It’s something we tend to grow up always assuming is real. This reality, this universe that we see and hear around us, is always with us, ever present. But sometimes there are doubts.
There’s a thing in philosophy called the Simulation Argument. It posits that, given that our descendants will likely develop the technology to simulate reality someday, the odds are quite high that our apparent world is one of these simulations, rather than the original world. It’s a probabilistic argument, based on estimated odds of there being many such simulations.
A long time ago, I had an interesting experience. Back then, as a Christian, I wrestled with my faith and was at times mad at God for the apparent evil in this world. At one point, in a moment of anger, I took a pocket knife and made a gash in a world map on the wall of my bedroom. I then went on a camping trip, and overheard in the news that Russia had invaded Georgia. Upon returning, I found that the gash went straight through the border between Russia and Georgia. I’d made that gash exactly six days before the invasion.
Then there’s the memory I have of a “glitch in the Matrix”, so to speak. Many years ago, I was in a bad place mentally and emotionally, and I tried to open a second floor window to get out of a house that probably would have ended badly, were it not for a momentary change that caused the window, which had a crank to open, to suddenly become a solid frame with no crank or way to open. It happened for a split second. Just long enough for me to panic and throw my body against the frame, making such a racket as to attract the attention of someone who could stop me and calm me down.
I still remember this incident. At the time I thought it was some intervention by God or time travellers/aliens/simulators or some other benevolent higher power. Obviously I have nothing except my memory of this. There’s no real reason for you to believe my testimony. But it’s one reason among many why I believe the world is not as it seems.
Consider for a moment the case of the total solar eclipse. It’s a convenient thing to have occur, because it allowed Einstein to prove his Theory of Relativity in 1919 by looking at the gravitational lensing effect of the sun that is only visible during an eclipse. But total solar eclipses don’t have to be. They only happen because the sun is approximately 400 times the size and 400 times the distance from the Earth as the moon is. They are exactly the right ratio of size and distance for total solar eclipses to occur. Furthermore, due to gradual changes in orbit, this coincidence is only present for a cosmologically short time frame of a few hundred million years that happens to coincide with the development of human civilization.
Note that this coincidence is immune to the Anthropic Principle because it is not essential to human existence. It is merely a useful coincidence.
Another fun coincidence is the names of the arctic and antarctic. The arctic is named after the bear constellations of Ursa Major and Minor, which can be seen only from the northern hemisphere. Antarctic literally means opposite of arctic. Coincidentally, polar bears can be found in the arctic, but no species of bear is found in the antarctic.
There are probably many more interesting coincidences like this, little Easter eggs that have been left for us to notice.
The true nature of our reality is probably something beyond our comprehension. There are hints at it however, that make me wonder about the implications. So, I advise you to keep an open mind about the possible.
Note: The following is a blog post I wrote as part of a paid written work trial with Epoch. For probably obvious reasons, I didn’t end up getting the job, but they said it was okay to publish this.
Historically, one of the major reasons machine learning was able to take off in the past decade was the utilization of Graphical Processing Units (GPUs) to accelerate the process of training and inference dramatically. In particular, Nvidia GPUs have been at the forefront of this trend, as most deep learning libraries such as Tensorflow and PyTorch initially relied quite heavily on implementations that made use of the CUDA framework. The strength of the CUDA ecosystem remains strong, such that Nvidia commands an 80% market share of data center GPUs according to a report by Omdia (https://omdia.tech.informa.com/pr/2021-aug/nvidia-maintains-dominant-position-in-2020-market-for-ai-processors-for-cloud-and-data-center).
Given the importance of hardware acceleration in the timely training and inference of machine learning models, it might be naively seem useful to look at the raw computing power of these devices in terms of FLOPS. However, due to the massively parallel nature of modern deep learning algorithms, it should be noted that it is relatively trivial to scale up model processing by simply adding additional devices, taking advantage of both data and model parallelism. Thus, raw computing power isn’t really a proper limit to consider.
What’s more appropriate is to instead look at the energy efficiency of these devices in terms of performance per watt. In the long run, energy constraints have the potential to be a bottleneck, as power generation requires substantial capital investment. Notably, data centers currently use up about 2% of the U.S. power generation capacity (https://www.energy.gov/eere/buildings/data-centers-and-servers).
For the purposes of simplifying data collection and as a nod to the dominance of Nvidia, let’s look at the energy efficiency trends in Nvidia Tesla GPUs over the past decade. Tesla GPUs are chosen because Nvidia has a policy of not selling their other consumer grade GPUs for data center use.
The data for the following was collected from Wikipedia’s page on Nvidia GPUs (https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units), which summarizes information that is publicly available from Nvidia’s product datasheets on their website. A floating point precision of 32-bits (single precision) is used for determining which FLOPS figures to use.
A more thorough analysis would probably also look at Google TPUs and AMDs lineup of GPUs, as well as Nvidia’s consumer grade GPUs. The analysis provided here can be seen more as a snapshot of the typical GPU most commonly used in today’s data centers.
Figure 1: The performance per watt of Nvidia Tesla GPUs from 2011 to 2022, in GigaFLOPS per Watt.
Notably the trend is positive. While wattages of individual cards have increased slightly over time, the performance has increased faster. Interestingly, the efficiency of these cards exceeds the efficiency of the most energy efficient supercomputers as seen in the Green500 for the same year (https://www.top500.org/lists/green500/).
An important consideration in all this is that energy efficiency is believed to have a possible hard physical limit, known as the Laudauer Limit (https://en.wikipedia.org/wiki/Landauer%27s_principle), which is dependent on the nature of entropy and information processing. Although, efforts have been made to develop reversible computation that could, in theory, get around this limit, it is not clear that such technology will ever actually be practical as all proposed forms seem to trade off this energy savings with substantial costs in space and time complexity (https://arxiv.org/abs/1708.08480).
Space complexity costs additional memory storage and time complexity requires additional operations to perform the same effective calculation. Both in practice translate into energy costs, whether it be the matter required to store the additional data, or the opportunity cost in terms of wasted operations.
More generally, it can be argued that useful information processing is efficient because it compresses information, extracting signal from noise, and filtering away irrelevant data. Neural networks for instance, rely on neural units that take in many inputs and generate a single output value that is propagated forward. This efficient aggregation of information is what makes neural networks powerful. Reversible computation in some sense reverses this efficiency, making its practicality, questionable.
Thus, it is perhaps useful to know how close we are to approaching the Laudauer Limit with our existing technology, and when to expect to reach it. The Laudauer Limit works out to 87 TeraFLOPS per watt assuming 32-bit floating point precision at room temperature.
Previous research to that end has proposed Koomey’s Law (https://en.wikipedia.org/wiki/Koomey%27s_law), which began as an expected doubling of energy efficiency every 1.57 years, but has since been revised down to once every 2.6 years. Figure 1 suggests that for Nvidia Tesla GPUs, it’s even slower.
Another interesting reason why energy efficiency may be relevant has to do with the real world benchmark of the human brain, which is believed to have evolved with energy efficiency as a critical constraint. Although the human brain is obviously not designed for general computation, we are able to roughly estimate the number of computations that the brain performs, and its related energy efficiency. Although the error bars on this calculation are significant, the human brain is estimated to perform at about 1 PetaFLOPS while using only 20 watts (https://www.openphilanthropy.org/research/new-report-on-how-much-computational-power-it-takes-to-match-the-human-brain/). This works out to approximately 50 TeraFLOPS per watt. This makes the human brain less powerful strictly speaking than our most powerful supercomputers, but more energy efficient than them by a significant margin.
Note that this is actually within an order of magnitude of the Laudauer Limit. Note also that the human brain is also roughly two and a half orders of magnitude more efficient than the most efficient Nvidia Tesla GPUs as of 2022.
On a grander scope, the question of energy efficiency is also relevant to the question of the ideal long term future. There is a scenario in Utilitarian moral philosophy known as the Utilitronium Shockwave, where the universe is hypothetically converted into the most dense possible computational matter and happiness emulations are run on this hardware to maximize happiness theoretically. This scenario is occasionally conjured up as a challenge against Utilitarian moral philosophy, but it would look very different if the most computationally efficient form of matter already existed in the form of the human brain. In such a case, the ideal future would correspond with an extraordinarily vast number of humans living excellent lives. Thus, if the human brain is in effect at the Laudauer Limit in terms of energy efficiency, and the Laudauer Limit holds against efforts towards reversible computing, we can argue in favour of this desirable human filled future.
In reality, due to entropy, it is energy that ultimately constrains the number of sentient entities that can populate the universe, rather than space, which is much more vast and largely empty. So, energy efficiency would logically be much more critical than density of matter.
This also has implications for population ethics. Assuming that entropy cannot be reversed, and the cost of living and existing requires converting some amount of usable energy into entropy, then there is a hard limit on the number of human beings that can be born into the universe. Thus, more people born at this particular moment in time implies an equivalent reduction of possible people in the future. This creates a tradeoff. People born in the present have potentially vast value in terms of influencing the future, but they will likely live worse lives than those who are born into that probably better future.
Interesting philosophical implications aside, the shrinking gap between GPU efficiency and the human brain sets a potential timeline. Once this gap in efficiency is bridged, it theoretically makes computers as energy efficient as human brains, and it should be possible at that point to emulate a human mind on hardware such that you could essentially have a synthetic human that is as economical as a biological human. This is comparable to the Ems that the economist Robin Hanson describes in his book, The Age of EM. The possibility of duplicating copies of human minds comes with its own economic and social considerations.
So, how long away is this point? Given the trend observed with GPU efficiency growth, it looks like a doubling occurs about every three years. Thus, one can expect an order of magnitude improvement in about thirty years, and two and a half orders of magnitude in seventy-five years. As mentioned, two and a half orders of magnitude is the current distance from existing GPUs and the human brain. Thus, we can roughly anticipate this to be around 2100. We can also expect to reach the Laudauer Limit shortly thereafter.
Most AI safety timelines are much sooner than this however, so it is likely that we will have to deal with aligning AGI before the potential boost that could come from having synthetic human minds or the potential barrier of the Laudauer Limit slowing down AI capabilities development.
In terms of future research considerations, a logical next step would be to look at how quickly the overall power consumption of data centers is increasing and also the current growth rates of electricity production to see to what extent they are sustainable and whether improvements to energy efficiency will be outpaced by demand. If so, that could act to slow the pace of machine learning research that relies on very large models trained on massive amounts of compute. This is in addition to other potential limits, such as the rate of data generation for large language models, which depend on massive datasets of essentially the entire Internet at this point.
The nature of current modern computation is that it is not free. It requires available energy to be expended and converted to entropy. Barring radical new innovations like practical reversible computers, this has the potential to be a long-term limiting factor in the advancement of machine learning technologies that rely heavily on parallel processing accelerators like GPUs.
One thing I’ve learned from observing people and society is the awareness that the vast majority of folks are egoistic, or selfish. They tend to care about their own happiness and are at best indifferent to the happiness of others unless they have some kind of relationship with that person, in which case they care about that person’s happiness in so far as it has an effect on their own happiness to keep that person happy. This is the natural, neutral state of affairs. It is unnatural to care about other people’s happiness for the sake of themselves as ends. We call such unnatural behaviour “altruism”, and tend to glorify it in narratives but avoid actually being that way in reality.
In an ideal world, all people would be altruistic. They would equally value their own happiness and the happiness of each other person because we are all persons deserving happiness. Instead, reality is mostly a world of selfishness. To me, the root of all evil is this egoism, this lack of concern for the well-being of others that is the norm in our society.
I say this knowing that I am a hypocrite. I say this as someone who tries to be altruistic at times, but is very inconsistent with the application of the principles that it logically entails. If I were a saint, I would have sold everything I didn’t need and donated at least half my gross income to charities that help the global poor. I would be vegan. I would probably not live in a nice house and own a car (a hybrid at least) and be busy living a pleasant life with my family.
Instead, I donate a small fraction of my gross income to charity and call it a day. I occasionally make the effort to help my friends and family when they are in obvious need. I still eat meat and play computer games and own a grand piano that I don’t need.
The reality is that altruism is hard. Doing the right thing for the right reasons requires sacrificing our selfish desires. Most people don’t even begin to bother. In their world view, acts of kindness and altruism are seen with suspicion, as having ulterior motives of virtue signalling or guilt tripping or something else. In such a world, we are not rewarded for doing good, but punished. The incentives favour egoism. That’s why the world runs on capitalism after all.
And so, the world is the way it is. People largely don’t do the right thing, and don’t even realize there is a right thing to do. Most of them don’t care. There are seven billion people in this world right now, and most likely, only a tiny handful of people care that you or I even exist, much less act consistently towards our well-being and happiness.
So, why am I bothering to explain this to you? Because I think we can do better. Not be perfect, but better. We can do more to try to care about others and make the effort to make the world a better place. I believe I do this with my modest donations to charity, and my acts of kindness towards friends and strangers alike. These are small victories for goodness and justice and should be celebrated, even if in the end we fall short of being saints.
In the end, the direction you go in is more important than the magnitude of the step you take. Many small steps in the right direction will get you to where you want to be eventually. Conversely, if your direction is wrong, then bigger steps aren’t always better.
When I was a child, I wanted, at various times, to be a jet fighter pilot, the next Sherlock Holmes (unaware he was fictional), or a great scientist like Albert Einstein. As I grew older, I found myself drawn to creative hobbies, like writing stories (or at least coming up with ideas for them) and making computer games in my spare time. In grade 8 I won an English award, mostly because I’d shown such fervour in reading my teacher’s copy of The Lord Of The Rings, and written some interesting things while inspired to be like J.R.R. Tolkien, or Isaac Asimov.
In high school my highest grades were reserved for computer science initially, where I managed to turn a hobby of making silly computer games into a top final project a couple years in a row. Even though, at the end of high school, I won another award, this time the Social Science Book award, after doing quite well in a modern history class, I decided to go into computer science in undergrad.
For various reasons, I got depressed at the end of high school, and the depression dragged through the beginning of undergrad where I was no longer a top student. I struggled with the freedom I had, and I wasn’t particularly focused or diligent. Programming became work to me, and my game design hobby fell by the wayside. Writing essays for school made me lose interest in my novel ideas as well.
At some point, one of the few morning lectures I was able to drag myself to was presented by a professor who mentioned he wanted a research assistant for a project. Later that summer, I somehow convinced him to take me on and spent time in a lab trying to get projectors to work with mirrors and fresnel lenses to make a kind of flight simulator for birds. It didn’t go far, but it gave me a taste for this research thing.
I spent the rest of my undergrad trying to shore up my GPA so I could get into a masters program and attempt to learn to be a scientist. In a way, I’d gone full circle to an early dream I had as a child. I’d also become increasingly interested in neural networks as a path towards AI, having switched from software design to cognitive science as my computing specialization early on.
The masters was also a struggle. Around this time emotional and mental health issues made me ineffective at times, and although I did find an understanding professor to be my thesis supervisor, I was often distracted from my work.
Eventually though, I finished my thesis. I technically also published two papers with it, although I don’t consider these my best work. While in the big city, I was also able to attend a machine learning course at a more prestigeous university, and got swept up in the deep learning wave that was happening around then.
Around then I devoted myself to some ambitious projects, like the original iteration of the Earthquake Predictor and Music-RNN. Riding the wave, I joined a startup as a data scientist, briefly, and then a big tech company as a research scientist. I poured my heart and soul into some ideas that I thought had potential, unaware that most of them were either flukes of experimental randomness, or doomed to be swept away by the continuing tide of new innovations that would quickly replace them.
Still, I found myself struggling to keep working on the ideas I thought were meaningful, and became disillusioned when it became apparent that they wouldn’t see support and I was sidelined into a lesser role than before, with little freedom to pursue my research.
In some sense, I left because I wanted to prove my ideas on my own. And then I tried to do so, and realized that I didn’t have the resources or the competency. Further experiments were inconclusive. The thing I thought was my most important work, this activation function that I thought could replace the default, turned out to be less clearly optimal than I’d theorized. My most recent experiments suggest it still is something that is better calibrated and leads to less overconfident models, but I don’t think I have the capabilities to turn this into a paper and publish it anywhere worthwhile. And I’m not sure if I’m just still holding onto a silly hope that all the experiments and effort that went into this project weren’t a grand waste of time.
I’d hoped that I could find my way into a position somewhere that would appreciate the ideas that I’d developed, perhaps help me to finally publish them. I interviewed with places of some repute. But eventually I started to wonder if what I was doing even made sense.
This dream of AI research. It depended on the assumption that this technology would benefit humanity in the grandest way possible. It depended on the belief that by being somewhere in the machinary of research and engineering, I’d be able to help steer things in the right direction.
But then I read in a book about how AI capability was dramatically outpacing AI safety. I was vaguely aware of this fact before. The fact is that these corporations and governments want AI for the power that it offers them, and questions about friendliness and superintelligence seem silly and absurd when looking at the average model that simply perceives and reports probabilities that such and such a thing is such.
And I watched as the surveillance economy grew on the backs of these models. I realized that the people in charge weren’t necessarily considering the moral implications of things. I realized that by pursuing my dream, I was allowing myself to be a part of a machine that was starting to more closely resemble a kind of dystopian nightmare.
So I made a decision. That this dream didn’t serve the greatest good. That my dream was selfish. I took the opportunity that presented itself to go back to another dream from another life, the old one about designing games and telling stories. Because at least, I could see no way for those dreams to turn out wrong.
In theory, I could work on AI safety directly. In theory I could try a different version of the dream. But in practice, I don’t know where to begin that line of research. And I don’t want to be responsible for a mistake that ends the world.
So, for now at least, I’m choosing a dream I can live with. Something less grand, but also less dangerous. I don’t know if this is the right way to go. But it’s the path that seems open to me now. What else happens, I cannot predict. But I can try to take the path that seems best. Because my true dream is to do something that brings about the best world, with the most authentic happiness. How I go about it, that can change with the winds.
I’m currently still trying to decide what I should even post here. I tend to post more personal stuff on Facebook and to a lesser extent on Twitter, but my fiancee thinks it might be unwise to publish personal details on a public facing blog like this one.
Possibly I could focus more on professionally relevant ideas, but I’m not sure what I can offer in that regard. Anything really worth publishing should probably go into a proper paper rather than some random blog on the Internets. I suppose I could write opinions about philosophical things, but that overlaps with the Pax Scientia wiki that I was working on building earlier.
I probably have too many of these projects that don’t get enough attention anyway. I’ve been trying to consolidate them recently, but I worry that the resulting web presence is still far too sprawling and even less clear to navigate without the delineations.
Another debate I’ve been having recently is whether to put more effort into my creative writing. I want to eventually write a novel. It’s a vague goal I’ve had since I was a kid. I have lots of ideas for stories, but I’ve always had trouble actually getting down to writing the ideas down into actual narratives. Sometimes I wonder if I actually have the writing ability to justify the effort, whether it makes sense to add yet another piece of literature to the ever expanding pile of books in the world.
I spent a long time working out in my head the worlds that I want to write about. In some sense, if I don’t write, it’ll have been a waste. But I’m not sure my imagination is that much more extraordinary enough to justify the effort in the first place.
I also claim to be a music composer and a game designer, the other two arts that I have some capacity in. To what extent would those be more appropriate uses of my time? To what extent is writing more worthwhile than composing songs for instance? I can hash out a song somewhat faster than a novel, but I also as yet don’t consider my songs to be particularly notable either.
My thoughts on why writing was my first choice in terms of artistic expression were originally and ostensibly because writing allows me to communicate ideas rather than just emotions like with music. And writing can be done on my own, rather than needing an artist and a team for game development. Admittedly, the creator of Stardew Valley did it on his own, but I don’t have the visual art skills for that, and I don’t see myself having the patience to become good at drawing at this point.
In another debate, I’ve also been considering a change of career path. Working in machine learning has been exciting and lucrative, but the market now seems increasingly saturated as the most competent folks in the world recognized the hype and adjusted their trajectories to compete with mine. Whereas a few years ago I was one of maybe a couple hundred, now there seem like thousands of people with PhDs who outclass me.
At the same time, I’ve wondered about whether or not the A.I. Alignment problem, the existential risk of which has been the focus of several books by prominent philosophers and computer scientists, isn’t a more important problem that needs more people working on it. So I’ve wondered if I should try switching into this field.
Admittedly, this field seems to be still in its infancy. There’s a bunch of papers looking at defining terms and building theoretical frameworks, and little in the way of even basic toy problems that can be coded and tested. I’m personally more of an experimentalist than a theoretician when it comes to AI and ML, mostly because my mathematical acumen is somewhat lackluster, so I’m not sure how much I can help push forward the field anyway.
On a more philosophical note, it seems the social media filter bubble has been pushing me more to the left politically. At least, I find myself debating online with Marxists about things and becoming more sympathetic to socialism, even though a couple years ago I was a moderate liberal. I’m not sure how much to blame the polarization of social media, and how much it’s the reality of disillusionment with the existing world.
I also have mixed feelings in part because the last company I worked for was, according to media outlets, controversial, but to me it was the company that gave me a chance to work on some really cool things and paid me handsomely for my time and energy. Admittedly, as a lowly scientist working in an R&D lab, I wouldn’t have been privy to anything untoward that could have been happening, but it was always jarring to see the news articles that attacked the company.
I left more for personal reasons, partly some issues of office politics that I wasn’t particularly good at dealing with. My own criticisms of the company culture would be much more nuanced, aware that any major corporation has its internal issues, and that many of them are general concerns of large tech companies.
The debates in my head are somewhat bothersome to be honest. But at the same time, it means I’m thinking about things, and open to updating my understanding of the truth according to new evidence, factored with my prior knowledge.
