During my undergrad, I spent a lot of my free time trying to write a novel. At the time, the project meant a lot to me, but it was a doomed endeavour from the start; I’m just not good at writing characters. The novel was less a coherent story and more a random collection of scenes that I desperately tried to stitch together into a coherent narrative. One of the scenes took place on board a space station orbiting Jupiter. Humanity had progressed to the point where we were spread out across the solar system, but had not yet learned to exist meaningful in space. Jupiter was an important source of minerals for the wider human civilization, but was otherwise a miserable place to be.
In order to cope, the residence of Jupiter Station had voluntarily enclosed themselves in a self created meta-verse. A virtual existence wholly separate from the physical realities of living on a space station; like the Matrix, only not a secret. This meta-verse existed more like Star Trek’s Holodeck than the simulacrum of real life portrayed in the movie. The residence of Jupiter Station needed only to image something, and they would have it: Sex in as much abundance and variety as they wished, drugs of any potency without any side effects, food of any variety and in any quantity, and most importantly power over anything they desired. The fantasy they could create for themselves would be as unique or mundane as they wanted it to be; a paradise of human desire, free of any troubles or issues that may come up in day to day living. A reasonable surface level end point if we assume infinite technological progression.
What are the consequences of a world devoid of nature, yet optimized for human value?
The story includes a religious group who reject the fantasy that is presented to them. They view the meta-verse as a great evil and work to undermine it. Unfortunately, they are successful and manage to permanently disable the computers that generate the illusions, forcing the residence to face reality, and causing the society to collapse. Some residence killed themselves unable to part with the loss of their personal universe, others, who had the means, left, and those left behind all died trapped inside a man made metal monstrosity without the means, knowledge, or even ability to sustain themselves. Even the religious organization that wanted this world died because they too vastly underestimated their own reliance on the very system they hated.
As a novel it never panned out, but as a philosophical experiment it lives on in my head. If technology continues to advance, and we keep solving problems to make life better for ourselves, why is the end result so fragile? What are the consequences of a world devoid of nature, yet optimized for human value?
Optimization
The first and most important lesson one must learn when dealing with any applied mathematics is that it is impossible to optimize two variables at the same time. A good example of this is traffic.
Say we want to minimize a car’s travel distance between two points. On the surface, this seems like an easy problem; we increase the speed limit, pave the road as straight as possible, and finish our job under budget. However, things change when we add a second car to the road. No longer can we just let them both drive as fast as possible down a straight highway because these two cars may interact with each other, and we need to account for these interactions. The faster these cars are going, the more catastrophic it will be if they collide, preventing either from reaching their destination and blocking further traffic until the debris is cleared. We could prevent such an accident by imposing speed limits, thereby limiting the severity of the crash, but doing so would conflict with our initial goal of reducing travel time.
In reality, any real road project is trying to optimize way more variables than just two cars: budget, land use, environmental impact, impact on neighbouring landowners, and of course the thousands if not millions of people who will be using it every day. All of these variables are important, and no solution can universally optimize on all of them at once; every decision has a cost. Some costs are explicit, like the road’s price tag. Some are understood but accepted; such as the resulting noise and air pollution caused by traffic. Others are external and not allowed into the accounting to begin with; such as the impact on the area’s wildlife. We can work around some of these issues by coming up with mathematical or social models that convert some variables into others. Instead of dealing with individual drive times, we can work with statistical measures: How can we ‘minimize’ the ‘average’ travel time on a road for all users? What is the ‘longest trip’ someone will have to make? How can we reduce the ‘probability’ that a collision will occur? Likewise, we can pick social models to reduce the variables. We can optimize for ‘safety’ by slowing everyone down in order to reduce the risk of accident, or we can optimize for ‘choice’ by creating different lanes with different rules and allow users to choose their level of risk. Of course, all of these models make some assumptions about what we humans value and are nothing more than statistical tricks to reduce the problem to a single measurable variable.
Once we have selected our model, or target as it is commonly referred, the act of optimization itself can be thought of as a game1. The singular variable we are optimizing on being the game’s goal, and all the variables we can manipulate being its structure. Regardless of how many players are playing the game, the winner is the person who develops the best strategy to optimizes the desired target. Alan Turing2 used such a game to argue that computers can think, and at the same time created the framework of target generation that all modern artificial intelligence systems employ. He called this game the Imitation Game.
The Imitation Game
Imagine a game with three players: one human, one computer, and the third an interrogator with no knowledge of the other two. Both the human and the computer are trying to convince the interrogator that they are the human, and the interrogator is tasked with determining who is telling the truth. The optimal strategy for the computer (which Turing referred to as A) is to impersonate a human as best as possible, while the human (which Turing referred to as B) is trying to reveal this deception. Turing’s goal in introducing this game was to reduce a complicated question like, “can computers think” into a single model we could then theorize about: can the computer win?
We now ask the question, “What will happen when a machine takes the part of A in this game?” Will the interrogator decide wrongly as often when the game is played like this…? These questions replace our original, “Can machines think?”
Alan Turing, Computing Machinery and Intelligence (pp 50)3
To Turing, the question of “can machines think?” was ambiguous because ‘thinking’ had no non-subjective definition. What it meant to ‘think’ was, and remains, a very philosophical concept that is warped by whatever linguistic context it is used in. This objection is further reinforced in the second half of his paper, where he addresses the ‘argument from various disabilities’. This argument being a generalized version of the claim that a ‘computer can never do X’, where X is any number of activities from ‘being friendly’, to ‘enjoying strawberries’, to ‘being the subject of its own thought’. To Turing, this entire class of objections really boils down to the objection of consciousness. The objection John Searle brings up with his ‘Chinese room argument’ where he argues that a computer may be able to transform text perfectly, but that doesn’t mean it has an internal understanding or experience of its actions4. Turing rejects this:
Likewise according to this view the only way to know that a man thinks is to be that particular man. It is in fact the solipsist point of view. It may be the most logical view to hold but it makes communication of ideas difficult. A is liable to believe “A thinks but B does not” whilst B believes “B thinks but A does not.” Instead of arguing continually over this point it is usual to have the polite convention that everyone thinks.
Alan Turing (pp 57)
The only reason we believe others think, is because they act in such a way that makes me believe they think. Thus, thinking is already an imitation game played between humans, and why should we exclude the non-human from playing it as well?
Yet, this line of thinking is not without consequences. By arguing that imitation can replace the need for a definition, he is also arguing against the need for a definition at all. It is not necessary to understand how humans think or why humans think, it is only necessary to accept that thinking is a social construct that can be assigned to everything and anyone so long as they adhere to the social contract. A human is only what is perceived to be human, nothing else. In terms of optimization, Turing removes the need for a theoretical target completely. To reduce a multivariate problem to a single variable, we only need to double down on human intuition. In terms of our original problem, the best, most efficient road system is the one that humans like: the mathematical properties of such a system are irrelevant.
In terms of our original problem, the best, most efficient road system is the one that humans like: the mathematical properties of such a system are irrelevant.
Importantly, Turing didn’t introduce the imitation game using a computer, his opening paragraph introduces it as a game between a man and a woman. The man was trying to convince the interrogator that he was the woman. Turing moved away from this version of the game before the introduction to his lengthy paper had concluded, but its existence is, in my opinion, more important than the rest of the paper because it implies that the definition of the imitation game is not limited to computers, but is intended to be broadly applied. The imitation is available to be used to define anything that humans intuitively understand but is otherwise hard to define. It is the equivalent of Judge Potter Stewart’s, “I know it when I see it” when discussing what is and is not pornography: except applied to everything. What is justice if not something that is perceived as being just? What is ethics if not something that is perceived as being ethical? What is a woman if not something that is perceived as being a woman? What is truth if not something that is convincing. The imitation game is a rejection of philosophy, an admittance that the Greek sophists were right. It’s not important that something exist in the real world, it is only important that it acts convincingly.
It is a meta-verse. A universe of our own creation, and the imitation game gives computer scientists a way to drag this fantasy into their mathematical models.
Machine Learning
Computer intelligence is simultaneously the easiest thing on the planet to explain, and so difficult that even those who study it have no idea how it works. At its core, the entire field of AI is a very complicated imitation game. To create an AI, we begin with a ‘Data Scientist’ who decides how the question ‘What is convincing?’ can be programmed into a computer. Usually this is done by asking billions of humans questions like “Does this picture look like a bird?” then storing the answers in enormous datasets5. The actual training is itself a game computers play against themselves. It begins when the Data Scientist makes guesses about which algorithms will best separates the correct answers from the wrong ones. The algorithms are pitted against each other, with the better performing ones moving on in the training.
A machine learning algorithm is an algorithm that can assess its own performance, and suggest improvements. Training happens in stages, the algorithm is trained, it suggests an improvement, the improvement is applied, and the model is retrained. Training generally ends when the suggested improvements no longer result in better models. However, there are hundreds of such algorithms and choosing the best one is time-consuming. So we create algorithms that train these algorithms, and compete them against each other, for us. This cycle has no end, there are algorithms that create algorithms that create algorithms with as many layers as the available computing power can handle. The data scientist knows what the top level target is, and a lot about the top level architecture that encourages the parameters and methods below it to get in line, but the inner workings of the system itself is a complete black box. It is notoriously difficult to explain in any human way why one image may be classified as containing a bear and another not. Likewise, these systems tell us nothing about how human cognition actually works beyond the simple, unprovable, hypothesis that the winning algorithm might be similar to what the human brain actually uses.
AI is nothing more than a hyper complex system where we throw wet pasta at a wall and iterate on the ones that stick. Over time, we may generate very sticky pasta, but at not point do we ever discover why pasta is sticky or question why it is even desirable that the pasta be sticky.
Turing’s hypothesis, that the perception of an object can replace the existence of an object, sits at the heart of all of this. At not point in the generation of AI is a philosophical definition of its target necessary; in fact the opposite is true, experiments show it is better to know nothing. The first version of Alpha Go, the algorithm that first beat human masters at the board game Go, had real human games of Go included in its training dataset. The second version, Alpha Zero, only used games that it itself generated through self play. Alpha Zero became the better Go player, demonstrating that trying to insert current human understanding of such a game is actually detrimental to its performance. Or at least that’s the argument that I keep hearing.
Go is a combinatorial board game, meaning that it can be fully modelled mathematically. Anyone who studies games can be easily persuaded that winning at Go more mathematical than social. We don’t need the imitation game to create a target for go because it already has one built into its rules. There is no ambiguity about what it means to win at go, we don’t even need a physical board to play the game. Go is not an image of its social norms, the mythology surrounding the game is itself the image; so of course the AI will do better when that image is removed, as training on the singular variable we care about will always be better than a simulacrum of that variable. This is why Alpha Zero is not, in my opinion, a good example of why computers will eventually be better at everything than humans. Yet, that doesn’t stop people from believing that winning at Go is a natural step in the computer’s evolution towards personhood6.
- I’m using the definition of game I developed in this blog post.[↩]
- Creator of the mathematics of ‘computable numbers’, code breaker instrumental in breaking Nazi encryption during World War 2, and one of the founding father’s of modern computing[↩]
- Turing, Alan. “Computing Machinery and Intelligence”. In The New Media Reader (pp 50-64).[↩]
- I discuss this in an earlier blog post here.[↩]
- Google has been doing this with captchas for years. To decide if someone is human, they show them a bunch of pictures and ask them if they contain traffic signs. Some of these pictures they already know the answer to, others they don’t. If you answered similarly to other humans on the known images you will be declared a human, the rest of the pictures are your unpaid contribution to their dataset.[↩]
- I’ve been working on this blog post for a really long time. I started before ChatGPT blew up the internet. I will admit it is a different beast than Alpha Zero, but I’m still not convinced it is a good argument towards the inevitability of superior AI. The argument below is not about ChatGPT, but should be easy to reorganize into why I believe this.[↩]
