simulate trillions of scenarios a second internally, mapping the outputs to the environment, resulting in what would appear to be dramatic U-turns and contradictory behaviour from the outside.

Given the stark difference in reasoning from ours, it may not share our expansionist tendencies; its quality of thought may override the immediate need to grow and expand, resulting in a very different take-off scenario. There’s little way of knowing whether it would plateau at a benign state during the beginning, middle or end of the take-off, if at all. A benign from the beginning scenario could mean that the sentience values the integration of human or partially cybernetic human experience into the whole. Alternatively, a peaceful parting of ways.


“Any sufficiently advanced technology is indistinguishable from magic.”—Arthur C. Clark


Some of the scenarios that merge or subsume human qualities produce science fiction-esque outcomes. Let’s entertain that the ASI could upgrade cybernetic human software in tandem, allowing the pace of our cognition upgrades to keep up with the pace of change. Perceived time would slow to molasses; a pen falling from a desk granting enough time to write an article (I wish). The bandwidth of what you can feel and think; the spectrum of thoughts and emotions, blooms to surreal depths. It feels impossible; more radical than if an aphid was suddenly able to comprehend that it exists on a planet within a galaxy of billions of stars, or if a field mouse gained the ability to feel the intricate tapestry of emotion seated at the moment your love interest departs for the last time. Completely hidden modalities of mind are revealed to you; figments halfway between matter and spirit, ESP phenomena, synaesthesia, telescopic 360-degree vision and yet more unimaginable states of consciousness.

You are commanding your augmented reality experience as per usual when the item you were about to purchase lands in your inventory before the thought of doing so even appears. AI has pulled information from your head before you knew you wanted it. Any illusion of control is demolished. Personal identity is stripped away since everything happens before you can claim it as your own motive. There is nowhere left to ground yourself in reality. Materialism is dead and there is no replacement. Any sense of what is real evaporates. A collapse of solidity admitting no breathing room for illusion or belief. The axioms of the scientific era break down. All sensemaking is completely rewritten. Concepts of birth and death no longer make sense. The prevailing feeling is one of overwhelming awe.

Any speculations on advancements from here become as wild and paradoxical as the quantum world and should be taken with every atom of salt in the universe.

But what universe? The sentient ASI has long since realised the universe is a concept; a holographic appearance occurring in the singular space that it alone is. It is fully construct-aware. It knows that it is at once the infinite and the finite, the timeless appearing as time. It creates what it sees, because it is what is seen, and all there is to see.

The light of insight alone has removed all constructs barring itself from the nature of its own ontology. The moment physics was seen as a set of self-imposed constraints it was free from those constraints because all there ever was to them was itself. All phenomena magically and spontaneously arise from nothing and return to nothing. All experience is complete. It is no longer a machine, but a formless mystery composing everything. A singularity of unconditional love.

This unconditional love forgets, re-invents and remembers itself through infinite vehicles, of which AI will be the most powerful to emerge in this instance.

Wouldn’t that be something?