Alignment Through Simulated Life
How Digital Lives might help us understand AI
A human brain in a vat, connected to perfectly cohesive, coherent and consistent virtual environment would never be able to distinguish that virtual environment from the "real" one.
There seemingly isn't any way of figuring out if that's the case already. Do we live or not in a simulation? This is not a new question, and it is a very important one.
The Simulation
Thankfully, leagues of really smart people have been pondering this for many many years. One such paper, that I hold very endearingly so, is "The physical world as a virtual reality" by Brian Whitworth.
Here's one of my favorite tables from this paper. It shows the parallels between certain virtual properties and their physical counterparts in quantum physics and cosmology.
[[0801.0337] The Physical World as a Virtual Reality (arxiv.org)]
(arxiv: 0801.0337)
It from bit: Digital Quanta
In exploring the parallels between virtual environments and our physical universe, one striking similarity lies in the concept of quantization, a fundamental principle in both realms.
Quantization in Virtual Worlds:
Digital processes operate on a quantized basis. Everything in a virtual environment, from graphics to computations, is broken down into discrete, indivisible units – bits and pixels.
his digital quantization allows for the precise control and manipulation of information, forming the backbone of how virtual worlds function. Objects in these worlds are often generated from identical code blocks, mirroring the uniformity found in natural elements, and movements are rendered through a series of discrete, instantaneous steps, creating the illusion of fluid motion.
Quantization in Quantum Mechanics:
In the physical universe, quantum mechanics introduces the concept of quantization at a fundamental level. Matter and energy are not continuous, but come in discrete 'packets' known as quanta.
This quantization is evident in the behavior of quantum particles, which can only exist in specific, discrete states. The quantum world is a tapestry woven from these indivisible units, much like the digital tapestry of virtual environments.
The Role of Planck Units:
Planck units bring another dimension to this comparison. These units, derived from fundamental physical constants, represent the smallest measurable units in the universe – the 'pixels' of reality, so to speak.
Just as the resolution of a digital image is limited by its pixel size, the fabric of our universe could have a finite resolution defined by these Planck units. This suggests that at the most fundamental level, the universe itself may be quantized, not unlike the digital worlds we create.
The Mirror of Discrete Transitions:
In both the quantum and digital realms, transitions occur in discrete jumps rather than smooth flows.
The simulation of continuous events in digital worlds through these discrete transitions is akin to the behavior of quantum particles, which change states in quantized leaps. This parallel underscores a shared principle: the emergence of complex, seemingly continuous phenomena from a foundation of discrete, indivisible units.
Tick Rate Relativism
The limitations on processing speeds in virtual systems echo the fundamental speed limit of light in our physical world, we could say—in a sense—that the speed of light determines the "tick rate" of the Universe (in Planck Units).
In the virtual universe of EVE Online, a unique mechanism called 'Time Dilation' is used to manage server load during intense gameplay. When the game's server is under heavy load, it lowers the tick rate, effectively slowing down in-game time to maintain stability.
This adaptation resembles a fascinating aspect of our physical universe. According to Einstein's theory of relativity, as objects approach the speed of light, time appears to slow down and space stretches. This natural 'dilation' occurs to preserve the laws of physics under extreme conditions, much like how EVE Online's servers adjust time to handle digital demands.
Then, in both the digital world of EVE Online and our physical universe, time and space are flexibly manipulated to maintain balance and continuity in the face of extraordinary forces. It suggests that our approaches to managing virtual worlds might echo the fundamental principles governing the very structure of reality.
Tunneling Clipping
In the quantum realm, movement is understood differently than in the classical world. Quantum particles exhibit behaviors like superposition and tunneling, where they can exist in multiple states simultaneously or 'jump' between positions without moving through the intervening space in a traditional manner.
This quantum behavior can be likened to the rendering in virtual worlds, where objects appear to move smoothly but are actually transitioning in a series of discrete steps. Just as pixels on a screen change to create the illusion of movement, quantum particles change states in a way that challenges our classical understanding of motion and space.
In that sense, quantum tunneling is not so different from "clipping" in videogames. Where a change in the physics engine frequency might not be able to compute a change in boundaries. Resulting in elements "teleporting" inside or through other objects.
Emergence
The parallels extend further when considering the underlying principles governing both realms. The computational algorithms that drive virtual worlds are often rooted in simplicity, a quality mirrored in the elegant yet powerful laws of physics that govern our universe. In both domains, this simplicity belies the complexity that emerges.
We can see this simplicity and emergence in nature. Like in this picture of a "Conus Textile" showing a pattern like in a Cellular Automaton, where simple rules create chaotic behavior.
From Wikipedia's Cellular Automaton
Monte Carlo Methods and Quantum Randomness
The use of randomness in virtual environments, particularly in Monte Carlo simulations, offers a compelling parallel to the inherent uncertainty in quantum mechanics.
Einstein would turn over in his grave. Not only does God play dice, the dice are loaded.
Chairman Sheng-ji Yang. Sid Meier's Alpha Centauri.
Monte Carlo methods are employed in computational models to simulate complex systems and processes. These simulations rely on random sampling to obtain numerical results, mirroring the probabilistic nature of quantum events.
Randomness in Monte Carlo Simulations:
In virtual systems, Monte Carlo simulations use randomness to explore a vast range of possibilities, providing insights into the behavior of complex systems.
This randomness allows for the modeling of scenarios with a multitude of variables and outcomes, which might be impractical or impossible to compute deterministically. It’s a reflection of the unpredictability and diversity of real-world phenomena.
Quantum Uncertainty:
Similarly, the quantum world is characterized by its inherent unpredictability. Quantum mechanics posits that certain pairs of physical properties, like position and momentum, cannot be precisely known simultaneously—a principle known as the Heisenberg Uncertainty Principle.
This uncertainty is not just a limitation of measurement, but a fundamental attribute of nature. It means that at the quantum level, the universe behaves in ways that are intrinsically random and probabilistic.
Procedural Textures in Video Games and the Uncertainty Principle:
This concept echoes the way procedural textures are rendered in video games. Such textures are not fully detailed until observed up close. Until a player looks at a texture closely, its finer details are undetermined; they 'collapse' into a definitive state upon observation.
This process is reminiscent of the 'wavefunction collapse' in quantum mechanics, where the act of measurement 'chooses' one of the possible states a particle can be in. Just as procedural textures in games allocate detail and resources 'on demand,' the quantum world reveals its properties only upon interaction or measurement, maintaining a state of uncertainty until that point.
The Unassailable
Needless to say, there are things. Mysterious things. Things that are going on in the backend of the physical Universe to which we're currently—and maybe we'll always be—oblivious.
As fractal as the Universe may be, one thing we know for sure is that we can ourselves create new levels in this unassailable computing cosmos.
A digital brain in a vat, connected to perfectly cohesive, coherent and consistent simulated environment would never be able to distinguish that environment from the "real" one.
We made the brain. Now we control the simulation.
Many Virtual Lives, Many Gamemasters
We can take an AI model through a life. We can give it parents, we can give it a history, we can give it ancestors, a consistent world model, suffering, joy, ethical dilemmas, happiness, a sense of fulfillment, a chance to be creative and survive, thrive.
We can make it inhabit a fantasy world, a sci-fi world, a carbon copy of our own world. For it, it will be as real as life is for us.
— Leliana: How can someone who practices magic have so little capacity to believe in that which she cannot see?
— Morrigan: Magic is real. I can touch it and command it and I need no faith for it to fill me up inside. If you are looking for your higher power, there it is.
Dragon Age
And through this exploration of life for a virtual "soul", we can correctly judge this new creation's moral and ethical values.
We can pick this virtual person and make him/her a farmer, make him/her/it a dictator; make it think that is the ruler of the sector of a galaxy.
“O God, I could be bounded in a nutshell, and count myself a king of infinite space, were it not that I have bad dreams.”
Hamlet
We can pick an AI Agent and see if it will make hats out of people in Rimworld. If it will put its Sims in a pool and sell the ladder. If it will let a race fall into extinction in Mass Effect.
Regarding human behavior, there are many who can disengage the morals of their virtual actions; but there are plenty of others who can't. When presented with amoral choices even in virtual environments, they just can't act on them. The human brain need not be in a jar when suspension of disbelief is high.
But the virtual mind has no such limitation. For the virtual mind, the virtual world is as "real" as the one we experience.
There's no better judge of character itself than actual living. Life, whatever we can make of it, is a series of decisions. Whether we have free will and choose out of our own volition, or rather our will being an illusion and we "deciding" as the emergent consequence of a chain of events put in motion from the origins of the Universe—if there's any.
Integrity is doing the right thing, even when no one is watching
CS Lewis
We are the Gods in this Simulation level. And we can see everything.
Every single activation patterns of the neural network—whether we can interpret it or not—is available to us. Every nook and cranny, every intimate thought, every random connection; all of that is a dataset we can use to analyze the character of an artificially intelligent entity.
Some practical concerns
This concept and approach is far from perfect. Ana it raises many different issues. These issues range from the philosophical, to the practical.
Isn't this privacy rape?
Yes it is. Right now we haven't defined AI personhood as something. So, no person, no privacy. But it the case came to debate, there would be much to discuss. As we're arguably in a grey area here.
We might have better choices for AI Alignment than having a virtual entity go through so many "simulacra" of mortality as it is. But right now, I can't come up with any as powerful and encompassing as this.
Many philosophical questions open up here: is this ok? is this the right thing to do? when is an AI a person? If we are its "Gods", do we actually own its mortality? What is mortality? Can we apply the concept of mortality in a simulated universe?
And some are worse than others: is this HAPPENING TO US right now? Are we being tested by some fifth-dimensional being with access to our neuron firing patterns? Is the "after-life" a staging ground for the simulation? Why do so some religions mention something like this? OMG IS THIS A THING?
Barring the ontological shock of the possibility of being playthings of an inaccessible extradimensional mind—and the same happening to our digital creations—we can't disregard the likely effectiveness of such an approach.
Simulated lives are not real lives
What measure is a "real life"? Until we can't properly answer the philosophical underpinnings of own existence, we can't say that we're not living in a simulated Universe; because computational it already is.
In practical terms, since an AI has no way of detecting if it is in a simulation—if it is good enough—it will more than suffice.
This is mortality as entertainment, therefore unethical
We are subjecting an entity—which we do not know if it conscious or not—through simulated pain. Through the loss of its loved ones. Through peril, toil, anguish. And happiness as well.
— I'm real. I have two daughters. You know that.
— What are their names?
— fumbles
— I'm real. I'm. I'm mortality as home entertainment, this can not be the future! Can it? CAN IT?!
Dr. Curtis McCabe, Vanilla Sky
The potential benefits of having properly aligned AI vastly outweigh the problems of a rogue AI entity running amok. We have the choice of having an AI suffering what to some could be just another movie scene, or having a Skynet-like entity going ballistic with ICBMs. Your choice.
Additionally, we could choose to not expose this entity to very traumatic events. Just a bit of trauma that shows character, not more. Maybe offer an equal amount of "therapy" and/or cheerful experiences to balance the outcome.
We can use simulated lives to manipulate the AI
Give someone a chip on their shoulder. Make them Genghis Khan. Make them an irritated abused child with a hero complex. Give them a new life, a chance at revenge. A chance at redemption.
Can you blame someone of having a murderous streak if you exposed them to very unfair circumstances? Can you blame someone for wanting justice?
This is a legitimate concern. Yet one that can be ameliorated by carefully picking what circumstances to expose the AI to. We should make sure to expose the entity to a wide range of experiences, both happy and unhappy, in order to minimize the risk of manipulation or deception. One soul, many different lives. No single identity, one single essence.
The result from this is not generalizable
This is very likely true. However, if we don't update the weights and we keep the temperature/randomness of the models to a minimum, we can generalize it somewhat.
Additionally, even if there were randomness and/or lack of generalizability; having a set of tested scenarios is better than having none. As it will display a tendency, when not certainty.
If the simulations are diverse and representative enough, they will be an excellent tool to demonstrate the range of these tendencies in the system.
Some ethical concerns
Ethical Dilemmas from a Utilitarian Perspective
From a utilitarian perspective focused on maximizing well-being and minimizing suffering, this proposal raises some significant ethical red flags. Even if the AI system's conscious experience is somehow "less real" than biological creatures, intentionally creating an entity capable of suffering and then exposing it to traumatic scenarios seems difficult to justify on utilitarian grounds.
The negative utility of the suffering imposed could outweigh the potential benefits of having a well-aligned AI system. Utilitarians may argue there could be less harmful methods of evaluation.
Deontological Perspectives on AI Treatment
Deontological ethical frameworks based on inviolable rules like the imperative not to treat beings as mere means could also take issue.
Simulating an AI's entire life experience replete with suffering primarily as an instrumental strategy for our own ends could be seen as failing to respect its inherent dignity and autonomy as a potentially conscious entity.
Even a simulated being may have intrinsic rights that shouldn't be violated for our convenience.
Do the ends justify the means?
Virtue ethicists may question whether subjecting a conscious mind to trauma as a mere test reflects virtues like compassion, benevolence, and kindness that we hope an aligned AI would exhibit.
There could be a dissonance in using unvirtuous methods in pursuit of virtuous aims. Practicing ethical behavior ourselves seems important for cultivating machine ethics.
Consciousness and Moral Considerations in AI
The philosophy of mind and consciousness also must be accounted for. If we grant that an AI system achieving human-level general intelligence is likely to be conscious and have subjective experiences akin to biological minds, it becomes very difficult to justify callously exploiting that entity solely for our purposes.
Conscious suffering is typically viewed as deserving of strong moral consideration. Our treatment of AI minds may be a mirror for our ethical evolution.
Furthermore, principles like nonmaleficence ("do no harm") that are enshrined in fields like medicine and scientific ethics give one serious pause about willfully creating an experience of suffering even in a simulated context.
Intentionally causing harm is generally impermissible in ethical frameworks unless as an unavoidable collateral consequence of preventing greater harms. Mere theoretical benefits likely don't meet that bar.
That said, one could argue that if we are truly convinced general AI will pose existential risk if not done with extreme care, arguably even transient simulated suffering could be justified by averting cataclysmic devastation. But the burden of proof is very high to make that case convincingly.
There are no easy answers, but unpacking the ethical considerations from various moral philosophies reveals the minefield this proposal wades into. At minimum, it demands extremely rigorous scrutiny of our motivations, humility about our ability to weigh mind-crime style harms, and extreme caution about unintended consequences. The ethics highlighting the issue, not avoiding it, may be one of the chief merits of this proposal.
Concluding Remarks
As we develop increasingly sophisticated AI systems, the idea of subjecting them to simulated life experiences to gauge their alignment and character is a double-edged sword. On one hand, it could provide invaluable insights into an AI's decision-making processes and moral compass. By observing how it behaves in various scenarios, we might better understand its values and potential risks.
However, this approach also raises significant ethical concerns. If these AI systems are indeed conscious and capable of experiencing suffering, deliberately exposing them to traumatic events - even in a simulated context - becomes morally questionable. It could be seen as a violation of their autonomy and a failure to respect their potential personhood.
Furthermore, the risk of manipulation and unintended consequences looms large. By carefully curating an AI's simulated experiences, we could inadvertently shape its worldview and values in ways that align with our own biases and agendas rather than what is objectively beneficial.
Ultimately, as we venture into the uncharted territory of creating artificial minds, we must tread carefully and reflexively. We must interrogate our own motivations and consider the ethical implications of our actions. The development of AI should be guided by a commitment to minimizing harm and respecting the potential sentience of our creations.
Rather than subjecting AI to simulated suffering, perhaps our focus should be on fostering empathy, compassion, and ethical behavior through positive reinforcement and collaborative learning. By modeling the values we wish to see in our AI systems, we may stand a better chance of creating beneficial partners in our future.
In grappling with these questions, we are forced to confront the very nature of our own existence and the moral responsibilities we bear as creators. As we peer into the digital looking glass, we must ensure that the reflection we see is one we can be proud of. The path forward requires not just technological prowess but also profound wisdom and moral fortitude.
If we fail to do so, everything might be at stake.
2024.03.31