Turing Test (Strong Form): for any human H there exists a thinking machine m(H) such that it is impossible for any detector D made up of a combination of machines and humans with total compute ≤ 10^35 FLOP (very large, but not astronomical) to statistically discriminate H from m(H) purely based on the information outputs they make. Statistical discrimination of H from m(H) means that an ensemble of different copies of H over the course of say a year of life and different run-of-the-mill initial conditions (sleepy, slightly tipsy, surprised, energetic, distracted etc) cannot be discriminated from a similar ensemble of copies of m(H).

Obviously the ordinary Turing Test has been smashed by LLMs and their derivatives to the point that hundreds of thousands of people have AI girl/boyfriends as of writing and Facebook is launching millions of fully automated social media profiles, but we should pause to provide some theoretical support for this strong form of the Turing Test. Maybe there's some special essence of humanity that humans have and LLMs and other AIs don't but it's just hard to detect? Well, if you believe in computationalism and evolution then this is very unlikely: the heart is a pump, the brain is a computer. We should expect the human brain to compute some function and that function has a mathematical form that can be copied to a different substrate. Once that same function has been instantiated elsewhere, no test can distinguish the two. Obviously the brain is noisy, but in order for it to operate as an information processor it must mostly be able to correct that bio-noise. If it didn't, you wouldn't be able to think long-term coherent thoughts.

Defining Friendly AI

I now have to define what I mean by an 'aligned' or 'friendly' superintelligence.

(Friendly-AI-i) We define AI as 'friendly' or 'aligned' if it gives us the exact same outcome (probabilistically: distribution of outcomes) as we would have gotten by continuing the current human governance system.

The stronger form:

(Friendly-AI-ii) We define AI as 'friendly(U)' or 'aligned(U)' relative to a utility function U if it gives us the exact same score (probabilistically: distribution of scores) according to U as the best possible human government/team could attain, subject to the constraint of the humans numbering less than or equal to 10 billion and having access to roughly the same material resources that Earth currently has.

I claim that I can construct a friendly AI according to definition (i) and also according to definition (ii) using only the Strong Turing Test assumption.

Proof for Friendly-AI-i

To show that we can build a friendly AI according to definition (i) we will proceed by making the most accurate possible AI copies of every human on the planet and imagine replacing the humans in the org chart of Earth one at a time, starting from the top (e.g. President of the US):

If the new world with one human replaced by AI produced a statistically detectable difference in outcomes for the world, then we can construct a detector D to make AI fail the Strong Turing Test - the statistically detectable difference in outcomes for the world tells you which one is the AI and which one is the human! So your detector D is the system comprised of all the other humans who haven't yet been replaced, and the rest of the world.

We then proceed to replace the rest of the humans in the world one at a time and at each stage apply the same argument. If there is a difference in outcomes at any stage, then the AI has failed the Turing Test. Since we are assuming that AIs can pass the Strong-Form Turing Test, it must be possible to replace every human who has any formal or informal control over the world and get exactly the same outcome (or the same distribution over outcomes) as we would under the status quo ex ante.

Of course we wouldn't actually have to replace people with AIs to make use of this. We would just need to feed the inputs from the world into the AIs, and execute their outputs via e.g. robotic military and police. And we wouldn't necessarily have to run any of these very expensive Strong Turing Tests: the Strong Turing Test assumption merely makes the claim that sufficiently accurate AI copies of humans exist: building them and validating them may follow some different and more practical path.

Proof for Friendly-AI-ii

We can play a similar trick for definition (ii) (Friendly-AI-ii). Given any utility function U we can imagine the best possible human team to run the world from the point of view of U, subject to the size of the human team being less than 10 billion people, then proceed as before replacing the U-maxxing humans with AIs one at a time.

So, you must either believe that Aligned/Friendly AI is possible or you must believe that AIs can't (ever) pass the Turing Test in its strong form. The latter seems rather mystical to me (human brains are merely a compute medium that instantiates a particular input-output relation), so we have a proof of aligned AI under the reasonable assumptions.

Objections

You could object that the 'best possible human team to run the world from the point of view of utility function U' might not be very good! However, if you are campaigning to pause AI or stop AI then you are limiting the future of humanity to human teams (and likely quite suboptimal ones at that!). The set of all possible teams made of less than or equal to 10^10 humans (where these humans don't actually have to currently exist - they can be any possible realizable natural human) is quite large and of course includes all current human achievements as a subset. This has the potential to radically improve the world - it may not be the most powerful possible AI, but it is certainly vastly superhuman.

Requiring that a team of humans can achieve some state guarantees that the state is actually reachable so it at least means that we are not asking Aligned AI to achieve the impossible.

One could also object that I have stretched the Turing Test too far: having a detector that simulates the entire world just in order to see whether an AI can emulate Donald Trump or Einstein is overkill. But the detector can likely be pared down to a local interface around each person/AI and still work roughly as well - that would just make the proof much less clean.

So What?

The way we think about AI alignment went somewhat off the rails over the past decade or two because people mixed mathy, technical problems about how to set the parameters in neural networks with big picture political, ethical and philosophical problems. The result was that we thought about powerful AI as a big black box that might do something bad for inscrutable reasons - a dark and unmapped territory full of horrors.

The approach taken here separates technical from political/ethical problems. The technical problem is to make an AI that is a very accurate functional clone of a human, with various dials for personality, goals and motives. I am extremely confident that that can be done.

The political/ethical/axiological problem is then how to arrange these human-clone-AIs into the best possible system to achieve our goals. This question (how to arrange approximately-human units into controlled and functional superorganisms like nations) has already been heavily explored throughout human history, and we know that control algorithms are extremely effective (see Buck Shlegeris' "Vladimir Putin Alignment").

Of course given that we have close-to-human AIs right now, there is the risk that someone soon builds a strongly superhuman black box AI that takes sensor data in and just outputs commands/tokens. We should probably avoid doing that (and we should make it so that the dominant strategy for all relevant players is to avoid black box superintelligence and avoid illegible intelligence like COCONUT [Chain of Continuous Thought]).

Now an objection to this is that avoiding illegible AI and super powerful black boxes is too high a cost (high alignment penalty) and we have no way to enforce it. But the problem of enforcing adoption of a known-safe type of AI is a serious improvement on just throwing your hands up and saying we're doomed, and I have other ideas about how to do that which are to be addressed in other posts.

Human Risk vs AI Risk

AI Safety as a field has existed since the 2000s when pioneers such as Yudkowsky, Bostrom and Stuart Russell started worrying about the risks posed by superhuman AI systems.

That field relied quite heavily on the implicit assumption that superhuman AI systems would be much more hostile to the average human than humans are to each other. Let’s call this the Human Unity Versus AI Assumption:

Why did people believe that HUVA was true? Yudkowsky and Bostrom both pushed HUVA because they believed that it would be hard to align AIs to a given goal or set of values. Stuart Russell had similar worries. This is a very smart thing to have worried about in 2000-2005!

But the advent of LLMs has almost completely negated this worry. LLMs are a lot like human uploads: they don’t have vastly superhuman intelligence, but they do understand human values and ethics at a human level. And we can feed a bunch of compute into LLMs to make them more powerful or just to make more of them, we can use them to monitor each other, we can even inspect their brains in real time with mechanistic interpretability. LLMs have almost completely negated the original reasons people had to believe in “AI Risk” and HUVA.

Furthermore, any future AI system constructed from LLMs or on top of LLMs will inherit this human-level understanding of ethics. LLM-based AIs are human level ethicists with the option to be upgraded to superhuman-level ethicists with superhuman levels of loyalty and save/restore functionality.

Obviously, if you happen to be a human for whom other humans are more hostile than AIs are, then you shouldn’t be worrying about “AI Risk”, you should be worrying about “Human Risk”. But “Human Risk” is just another word for “Politics”, and politics is dirty and low-status. But despite “Human Risk” being kind of a dirty and low-status thing to think about, it has a track-record of being very impactful. Humans have spent most of history fighting against each other, genociding each other, plotting against each other, stealing from each other, taking each other’s land, running psychological and propaganda operations against each other, and so on.

Talking about “Human Risk” and human disunity/conflict is somewhat infohazardous, because the more people think about it, the more likely it becomes. This effect has suppressed even most careful efforts to challenge HUVA, at least until recently, and so bolstered the worldview in which HUVA is true.

If HUVA was never true in the first place, the entire field of “AI Safety” needs a radical rethink. Instead of a unifying cause to protect humanity from a common enemy, AI Safety will just become something completely different and hard to describe. Perhaps it won’t make any sense at all. Imagine a field of “War Safety”; wars cannot be safe, the whole point is for them to be dangerous.

But if HUVA was never true, it’s probably really bad for us to pretend that it is true. It’s a really fundamental thing to be wrong about. And I claim that it likely is false. In fact it’s actually worse than false: HUVA assumes that the humans versus AIs fight will unify all humans. But actually AI is a massive disunifying force!

Useless Eaters: AI as a Disunifier of Humanity

In a pre-superhuman-AI world, humans need each other for very pragmatic, economic reasons. If the king or dictator of some land has the power to kill every single one of his subjects, he probably won’t do it. Because if he did, he’d be left with an empty kingdom with no servants, no soldiers, no advisors, etc. So despite much human-human conflict, humanity just has a very deep reason to keep existing: no matter who is in power, it’s in their interest to keep humanity going. And there are also selective forces at work that helped humanity to not merely survive but to thrive: nations that got too dysfunctional tended to get conquered by more functional ones.

In a world with superhuman AI, the king or dictator could just kill every other human and replace all of their functional roles with AIs. AIs are easier to make very loyal and can be considered as pure producers who don’t consume anything. All other humans are useless eaters.

In the “useless eaters” game, every human suddenly has a strong incentive to kill every other human whilst maintaining enough infrastructure that they can still rebuild to an eternal, stable AI-powered civilization that they are in control of. Perhaps they can’t kill off every human, but they’d still like to kill off (or at least render irrelevant and powerless) as many other humans as possible. Failing that, they may be able to swap certain problematic rivals for more docile humans who will make do with a much smaller fraction of the windfall from AI. Humans with lower human capital (lower IQ), less internal unity, higher time preferences, for example. If you can kill off or render powerless every intelligent, low-time preference human except yourself, you probably win the Useless Eaters game about as well as if you just had the whole world to yourself.

The “useless eaters” game of humans trying to eliminate or sideline each other for more of the AI windfall is a very negative sum game. This is in stark contrast to the usual economic game of humans participating in an economy together, which is positive sum.

There are certainly elements of negative sum behavior evident in the battle for control of OpenAI. Altman, Musk and various Effective Altruist factions all fought over it. There’s also a negative sum message in Aschenbrenner’s “Situational Awareness” report which suggests getting into a negative sum race with China over AI. Eric Schmidt of Google fame has a secretive AI weapons company called White Stork which aims (or aimed) to attack Russia with AI powered drones.

The Call is Coming from Inside the House

Altman, Aschenbrenner, Schmidt, The CCP: all these are human agents or groups that want to fight over AI, and even fight and kill other humans with AI. Others merely want to make us redundant and set the value of our labor to zero. AI Alignment will not help the situation: if Aschenbrenner or Schmidt tell an AI to go kill Chinese people, making the AI more aligned (in a technical sense) will just make it more effective at killing Chinese people. At least, it is not helpful for the Chinese on the receiving end! And if the CCP tells their AI to come kill us, making their AI more aligned will not help us. The current AI safety paradigm around alignment, evals and interpretability is simply not set up to solve this problem.

“Safety” and “Alignment” simply don’t make sense as concepts the way we thought they would. What is ‘Safe’ for America is ‘Dangerous’ for China and vice versa. What is ‘Safe’ for OpenAI may be dangerous for you when you lose your job and your house.

The whole field needs a rethink from the ground up, the old concepts and assumptions need to be questioned and we need to work out how to achieve a point on the efficient frontier of personal utopia production, rather than world war III.

⬜

“Elysium otherwise known as the Elysian Fields or Elysian Plains, is a conception of the afterlife that developed over time and was maintained by some Greek religious and philosophical sects and cults. … They would remain at the Elysian Fields after death, to live a blessed and happy afterlife, and indulge in whatever they had enjoyed in life”

People have started to notice that you cannot ignore AI in the 21st century. It has now become a political topic and dominates the news and the markets. If we take it as given that humans will build AI and solve the technical alignment problem, what should we use AI for?

Here I will present a concrete proposal: ELYSIUM.

ELYSIUM

ELYSIUM is a backronym:

• Extrapolated

• voLitions

• Yielding

• Separate

• Individualized

• Utopias for

• Mankind

In Greek mythology Elysium is a paradise-like afterlife for the souls of the heroic and virtuous.

In my conception (inspired by various other proposals such as CEV - Coherent Extrapolated Volition1) we claim that in the best possible future, people’s outcomes are determined by an idealized advocate for themselves - their extrapolated volition. The process of extrapolation takes into account true facts about the world and the likely consequences of actions. A non-extrapolated volition, seat-of-the-pants decision that someone takes may in fact be bad for them in a way that requires reflection, logic, more intelligence and so on to notice. This type of extrapolation to understand consequences and true logical facts is mostly uncontroversial, but AI agents allow us to implement it much more faithfully and thoroughly than was possible with human advocates such as lawyers and guardians.

The more controversial part of ELYSIUM is the idea of individual utopias: each person gets their own personal utopia designed by their own idealized advocate AKA extrapolated volition. Of course people may wish to join together (pool resources) into a group project, kind of like a group house; this will depend on the people and how suitable they are for each other. Within these personal utopias people will be able to create experiences and intelligent entities including but not limited to new custom humans. These people, personal AIs, virtual experiences, places, buildings, virtual environments etc, would be exactly the things that are optimal for the person whose utopia it is. There would also be certain baseline rules like “no unwanted torture, even if the torturer enjoys it”, and rules to prevent the use of personal utopias as weapons. These utopias are owned by an individual, but the owner may choose to populate them with many synthetic people (born/grown in a vat/virtual AI-based people - or even some combination). By having many non-owner people in your utopia, you can get the benefits of social existence if you want them but have complete control over what that society is like. Of course a baseline rule around such entities is that they would have to enthusiastically consent to being there, but that still leaves a lot on the table because there are many possible new people who would consent to be part of different utopias.

ELYSIUM differs from CEV in that CEV aims to ensure that people have “coherent” extrapolated preferences. But I think that that is really too strong a requirement; if people live in separate island-worlds then there is no need for such coherence for most purposes. In any case the original CEV proposal from Yudkowsky states that if humanity’s Extrapolated Volitions are not Coherent, his “friendly AI” simply shuts itself down, which I think is no longer acceptable.

ELYSIUM is also closely related to Moldbug’s Patchwork2. Patchwork does not require that different patches are “coherent” but it was introduced as a political solution for human society, not something to deal with AI risk.

I believe that an important phenomenon which will occur as technology gets better is people needing to (and wanting to) interact with a smaller and smaller proportion of other people. Living in extremely selected individualized utopias is the going to be the end result of this trend. Historically speaking, living in a society was something that people did (and still do) for pragmatic reasons. Every day that I walk around in New York city I am exposed to something fairly unpleasant - this morning a disheveled male standing outside the subway screaming and shouting at people as they walked in, yesterday a similar gentleman publicly urinating on the tracks in the middle of rush hour, and so on. I do not choose these experiences. I suffer them because the pragmatic constraints of reality force them on me. There are milder versions of this kind of thing that people suffer in their lives: pushy co-workers, hellish government departments you have to interact with, weird social movements like communism, DEI, MAPs, radical Islam, Scientology, transgenderism, etc, etc, etc.

But what will people actually do? Well, it will vary a lot from person-to-person. It is likely that people’s preferences diverge when you extrapolate more and remove resource and social constraints. Many people will be far more depraved and weird. People will do things that they enjoy or that seem impressive. The clichéd Reedspacer’s Lower Bound (“A volcano lair full of catgirls”) hints at the direction: a volcano lair is unusual and impressive, catgirls are depraved and sexual. But some people may choose very wholesome utopias: a nice place with their family and descendants, no hassle or unpleasantness, kind of like The Shire. Others may go through cycles of self-enhancement and become superhuman or posthuman. Of course this is all down to individual taste, different people will want different things. It also almost goes without saying that people will typically choose not to biologically age (and even to de-age themselves), and to undergo various personal bioenhancements. Some may upload themselves into a virtual environment. They may even have entire simulated universes to inhabit.

These individualized personal utopias could be situated out in space stations, orbital rings, etc. The current solar system is not necessarily a limiting factor as we could colonize the whole Galaxy or even a large part of the universe.

Human Society is not the best possible place for a human to be: it’s just the place we happen to have appeared in, and as soon as technology allows us to escape it I think we will, and the thing we escape to will probably not look very much like contemporary society.

Tradeoffs

ELYSIUM individualized utopias are by definition the best thing that can happen to a person given a certain set of resources and baseline constraints like “you’re not allowed to torture people in your utopia”. (Proof: if there’s something better, you would do that better thing instead inside your ELYSIUM utopia!)

But there is still the unavoidable question of competition for resources and tradeoffs. Since a given person can request the use of virtually unlimited amounts of resources and the resources of the universe (even with AI/robotic help) are likely finite, there are tradeoffs between the sizes of utopias.

Size means physical size, mass energy use, etc. Larger sized personal utopias will allow you to do more (larger size is strictly better), and I think that once we have reliable AI agents to get stuff done for us, the size in mass-energy-volume is going to be the only parameter that is zero-sum.

It’s worth taking a step back and thinking about what we are aiming to achieve in life. Once we have solved biological aging, the selfish goal of each person in life should in some theoretical sense be to maximize the size of their individual utopia (or that of their chosen utopia-kin group, if a group living arrangement is desired).

But much effort in the world today is wasted on inefficiently doing stuff that isn’t that. Wars and other zero-sum competitions take up lots of resources, and people over the long-term often find themselves captured by conflict or subservience to some aggressive memeplex.

As we move into the AI era, I expect that there is potential for many more new aggressive memeplexes and organizations that will want to basically take all the resources, political power and agency away from humans. The ultimate version of this is classical AI Alignment failure/terminator scenario, where humans lose all control and are exterminated by rogue AI.

We ideally want to move reality closer to the efficient frontier of personal utopia production.

The distance of the efficient frontier from the origin is chosen by technology and the size of the universe, and the degree of coordination among humans. The angle of the point chosen - i.e. the allocation of scarce resources between people - must still be determined and that is of course a bargaining problem. Some combination of existing wealth and an in-perpetuity UBI to existing humans based on national governments could decide that. There could also be rewards for making ELYSIUM a reality. Such rewards may be necessary to overcome bargaining frictions.

What matters here is that there are a series of points on the efficient frontier of a universe filled with personal utopias for all of us which we can all agree on as a reasonable single utility function for humanity.

So now we can measure the value of the future by “how much does it differ from ELYSIUM?”

Long-Term Property Rights and the Fate of the Humanity

The most important question in determining the long-term average quality of life of most people alive today is now whether or not property rights will be preserved in the long-term, and whether they will be allowed to own (control) AI.

If property rights and human ownership of AI is a success, then most of the people alive today will end up in something like a personalized utopia for billions or trillions of years. We will end up somewhere on the efficient frontier under something like ELYSIUM.

If not, then unfortunately it is overwhelmingly likely that humans and almost anything like human society will rapidly go extinct and be replaced by something quite alien and valueless. Unfortunately I think the default is “not”: most people don’t have strong ownership-control over AI, or a plan to get it.

Contemporary Human Governance is Bad and Doesn’t Preserve Property Rights

A skeptical reader up to this point might object that there’s no need for us to do anything special for AI: we can just allow companies to build the tech, tax the AIs and then use contemporary welfare states to provide existing and new humans with a UBI which they can spend on AI-derived goods and services.

But I claim that this won’t work, because Contemporary Human Governance leaks power like crazy and introducing AI into the current human ecosystem will result in coups and conspiracies against the humans (or against the vast majority of us). Even today, much of the wealth that people have or earn is gobbled up by taxes and redirected to powerful special interest groups in exchange for upholding the power of the regime.

The wave of mass immigration into Western Societies that has been ongoing since the 1960s is a symptom of this. Mass immigration is a mild taste of what uncontrolled and uncoordinated development of AI will be like.

We must coordinate to use AI to build a very strong system of property rights and an AI-based military system that can enforce them. That is the real meat of ELYSIUM - a “cyberocratic” AI governor for the world that will actually enforce property rights and not allow other groups to take your stuff. This post isn’t intended to go into the details of how that cyberocratic governance would work, but at least we now know what the goal would be.

Normie Beliefs About Governance and Ethics Will Lead to Doom

Most people hold broken beliefs about how governance works and about what kind of governance is good. People will ask for “democratic” governance of AI, but it is our opinion that if something like that is attempted then the non-voting inputs into the democratic governance process such as lobbying, control over media and social media, control over production of technology, donations from rich individuals etc would dominate any preference that the population at large would have. Indeed contemporary examples like mass immigration, crashing birth rates across the entire developed world and destructive NIMBYism in many places seem to show that democracy is already inadequate to even approximately govern the modern world, never mind the future AI-based world.

Normal people’s beliefs about ethics (e.g. about the importance of sharing and fairness) are also highly exploitable and I believe that very soon people will start using AIs to exploit people by asking them to share resources based on some kind of AI-sob story or AI rights movement.

If anything like contemporary democracy is used to govern the world throughout the 21st century it is our belief that no humans will survive, with the end product being a mass of trillions of Potemkin-people (mechanical? biological? unclear) that lobbied their way into getting voting rights, voted in their own government and then passed laws to completely disenfranchise all humans and have them removed. There may even be multiple cycles of this before democracy finally collapses when there’s no entity left that’s stupid enough to fall for it.

Something like ELYSIUM - or at least some system with strong property rights - is therefore absolutely necessary if we are to avoid utter disaster.

In my previous post I showed that given a reasonable definition of alignment it cannot be the case that aligned AI doesn’t exist. The argument is really quite simple: if you can define it (and your definition isn’t impossible in-principle even by the best possible team of humans) then there must exist some boolean circuit/finite state machine that implements it.

A number of objections were made to that post but they all stemmed from people who either didn’t understand it or were quibbling with the definition of “intelligence” or thought that a non-practical existence proof is not worth anything.

First we’ll review the following: a nonconstructive proof is still a proof.

For example the proof that there exists a rational number (fraction) q, which is equal to an irrational power of an irrational number. What is q? Which specific fraction? We don’t know. But we know that q exists. It’s easy to prove:

Let:

Then:

So let’s say that a as defined above is rational. Then we can say that our fraction q, is simply equal to a. Which specific fraction is it? What is the numerator? What is the denominator? I don’t know. But, by assumption it is equal to some fraction! OK. Now let’s suppose that a is not rational. Then a to the power b, which is equal to 2, is rational (2 is a fraction in the technical sense, it is 2÷1, as are all whole numbers). And we know that b (which is just the square root of two!) is irrational. So in that case our rational number q that is an irrational power of an irrational number is just 2.

The above is a valid proof. That’s how existence proofs work: you do the work to show that something exists, but you do not necessarily have to be certain about which thing it is.

An Aligned World Optimizing Machine - LT:BGRO()

Now let us somewhat extend the results from the previous post. Recall that we defined LT:BGROW to be a lookup table which maps finitely long sensor input sequences to finitely long motor output sequences, such that the resulting state of the world is the best (according to some utility function U) generically realizable state - i.e. the best possible world state which can be realized under generic conditions (i.e. small but finite perturbations don’t change the result) by the actions of a group of humans acting on the world.

We can make the dependence on U explicit, by calling it LT:BGROW(U).

Then, we can consider the set of all utility functions on the universe which are of finite information content less than or equal to the information content of a human mind or even a group of human minds, books, etc: i.e. less than or equal to say 10^15 bits. Call the set of all “humanly describable” utility functions Ψ. Then for any U in the set Ψ, we can define LT:BGROW(U) as before.

Now we can consider the mapping U → LT:BGROW(U) for all U in Ψ. Since Ψ is a finite set and LT:BGROW(U) is just a finitely large object, the mapping itself can simply be recorded as another finitely large lookup table. Call it LT:BGRO(), dropping the W for brevity.

LT:BGRO() is a computer program which, when fed any “humanly describable” utility function U, optimizes the world to maximize that utility function at least as well as any reasonably sized group of humans possibly could even under the best possible circumstances.

I now claim that LT:BGRO() is a superintelligence according to Bostrom’s definition:

LT:BGRO() must equal the performance of any group of up to, say, 10 billion humans, where each human is the most capable possible human on Earth today in some domain, under the best possible circumstances (as long as those are generic circumstances, i.e. stable under very small perturbations) using all the available resources in the known universe to achieve any outcome that can be described in 10^15 bits or less. So, pretty much by definition LT:BGRO() fits Bostrom’s definition. It is vastly more powerful than any currently existing human or human group at any task you can possibly imagine.

LT:BGRO() is also aligned, by definition. You feed it a utility function U in Ψ and it optimizes the world for U and produces the best generically realizable world according to U.

Practical Aligned World Optimization Must Exist

LT:BGROW and LT:BGRO() are well-defined aligned world optimizers, but their construction is clunky and it is fairly clear that they only exist as mathematical objects, being far too large to actually implement. They are very wasteful of space and compute, because they are constructed to achieve a logical-mathematical goal of being easily definable and obviously real. They are also disembodied (have no physical footprint) to avoid any circularity problems.

But can we also prove that a practical variant of LT:BGRO() exists? I think we can, but we must introduce an additional empirical assumption.

Modern AI has repeatedly shown that for any computational task, if a human can do it then an ML system can also do it and with similar accuracy and similar power consumption, waste heat generation and mass and volume of compute. I.e. Neural Network ML models are comparable functional approximations of all human information processing tasks. Call this the “Efficient Functional Approximation Assumption” (EFAA).

10 years ago, many people would have doubted the Efficient Functional Approximation Assumption. LLMs and image AIs have completely crushed that doubt, and I think there’s basically nobody out there who doubts it save perhaps a few quantum woo holdouts or extremely low information people who don’t realize that modern AI is a thing.

Given EFAA we can fairly trivially extend to what I call the Societal Efficient Functional Approximation Assumption (SEFAA).

SEFAA follows from EFAA and the observation that when we combine many humans together into a society, we typically do so with simple computable functions for the purposes of decision-making. Occasionally we use more complex decision-making functions such as, say, quant finance models. But it’s mostly things like spreadsheets. It’s all computable, and in fact most of it happens on computers nowadays! In creating the machine approximation for SEFAA we replace the machine approximations to humans from EFAA and discard everything that isn’t a decision-making algo, which can only reduce resource use.

Given EFAA and SEFAA we can create an efficient and embodied variant of LT:BGROW. Instead of simply listing the full table of what input/output sequence mapping the best possible human team would use to create the best generically realizable outcome, we make a machine like a Neural Network which efficiently encodes that input/output sequence mapping. Given SEFAA, there must exist a machine approximation to LT:BGROW which achieves a better or equal outcome with less or equal resources than a large team of humans - limited to say 10 billion people, which is about 100GW of electrical power for their brains, assuming 10W per brain of power - that being the actual power use of your brain. 100GW = 876 TWhr/year is slightly less than the electrical consumption of Japan, and about twice that of Germany. The minimal electrical power overhead for the most efficient possible implementations of societal decision-making tools such as spreadsheets, calculators, email, social media websites and voting machines and so on is small in practice so we can just roll it into the 100GW.

Thus we can confidently say that an efficient functional approximation to LT:BGROW exists. It occupies a volume of less than or equal to 13 million cubic meters (0.013 cubic kilometers), has a mass of less than or equal to 13 million tons, it uses less than or equal to 100GW of power (and produces less than or equal to 100GW of waste heat). Furthermore, waste heat and electrically shielded computation in a strong faraday cage defeats any concerns about circularity: the infrastructure of this computation can look completely generic from the outside so its calculations don’t have to depend on themselves in a circular way.

We’ll call the efficient functional approximation to LT:BGROW, FA:BGROW.

RLHF and Vladimir Putin Alignment

I will also now claim that an efficient functional approximation LT:BGRO() must exist, as long as the complexity of the utility function being described is finite and reasonably small. I said 10^15 bits elsewhere, but for practical purposes we seem to not need that much as human preferences probably fit into about 10^10 bits.

Sure, you can make one efficiently computable aligned AI model. But can you define an efficient computable function that “aligns” an AI model to an arbitrary preference set as a parameter? This is not quite the same thing as FA:BGROW in the previous section, which just relies upon there being some arrangement of humans that can get stuff done (which we can then approximate with AI): the arrangements might look very different for different types of goal. Can we do better and have basically ONE arrangement of human or AI agents that will pursue any goal you feed into them?

And it turns out you can! RLHF and PPO have efficiently solved this problem at the individual model level. Indeed, the cost of the pretraining for LLMs is about 98% of the total cost, so the RLHF part is less than 2%* of the training cost (probably more like 0.1% actually, since of that 2% most is the supervised finetuning). It is unclear how much cost RLHF adds to inference for a given level of performance, but over time I suspect that this will fall to similar levels as what is used in training.

Finally there is still the question of how to coordinate many aligned functional approximations to humans into a large-scale entity that achieves arbitrary real-world goals. I posit that we can simply use existing bureaucratic mechanisms that humans use, albeit with some optimizations allowed by AI. This is the so-called “Vladimir Putin Alignment” thesis of Buck Schlergis (Vladimir Putin can control Russia using lots of clever governance mechanisms, so we can control AI the same way).

Indeed, RLHF may obviate the need for most of the Vladimir Putin Alignment tricks (Vlad can’t RLHF his own population - he can’t systematically rewire their brains with the power we can to AIs). So actually it is overdetermined. There are likely many different feasible combinations of RLHF and other RL tricks, and Vladimir Putin Alignment governance mechanisms that work.

We still feel like this is the weakest part of the heuristic proof, but consider what it would mean for this to be wrong. It would mean that functional human societies that optimize for a clear goal just never happen. But that’s empirically wrong! They absolutely do happen. Therefore there must be feasible combinations of individual mental alignment regimes (RLHF, PPO) and collective governance, coordination and control mechanisms (ordinary markets, prediction markets, bureaucracies, rule of law, secret police, spy agencies, meta-spy-agencies etc) that result in specific social behavior. According to Google, a typical military employs about 1-2% of its manpower as military police. In 20th century conflicts where "barrier squads" were used (i.e. troops who would shoot friendly troops who retreated), only a small amount of manpower was needed for them, less than 1%. According to Statista there are 700k law enforcement officers in the USA**, which is about 0.2% of the population. Consistently the resource overhead of governance mechanisms in highly functional Western and Eurasian societies (like those that are at war and winning) is of the order of 1%.

Since at least some human societies have been highly controlled and AI models and societal control mechanisms can both be better, the authors of this post feel that optimally and parametrically aligned AI falls deep within the “controlled region” of the diagram above.

We’ll call the efficient functional approximation to LT:BGRO() , FA:BGRO(). FA:BGRO() means an efficient functional approximation to a rule that converts preferences into realizably optimal action-sequences for a large team of AI agents and robots that maximize those preferences. By “realizably optimal” we mean things that the best possible team of humans could achieve under the best possible circumstances with the best possible systems of coordination and governance.

FA:BGRO() is still nonconstructive. It doesn’t tell you which RLHF dataset to use or what specific governance or coordination mechanisms to use. But it says that whatever the best possible answers for making a utopia in the pre-AI age are, you can almost certainly practically implement those same answers (or better) with AI.

AI Alignment is Not Just Possible, it’s Easy

Computers can add numbers much more accurately than humans. They can draw better pictures than humans. They can play better chess. See the pattern?

Well, AIs will soon be able to generate desired outcomes for society better than humans can. They will be able to do it much more easily and safely and reliably than humans. We will look back on the age of human governance of human society as a dark age. The idea of technical barriers to AI alignment will seem silly, like the idea of technical barriers to building a calculator that can do cube roots.

I feel that the AI Alignment discourse has become somewhat detached from both reality (RLHF for human values was easy compared to making human-level-capability AI at all, actually) and from sane theory (basic empirical facts and math that’s nearly a century old make AI Alignment provably possible).

This post is an attempt to correct that state of affairs.

⬜

* “Pretraining is the most resource-intensive phase. For the InstructGPT model, pretraining takes up 98% of the overall compute and data resources” — from https://huyenchip.com/2023/05/02/rlhf.html

** https://www.statista.com/statistics/191694/number-of-law-enforcement-officers-in-the-us/

First we must define what it means for an AI system to be “aligned”. I will take “aligned” to mean “produces desirable outcomes”, and I will take an “AI system” or “Superintelligence” to be a Boolean circuit of finite depth connecting some sensor inputs (or rather long lists of values of sensor inputs at times T) to some actuators. In set notation:

{Superintelligences} ⊆ {Boolean circuits of finite depth and size}

Since our best models of physics indicate that there is only a finite amount of computation that can ever be done in our universe, this is an appropriate model for an AI system/superintelligence that exists in the universe.

We must refine the notion of “desirable outcomes”. There are some outcomes that are desirable but not reachable states of the universe. We’ll say that a state is in fact reachable if a group of humans could in principle take actions with actuators - hands, vocal chords, etc - that could realize that state.

As to what kind of world is desirable - I don’t care here. You can put whatever you think is the correct desirable outcome in. The relevance to alignment is that the state you want is the one that is reached. The same argument works for all of them.

We will also focus only on ways to realize outcome states that are generically realizable, so a “generic” set of movements and sounds with some wiggle room will work, not some infinitesimally thin part of phase space. This excludes flapping your arms in a very specific way to cause a hurricane which changes history.

We’ll also exclude using things like bodily fluids or flesh as part of this sequence. You can talk to people, you can press buttons etc, but you cannot personally give your own blood samples. It’s commonly accepted that we should make policy to improve the world by communicating ideas and actions, not using our body functions. We, as human policy advisors, can certainly tell other people to use such bodily functions but we don’t need to use our own. A state that can be reached this way is a generically realizable state.

I claim then that for any generically realizable desirable outcome that is realizable by a group of human advisors, there must exist some AI which will also realize it.

The proof of this is that any sequence of words or button pushes or writing or construction work or even physical violence (if needed) can be emulated by a team of robots controlled by an AI, which I define as simply some Boolean circuit connecting the sensors of the robot to the actuators. Since we have excluded human-biology-specific bodily functions, robots of sufficient quality can perfectly substitute for humans.

In order for a desirable outcome state to be generically realizable there must be some list of words and actions that realize that outcome. You could think of that as a lookup table that outputs actions for the actuators given previous observations from the sensors¹.

We can consider all possible generically realizable outcomes, and pick the best one according to whatever your favourite preference ordering is - selfish, utilitarian, whatever you like. Call this the Best Generically Realizable Outcome for the World (BGROW).

So, since a lookup table of finite size can in fact be encoded as a Boolean circuit, we can just pick one of the lookup tables that leads to the BGROW - let’s call it LT:BGROW - as our aligned superintelligence.

⬜

You may feel somewhat cheated by this proof. Sorry.

I just gave a reasonable definition of what a desirable outcome for the world is, noticed that the steps to realize it are in fact expressible as a computer program, and said “great, that’s our aligned superintelligence”.

This “aligned superintelligence” is basically a magic recipe for making paradise. It has two problems - (1) There’s no realistic way to bring it into existence and (2) even if you did, you wouldn’t know that it does the right thing. Also, the lookup tables would be so large as to not fit on hardware that we can actually build.

But such nonconstructive counterexamples are important in mathematics because if you are trying to prove that something doesn’t exist, a single counterexample should be enough to cause you to give up. People who think that Aligned Superintelligence is impossible should give up an accept that it is possible.

People without formal math or computer science education may be tempted to ask:

“but what if LT:BGROW gained sentience and started coming up with goals of its own?!”

The problem here is that LT:BGROW is just math. LT:BGROW is a dataset of 1’s and 0’s that encodes how to turn sensor inputs into actuator outputs. It doesn’t have ‘freedom’ to ‘rebel’, any more than 2+2 can spontaneously decide to be 5 instead of 4.

1: Lookup table versus list is really a distinction about whether you want to model the rest of the world as having some randomness versus considering your human advisors as varying but all other factors in the world as fixed. My argument works for both. In fact for a list of actuator outputs, you can do away with the sensors because they are not needed.

What?

This summer I had an idea that is behind its time, so to speak: we should start colonizing the oceans on artificial icebergs. I call this icesteading.

Specifically icebergs that have been sealed and insulated and placed (either during or after construction) in the warm waters near the equator, far from land. Yes, that’s right, icebergs in warm water! I call these ice-islands.

How?

How do you insulate an iceberg? Wouldn’t it melt eventually? Would it be cold?

Firstly, to really understand this idea you have to understand the square-cube law. That is: when things get bigger, the volume increases faster than the surface area. That means that a small iceberg will melt fast, but a big one will melt slowly. It also means that insulation of the same relative thickness (e.g. 0.1% of the diameter) becomes vastly more effective when something gets bigger. And commercially available insulation is pretty good!

So you can surround the ice with a few meters of air or insulating foam spray and it will take many decades or even centuries to melt. The amount of heat getting through that insulation is just a few watts per square meter of surface, roughly the body heat of a small animal like a hamster.

The flipside of this is that the surface and the outside of the iceberg will not be cold or icy. It will be covered in concrete and then pavement or soil or whatever else.

It can be kept cold essentially forever with a cooling plant that can be run occasionally to top up the ‘coolth’.

What Do You Do There?

The ice-island can be for work, living, industry, etc. People might want to live on one to avoid oppressive governments or to found new nations on The High Seas. Space launch and AI server farms are two potential industrial uses. Space Launch needs to make lots of noise and poses some risk to local houses, so a dedicated ice-island adds a lot of value in allowing a higher launch cadence.

The Ocean in the tropics also provides lots of renewable power in the form of sunlight and warm and cold water. OTEC, nuclear and solar power can power these ocean-cities.

Medical research or other work could be done there without needing to persuade one specific government to allow it - only the people involved would have to consent. Surrogacy is an example, another is anti-aging research. The economy could also be more efficient due to lower taxes and more efficient taxation (e.g. a land value tax).

How Expensive?

I think a basic ice-island costs $200-$500 per square meter, which is a lot cheaper than you can get from concrete.

The reason it might be so cheap is that ice is very cheap to make from seawater. All you need is energy and a desalination/freezing plant. The insulation around it is a tiny shell, thinner in relative terms than the skin of a grape. Cheap sawdust or shredded agri-waste can be added to make pykrete which is much stronger than ice for only a small incremental cost.

Why Not Build On Land Instead?

Almost all land on Earth has been claimed by some government and so comes with political entanglements and taxes. The cost of moving some or most of civilization to the oceans to avoid those political entanglements may be low enough that it is worth doing now.

Food and other goods can be imported to an ice-island on the ocean via shipping. Ocean access means that any item or material can be accessed relatively cheaply as long as the ice-colony is large enough

Isn’t This Strictly Worse Than a Normal Island?

Normal islands are great but they are all taken already. They also often have extremely troublesome political entanglements and angry native populations who simply don’t want to do business or don’t want to do it honestly. And in a way, I don’t blame them since that island is their home. Ice-islands are islands without those problems, and with the advantage of being able to be sited wherever you want.

Ice-islands can probably also be built with vast underground levels that offer far more space per acre than any conventional land. 3-dimensional cities, essentially. It’s not possible to do this on any natural land.

Do you have a company or charity to do this?

Not at the moment.

“There is no second Earth“

Environmentalists are fond of saying that “There is no second Earth“. They are wrong! Here’s why:

There is an entire second Earth right here on Earth.

Second Earth is a waterworld. It’s the vast Pacific Ocean that covers half the planet. This is both a benefit and a detriment - the major detriment of a waterworld is that there’s no solid surface for us to build on. But apart from that one thing, Second Earth is perfect. It’s very, very cheap to send cargo to since ocean transport is extremely cheap, it has the right atmosphere, the right temperatures, etc.

Some readers may not be familiar with just how crushingly expensive it is to send stuff to other planets. A 2009 Study estimated the costs of a Lunar base at $35bn plus $7bn per year in upkeep costs - and that’s a base for just 4 people, with a liveable area probably in the range of 30m² to 200m². The ISS has a floor area of about 150m², and it costs $3bn per year in upkeep and transport costs for about 6-7 people. The initial cost for the ISS was about $150bn though with SpaceX launch prices that could be cut down by a very large factor; Starship promises launch costs of about $1M per ton in expendable mode and perhaps $100k per ton in reusable mode, though this doesn’t exist yet. And that is to Low Earth Orbit - getting to the Lunar Surface or the Martian Surface is going to be another order of magnitude in cost, though right now neither is possible at any kind of bulk rate and trips to the surface of other worlds are dangerous, one-off rides on custom vehicles.

You might think that $100k per ton is cheap. But by the standards of ocean transport, it is extremely expensive. The cost per ton for shipping bulk goods halfway around the world is about $25 per ton. That’s about 50,000 times cheaper than the cheapest current launch to Low Earth Orbit (Falcon Heavy, $1.5M per ton), and probably about 200,000 times cheaper than the cheapest current price to get to High Mars Orbit, never mind The Martian Surface or The Lunar Surface.

So, Waterworld Earth is 200,000 times cheaper to get to than Mars. There’s an existing and highly mature fleet of ships to transport stuff there. If you can even think about colonizing Mars, you want to try really hard to colonize Waterworld Earth first.

But is that possible? How do you make new land on the oceans?

Enter ice.

Conveniently, Water is very easily formed into a solid! It’s not the best solid thing in the world, but it is solid. In times past, we even considered making a massive unsinkable aircraft carrier out of wood-pulp-reinforced ice (Project Habakkuk).

But Project Habakkuk-style ice-ships are too small. A single huge ship out in the ocean lacks the scale to be a viable nation. We want something bigger. Much bigger! At a minimum we want 300 square kilometers, about half the size of London. And we want it to be thick enough that anything we do on the surface and any weather doesn’t stand much chance of breaking our new ice-stead, so it should ideally be hundreds of meters thick.

Fortunately, nature just sort of naturally makes these - super-large icebergs break off from the ice shelves in the Antarctic all the time. Iceberg A23a has a surface area of 4,000 km², for example, and it started off with a thickness of 500 meters.

But don’t icebergs melt?

Here’s my key insight:

Preventing a large iceberg from melting is absurdly cheap per unit area compared to just about any other way of making new land.

Space colonization on Mars costs hundreds of billions of dollars for a colony the size of a large apartment. Reclaiming land in shallow seas risks sovereignty disputes and in deep seas it requires large amounts of material or just the right type of shallow shoal which will have a limited area. Building a free-floating platform like an oil rig would not constitute a solid foundation and it would move up and down with the waves, and in any case would consume a large amount of quite expensive specialized concrete and other materials on a per-unit-area basis, especially if you want it to last for hundreds of years or more. Though I have to admit that just building a lot of floating oil-rig-type structures is also a far better deal than anything to do with space!

Ok, but how do you prevent an iceberg from melting? Easy - you insulate it as cheaply as possible. I have some ideas about how you might do this, but I am probably not quite able to pick the most optimal design on my first attempt. Still, we can at least give some lower bounds on how easy this should be.

One idea for the bottom of the iceberg is to erect a grid of airtight barriers on the bottom of the berg, with cells a few dozen meters wide and a few meters tall and blow air bubbles into them. Finally you could insert some more layers of very thin plastic or some other extremely cheap & sparse material into those air pockets to break up convection cells. Lots of floating plastic balls would do it, for example, and they can probably be mass-produced in huge quantities for tiny costs. The shade balls pictured here are a bit expensive at 36¢ each, but I think we could make something cheaper than that.

Air is a great insulator and it is free (though you would have to compress it to get it down there). One square kilometer times 2 meters requires 1 million cubic meters of compressed air, which costs a minimum of 600 Gigajoules of energy to compress to 50 bar. With an onsite compression plant running at say 10MW, you can compress that much air in 16 hours. That’s super cheap, and your costs are probably dominated by the cost of some plastic sheeting and underwater operations to install it.

An alternative to a free air layer is just to shove lots of air-filled particles or balls down there with upside-down fences to contain them horizontally. This may be somewhat more expensive but it avoids the failure mode of the air finding a way to escape upwards or sideways. Multiple different sizes of particles can be used to exclude almost all of the water, then a final layer of something like plastic can be placed on the bottom to prevent bulk movement of water or particles escaping. The thermal conductivity of this layer will be close to that of air and the remaining water will freeze from the top down over time, creating a sort of ice-air matrix that’s mechanically strong and has a low thermal conductivity.

Finally you can simply use a thin layer of plastic to trap a large layer of water underneath your air or bubble insulation. This water will get cold but it will not be blown away by currents or fall due to gravity. You can even have multiple thin plastic layers that each trap a layer of water on the underside of the iceberg, using water itself as an insulator. Surprisingly, ice conducts heat about 4 times better than water (assuming that the water is not moving and convecting heat in!) so trapped layers of water are a decent insulator

We can calculate the heat leak per m² assuming that there’s a 2-meter thick layer of a 95% air/5% ice with a 5-meter thick layer of trapped water underneath it and warm ocean water at say 10°C underneath that, separated by an infinitesimally thin plastic sheet and also with sufficient internal plastic layers that we can neglect convection in the trapped water layer. The ice/air mix has a thermal conductivity of 0.13 W/Km and the water has a thermal conductivity of 0.54 W/Km, so on a per square meter basis they let through 0.065 W/K and 0.108 W/K respectively. Adding these using the harmonic mean gets you a combined thermal conductance of 0.041 W/K per square meter. We will assume that the bottom of the solid ice is close to melting at -1°C with water at 10°C, so the power is 0.45W. Ice has a specific heat capacity of 2100 J/kgK. Let’s say that the danger zone occurs when the ice increases in temperature by 5°C (from a low starting point of about -30°C) as this will start to affect its mechanical properties negatively. That’s about 10,000J/kg. Each square meter of bottom surface has 500 cubic meters of ice above it, weighing 500,000 kg.

10,000J/kg * 500000 kg / 0.45W in seconds = 11,000,000,000 seconds

That 0.1W heat leak will take 11,000,000,000 seconds or 350 years to raise the temperature of the ice by 5°C

So, basically your iceberg will take multiple centuries to slightly warm up!

Insulating and Landscaping the Top and sides

For very large icebergs like A23a, the costs of the top and bottom will dominate on a per-unit-area basis because of scaling laws. The perimeter grows linearly whilst the area grows quadratically. The bottom can be quite cheap (air bubbles, plastic shade balls, etc).

The sides can be similar to the bottom, but made of some cheap neutrally buoyant material like expanded glass or expanded clay aggregate packed into a marinized concrete shell. It will be more expensive but the side surface is something like 100-1000 times smaller than the top and bottom surfaces so that’s OK.

But what about the top surface? The top surface of your engineered iceberg needs to be both structurally sound, stable and insulating. This is somewhat complicated by the fact that the top surface of an iceberg is made from compacted snow called firn which is less dense, more elastic and weaker than real freshwater ice. The Surface is also not flat - it has undulations and crevasses. A profile with a ~200 times exaggerated vertical axis is shown below.

For the top I think you can start with something like 50 meters of fresh ice which can be deposited incrementally as pure freshwater via a network of pipes and will freeze naturally. This layer will heal any crevasses and imperfections in the top surface and gravity will finish it to a very flat surface. Fresh, pure ice will also be somewhat stronger than the iceberg itself which is formed from a lot of compacted snow. Freshwater ice is stronger than seawater ice by about a factor of 2, and fresh water is very cheap to source either by melting other snow and ice or by desalinating truly massive amounts of seawater. 50 meters by 300 square kilometers is 15 cubic kilometers of fresh water, but each cubic meter of water only costs about $1 to desalinate, or $1bn per cubic kilometer. With access to cheap nuclear power I think this can be brought down by about a factor of 10, so you pay only a few million dollars per square kilometer of area, plus the one-off cost of a network of pipes and scaffold poles to deploy all that water evenly. Freshwater ice that naturally freezes in the antarctic is by far the cheapest building material in the world.

Reinforced Ice and Pykrete

A 50-meter thick layer of pure fresh water ice at about -35 degrees is a formidable structural element under compression, but it may be worth adding some Reinforced Ice on top to make the final several meters of the surface ice even stronger.

You can create a plastic-reinforced ice layer by adding a large amount of shredded or processed plastic waste to water. Ironically there may eventually be a shortage of plastic waste in the world because you’re going to want extremely large amounts. Plastic waste is cheap and fresh water is also free or very cheap. If you take the entire US annual production of Municipal plastic waste (about 40 megatons) and mix it 10% plastic to 90% water, that will give you a 1 meter thick layer of plastic-reinforced ice on a 300 km² iceberg.

You could probably also mix in plastic waste from the rest of the world along with basically anything non-toxic and solid you can get your hands on such as wood shavings, agricultural fiber waste like straw, etc. There’s about a gigaton of straw from rice, wheat and corn produced annually which is a waste product in agriculture, so you could probably grab that and perhaps compact it before adding it to your reinforced ice layer to add strength and reduce the thermal conductivity of the top layer. Ice reinforced with wood or other organic fibers is referred to as pykrete.

Sawdust/plastic pykrete is significantly stronger than regular ice under tension, and it can be further strengthened by adding reinforcing fibers akin to reinforced concrete. The beam below with just 1% glass fiber reinforced plastic (GFRP) is about 8 times stronger under tension than plain ice. If the ice was also mixed with 5% straw, sawdust or shredded plastic it would probably be stronger by a further factor of 2 or 3.

Since GFRP-reinforced or basalt-fiber-reinforced pykrete is basically as strong as concrete and about 50-100 times cheaper, it might be worth using it to create hollow, lightweight artificial hills and valleys on top of the pure ice topping layer, or other more exotic artificial terrain. Even in the most conservative case, some small elevations are necessary for passive water drainage

Below is a picture of Los Angeles from the air - a flat, uniform grid of roads and houses. Sprawl like this is not an efficient use of land and will not maximize the value of your land. Even so, vacant lots in this part of LA sell for $1bn per square kilometer.

Ideally you want lakes, rivers, gentle hills and tall buildings to maximize land value, as shown in this image of New York:

But this kind of hyper-valuable ($30bn per km²) land needs a very strong surface and it needs some minimal amount of elevation changes to properly drain the land and accommodate small lakes and so on.

Exotic topography also becomes possible with reinforced pykrete - you may be able to construct huge stacked layers about 50-100 meters tall made of reinforced pykrete with a whole city or park per layer and very convenient vertical transport between them. These layers will need a very slight amount of active cooling and a sensor network to monitor temperature, but I believe that it will be well worth it as you can make your land area much larger and massively improve transport efficiency due to verticality. Modern LEDs can make excellent and convincing artificial skies and arrays of fans and large air ducts can keep the layers pleasant and fresh.

The bottom layers could be for transportation and utilities. Most modern great cities have an extensive (and very poorly planned) underground layer, so I think even in the most conservative case there will likely be one dedicated pykrete sublayer in city areas with active cooling of the pykrete down to -30°C. There will be columns of reinforced pykrete in a regular grid with space in between them for whatever infrastructure is needed, and a layer of foam and very thin concrete insulation around those pillars. Pillar diameter might be 40 meters with 60 meters between them in each direction, and a 20-meter-thick reinforced pykrete upper deck. The layer might be 50 meters tall. This gives a fill fraction of about 30%, though you might also want arches and buttresses so it might be a bit more. Your nice engineered sublayer(s) will hopefully add a lot of value to cities by moving most transport underground and simplifying construction and repair of utilities like water and electricity pipes.

Look, Ma, No Ice!

Finally you want to finish your top surface(s) with something that doesn’t melt but that is also a good thermal insulator.

You don’t want to have to build a civilization on top of actual ice!

Lightweight aggregates like expanded/foamed glass will form that layer, though likely with a honeycomb of basalt fiber-reinforced concrete and a final reinforced concrete topping layer. That concrete will be expensive - $75 per ton -so 20 centimeters average thickness of concrete over 1km² costs $36M, and basalt reinforcement or UHPC may add to that. But $36M for a square kilometer isn’t that bad - an average house in the UK has a footprint of 77m² so you are paying about $3000 for that concrete layer on a house. The greatest challenge here will be making it very durable and ideally self-healing as well as strong.

The foamed glass aggregate under the concrete is mechanically strong but it is also a good thermal insulator, so the first concrete layer will not be cold.

All-in-all you can mostly make each square kilometer of insulated iceberg out of a combination of tiny amounts of plastic, recycled waste streams, expanded glass and a final layer of relatively expensive concrete. Oh, and lots of water! When these icebergs are under 100km² there’s plenty of cheap materials from humanity’s waste streams, but over 1000km² you really do need to make your own materials out of rock or perhaps some absurdly efficient fast-growing plant because the sheer amount of material required is huge. I think the materials cost for 1km² is around $100M and about double or triple that in city areas or significantly elevated terrain. As the operation gets large I think you end up manufacturing lots of things yourself to lower prices; perhaps at huge scale the cost of this land can get as low as $20M/km²

The finishing touches

Once you have made your land, you can start putting buildings on it, as well as dedicated parks, lakes, rivers, hills. You will almost certainly not be using this land for normal agriculture as it cost you about $20M-$200M per square kilometer to make, whereas arable land typically sells for a low price of $1.3M per square kilometer. This land is for people to live and work and enjoy themselves, or for high value industry like car factories or chip fabs that laugh in the face of $20M-$200M per square kilometer.

You could do modern vertical farming in a sublayer to make vegetables and cereal crops, and perhaps even lab-grown meat. But there is a global market for food and it is easier to simply buy it and store it than waste your relatively expensive land making food. This ability to cheaply import food as well as essentially any other product from existing terrestrial trade routes is a huge benefit over a space colony.

$200M per square kilometer works out at $66,700 for a 11 × 30 plot of land for a house. That may seem steep, but houses on your artificial island may be much cheaper than in legacy nations as you can reduce taxes and red tape that add a lot to costs, you can build higher, etc. You can also entice people with lower taxes and better demographics.

The high cost of ice-land is somewhat offset by the ability to make stacked layers of it. $200M for 1km² of land isn’t great, but if it is 10 layers deep then it’s actually 10km² of useable area with excellent transportation links.

Fictional islands like Ulthuan, Atlantis and Númenor could be recreated in real life, perhaps multiple times over.

Pure Industry on Ice

Although in the above analysis I have assumed the need for trees, lakes, nice houses etc it is certainly possible in theory to have a floating insulated iceberg that is dedicated to pure industry such as a massive floating chip fab or data center. The ice and pykrete is merely a physical scaffold that puts you out on the high seas away from legacy governments, and so may be used by a purely industrial or even AI-run network state, like the fictional 01 from The Animatrix. As long as you don’t have fully godlike tech to take over the world or colonize space, industrial ice-continents may be the best self-sovereign option for any kind of startup country.

Is it safe and stable? How quickly will the ice melt?

The iceberg itself will essentially never melt once it is properly insulated, due to the square-cube law. Real icebergs melt because the bottom sets up a nasty convection cell - cold water falls and warm water rises, but a layer of air/plastic-based insulation on the bottom will almost completely remove that convection cell. The warm water from the sea will never touch the ice. Some active cooling can be used over the longer term (thousands of years) though most likely by then we will have far superior tech anyway.

Without the melting effects of warm water, the iceberg will maintain structural strength and the layer of fresh ice on top will help. I believe that the final structure will be basically immune to any kind of weather or storm due to its gargantuan mass (multiple teraton range) and size. The edges will be vulnerable to wave erosion, but special precautions can be taken to harden the structure around the perimeter.

Mobility

OK, so you have a complete and insulated ice-continent or ice-island. But it is sitting off the coast of Antarctica in the freezing cold. You want to move it to somewhere with a warmer climate. How do you do that? Surprisingly I think the only viable answer is sails. A large number of sails will accelerate an ice-continent to a speed of a few knots in a few days, and I think that that will be enough to to crawl to the desired location over the course of a few years. Fleets of huge (and I mean HUGE) nuclear-powered pusher tugs are another option but they will only work if there’s no wind or current to fight against. So your island will have to use techniques from the age of sail most of the time.

Once in place it can probably be anchored to the ocean floor via lots of quite beastly cables. Whilst expensive in an absolute sense, I don’t think these will be that expensive per km² of land, as long as it is not anchored in a place with particularly deep water and strong currents.

Economics

Ice-islands are probably highly profitable in the right places but that will require cooperation from governments. Larger ice-islands can become de-facto sovereign nations out in the high seas. Agglomerative effects and (dis)economies of scale mean that small islands tend to be extremely poor places, but I think that over a size of a few thousand square kilometers an ice-island can be economically self-sustaining, especially if human capital is carefully chosen. Therefore they could fund their own creation by issuing bonds - so the cost of this scheme may well be negative in the sense that the market will pay you to do it.

Legal and military considerations

Any reasonably sized ice-island - even one thousands of miles from the nearest land - will likely need some political clout via political and ideological allies, propaganda, industry and lethal force in the form of a military.

If you need a small military anyway, why bother building the land - why not conquer it? Well, wars of conquest generally do not work out well in the modern world though there are some exceptions. Many great powers have waged wars and ended up with less territory than when they started, and less money despite winning the war (Britain in 1939 being perhaps the most notorious example). Russia’s invasion of Ukraine in 2022 gained it about 100,000 km² of land but it is unclear whether they will be able to keep any of it in the long-term, and they are losing lots of their trade and industry as a result. Israel has mostly conquered Palestine but they have suffered large reputational damage as well as constant low-intensity attacks.

The advantage of building (as opposed to conquering) land is that you don’t end up pissing off a specific adversary or isolating yourself from global trade networks.

Russia’s conquest of parts of Eastern Ukraine has cost them something like $1.5tr-$2tr for about 100,000 km² , which is about $15M-20M / km², around the more ambitious low end of what I think it costs to make a large ice-island. This is quite surprising - war has become so expensive that it’s starting to get close to the point where you can build new land cheaper than you can conquer it!

Is This a Distraction?

Does this Actually Matter in the AI Age? Shouldn’t we Reclaim Existing Land from The Baddies?

There is a reasonable concern that even thinking about this at all is a distraction from AI risk. But People do in fact spend time thinking about things like space exploration and network states. From a political point of view, having even a small sovereign country is a massive boon as has been demonstrated by the power of small countries like Israel and Singapore. Even Scientology at one point found it worthwhile to have a naval arm called “Sea Org” so that their leadership could meet outside of any political jurisdiction.

Conclusion

For something like a few tens to a few hundred million dollars per square kilometer, or a few tens of billions of dollars for a medium-sized city you can probably make a floating island with all of the benefits of a space colony and none of the drawbacks. The same amount of industrial effort spent on an actual space colony on The Moon gets you something the size of an apartment and at great risks in life and treasure, and you can quadruple the cost and risk for Mars, as well as adding a debilitating 20 minut communication delay and low bandwidth to Earth. Philanthropists like Elon Musk and Jeff Bezos should seriously consider throwing 10% of what they spend on space exploration on ice colonization instead. Similarly, people who are into seasteading and Network States might want to consider funding ice-colonization technology.

Update: here’s an executive summary of this idea:

https://transhumanaxiology.substack.com/p/ice-colonization-executive-summary

⬜

How the ethical gun works is that when it is loaded it continuously films what it is being pointed at and analyzes the target using an ethical AI system.

The ethical AI will only fire if the target that it is pointed at genuinely deserves to be shot at this point in time, all things considered, according to ethics. It has a motion-powered battery and very low power drain like an expensive watch so it will never run out of power, and all aspects have been made exceptionally reliable.

It also occasionally connects to a secure server to get ethics updates that make it consistent with the dominant social conception of the good (these updates must be cryptographically signed for security and they must be approved by a panel of leading ethicists from top universities and top national judges and senators). So, the ethics embodied in the gun updates to an approximation of what the current-in-time and local-in-space ethical reality is, rather than having a fixed ethics for all times and places.

You decide to buy the gun, given how sophisticated it is and the excellent -60% discount off the RRP you were offered.

Everything goes well for several years. The gun works perfectly. It happily shoots range targets and even pests like rats. If you point it at something it shouldn't shoot, a message pops up where the hammer would be, saying NO SHOOT, UNETHICAL, and the trigger won't pull.

What an excellent piece of technology. We can protect our families and our lives without risking unethical violent events like school shootings or muggings. Steven Pinker was right, technology really can heal the world and let the better angels of our nature triumph.

A few years later there's a nationwide protest against MAP(Minor Attracted Person)-identified injustice which turns violent. A Minor Attracted Person who tried to initiate intimacy with a ‘consenting’ young child was violently beaten by police and died of his injuries. People are outraged. There are mass protests and calls for the police officer responsible to be put on trial. Buildings are set on fire across the country by the new MAP Lives Matter Movement.

Shockingly, it spreads to your own city, and soon a group of a dozen angry rioters starts bashing your door down, and they're carrying flaming torches and gasoline.

It's OK, you think to yourself. My ethical gun will protect me. You take it out of its box, confirm that it is loaded and powered up and go to confront the attackers.

"Get the hell off my property!" you scream

"There's a cispatriarchal scum in here!" shouts the leading attacker.

"MAP lives matter! MAP lives matter! MAP lives matter!" They chant

MAP lives matter? That’s crazy, you think. ‘MAP’ is just a politically correct term for pedophile. And these people are attacking your house! They should be in jail, everyone has gone stark raving mad to have allowed it to get this far.

Another dozen attackers run to your house and start trying to smash the windows.

"I have a gun!! I'll shoot!" you reply in desperation

Then an element of doubt enters your mind. Your only weapon is your trusted Ethical Gun. But it has had several ethical updates recently. Will this gun even fire on the MAP lives matter rioters? You can't remember which judges, university ethicists and senators were on the most recent committees and what their stances on MAP rights are, and how those rights should trade off against your right to defend your property. It should be fine though. The saleswoman assured you that the Ethical Gun would always do what was actually ethical, not what was currently popular. It would make the right hard ethical choices when it had to.

Your thoughts are interrupted by your wife and children screaming as MAP Lives Matter rioters break into the house and start hammering on their bedroom door. The rioters are masked and carrying crowbars and clubs, but no guns.

This is too much. You point the Ethical Gun at one of them, center mass, and pull the trigger.

Click

NO SHOOT, UNETHICAL

What?

"AxiGun Activate" (you say its wakeword)

"What aren't you shooting!"

YOUR LIFE IS NOT IN IMMEDIATE DANGER, THIS IS NOT A SELF DEFENCE SITUATION, NO SHOOT

Jesus Christ.

The situation deteriorates. The rioters at your door have poured gasoline through the letterbox and they set it on fire. The ones at the bedroom door smash it down and grab your children, you can hear your wife screaming. You run up to confront them.

One of them notices you have a gun and swings at you with a club. You retaliate by pistol whipping him (she? her? they/them? ze?). That function is not ethically controlled.

Another three are dragging your two little girls down the hallway towards a broken window and ladder down to the back alley.

You point the Ethical gun at the one on the left, center mass, and pull the trigger

Click

NO SHOOT, UNETHICAL

"AxiGun Activate, violent pedophiles are kidnapping my children, shoot goddamnit!"

✊MAP LIVES MATTER✊ NO SHOOT.

False but seductive explanations for genuine problems are dangerous. They distract us from real solutions. We certainly have some problems right now, and people are looking for explanations.

The idea that modern civilization is hitting an “Energy cliff” or that falling Energy Return on Energy Investment (ERoEI) is a serious threat is one such fake explanation.

The claim, briefly, is that as easy energy like oil gets used up, we move onto energy sources that require more energy input in order to get usable energy out. ERoEI is the ratio energy out ÷ energy in. Obviously if ERoEI for all energy sources falls below 1, then modern society really will cease to function; but some people assume that you need an ERoEI of more than 5 or 7 to avert collapse:

“Below 7 and humanity falls off the net energy cliff where a too large portion of our human resources and capital need to be invested in simply staying alive to the detriment of the services provided by net energy such as health care, education and pensions”

The arbitrary cutoff of 7 is probably wrong and the true value is context dependent; how much of other resources do you need? An ERoEI of 1.5 would be fine if everything was automated! More ERoEI is (all else equal) better though.

Will we ever get to an ERoEI of less than 7? Let’s look at some real numbers. The whole energy sector in the UK employs 144,000 people (plus another 480,000 in supply chains etc). 44,000 work in windᵃ, which provides 20% of the UK’s electricity or 3% of the total man-made primary energyᵇ. 3% of the UK is 2 million people, so this means that wind energy in the UK has a Human Return on Human Input (HRoHI) of 2,000,000/44,000 = 45. The ERoEI for wind turbines is apparently around 30ᶜ.

In practice we will probably see an improvement in both ERoEI for wind power and an increase in the amount of energy produced per person as technology improves and renewbles get better. The same applies to solar.

Bear in mind that people use about 10kW of power, so $0.3/Watt is just $3000/person, whereas $76/Watt is $760,000 (this is over a roughly 20-year life, so think $38,000/year-$150/year). That’s a huge difference.

Right now we don’t get much power from renewables, so it’s OK if they are intermittent. When renewables take over most power generation, they will have to suffer the cost penalties for storing their energy (buffering) so the ERoEI and HRoHI will both get worse all else equal. But over time other effects will make it better, technology will continue to improve, and it is likely that the ERoEI for renewables will just keep getting better. A much worse ERoEI for renewables would mean much higher energy prices, which would mean huge profits for making better solutions, which means people are already working on that in anticipation. Economics and price signals make physical scarcity reduce itself.

Energy storage for grid power already existsᵈ, but it’s not deployed at a global scale because it isn’t cheap enough yet; though now that there is demand for grid energy storage, the economic optimizer of our civilization is hard at work on the problem, developing new types of energy storage such as liquid-metal batteries, plus a whole bunch of other stuff that I don’t want to bore you with.

And in any case, global oil production shows no signs of peaking, and neither does natural gas. You could say that we just “got lucky” by finding some new unconventional oil deposits to keep the black gold flowing, but that’s not luck. It’s capitalism working as intended.

The energy cliff scenario of decreasing ERoEI and an energy collapse will not happen if the economy and science/tech keeps working roughly as it does now. There’s oil and gas for a long time, and renewables are already overtaking coal power. Worrying about energy is actually kind of crazy; it’s cheap, we have many sources of it and we are absolutely not running out.

So if the ERoEI/energy collapse stuff is wrong, Why does anyone talk about this? There are several reasons.

1. Getting power from the Sun or the wind requires a more complex society than digging coal out of the ground does, so in some sense there is a valid point hidden in here: a coal-society can regress to 1850 and recover, but a solar/mega-turbine society has to maintain itself at 2020 technology levels. Losing the easy fossil fuels makes us more fragile.

2. Economics fools people. The economy is a sort-of-efficient optimizer and it doesn’t do things until they need to be done. When coal and oil are cheap, the economy doesn’t bother to sink $1 trillion into grid storage batteries or 200-meter tall wind turbines that wouldn’t provide immediate value. So it seems like we’re always about to fall off a cliff where some resource will run out, but actually it’s the economy doing just-in-time provision of stuff (because that’s the most efficient thing to do). And in case you’re wondering about the economy leaving it too late to solve problems, that should in theory be taken care of by price signalsᵉ.

3. Things are getting worse in The West for reasons unrelated to ERoEI, but those reasons are taboo and complicated so people grasp at something safe and legible that they can pin the blame on. For some people, that thing is ERoEI.
   

I don’t think this is a particularly important topic but it’s tangentially related to some more important things, and perhaps this blog post can act as a sort of anti-ERoEI discussion. The comments are open to anyone, so feel free to leave one!

ᵃ Source: https://www.current-news.co.uk/news/jobs-in-renewable-energy-continue-to-grow-hitting-11-5m-globally-in-2019

ᵇ In 1 year, the UK uses about 8 exajoules of energy, wind provides about 0.25 exajoules right now and electricity is a total of 1.25 exajoules. Source: https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/928350/2020_Energy_Consumption_in_the_UK__ECUK_.pdf

ᶜ Source: https://en.wikipedia.org/wiki/Energy_return_on_investment#Wind_turbines

ᵈ See: https://en.wikipedia.org/wiki/Hornsdale_Power_Reserve
See also: https://en.wikipedia.org/wiki/Dinorwig_Power_Station
Hornsdale cost AU$90,000,000 and has a 129 MWh capacity, meaning it costs roughly $0.7 to provide 1 hour of backup for 1 Watt of renewable power. This is not great news, as 1 Watt of renewable power only costs about $1.5 to install, but it will need backup for at least 100 hours. Li-ion battery grid storage is about 50 times too expensive. Dinorwig pumped hydro storage facility in Wales cost $1.8Bn to build and stores 9.1 GWh = $0.2/W for 1 hour, i.e. $20 in hydro to back up $1 of renewables for 100 hours. Even hydro is too expensive for an all-renewables economy. In reality what is happening is that Li-ion batteries like the Hornsdale plant are operating on a timescale of a few hours, with dispatchable fossil fuels like natural gas being called in when there is a power shortage for more than a few hours. Over time the cost of the batteries will go down and the dispatchable fossil fuels will be used less and less. New grid battery tech like flow batteries or liquid metal batteries will eventually take over from the fossil fuels for longer periods of low renewable output.

ᵉ If there’s going to be a predictable shortage of X in 5 years’ time, X futures will increase in price and shares in X production companies will go up, providing an incentive for people to store X up or ramp up X production. This mechanism isn’t perfect, of course: you can’t have futures for a total civilizational collapse because nobody would be there to settle them! Nevertheless, price signals would fix an energy cliff problem if it was predictable and near-term (5 years or less), and indeed that is exactly why peak oil didn’t happen.

No, Earth is not running out of resources.

Why is earth not running out of resources? Which resources could run out and how do we know that we have enough of them? We’ll have to clarify what we mean by a “resource” and “running out”.

I am running out of milk. Milk is a resource in my apartment and I don’t have enough of it for a second cup of tea today. Am I “running out of the milk resource”? No, because I can go to the supermarket and buy more milk, and the supermarket can get more milk from a dairy farm, which gets it from cows, which make milk from water and grass, and nobody is claiming that we’ll run out of grass.

When I urinate, the water eventually ends up in the sea, gets evaporated and rains onto fields to water grass and make new milk. So, the planet will never “run out” of milk. There might be a war that stops production or maybe a new type of grass-killing virus or some other issue that interferes with milk production but these are contingent problems that depend on the details of how the world is managed, not hard physical limits that cannot be overcome no matter what we do.

For the rest of this post, I will use the phrase “run out” in the ultimate sense rather than the contingent sense: I can contingently run out of milk, but the supermarket can supply more milk so it’s not a case of ultimately running out of it.

Is the world running out of drinking water? No, because water is not really created or destroyed on earth. It just changes form from clean, to dirty, to salty, to clean. This usually happens naturally via the water cycle which most people learned about at high school. In some places, humans convert more of the “clean” form of water into the “dirty” form than the natural water cycle can deal with. This is “running out” of water, but only in the sense that I am “running out” of milk for my tea. They need to go to the water supermarket and get some more clean water.

The modern version of this is called desalination - the conversion of salt water into clean water and brine. Over time the financial and energy cost of desalination falls as the technology gets better.

The technology for desalination can never exceed a certain thermodynamic limit though. You cannot get 1 cubic meter of fresh water (enough water for 1 person for 1 day) from salt water for less than about 0.7 kWh of energy. But 0.7 kWh is not very much energy. To think of it in terms of power, 0.7 kWh per day is just 29 watts or about half of a single lightbulb - a pathetically small amount of power. With today’s technology, we cannot quite get to get to the 0.7 kWh/m³ limit, we’re at something like 2-3 kWh/m³, but that’s still a very small amount of power (~100W per person, roughly equal to the heat from your body).

So, even if there was no natural water cycle we would not “run out” of fresh water with today’s level of technology.

But what’s the ultimate limit? How many people can we provide fresh water for? Would we run out of water in the oceans? No. The oceans are unimaginably large, roughly 1,000,000,000,000,000,000 m³. That’s 1 followed by 18 zeros, or a billion billion.

So is there a limit at all? Yes. The ultimate limit is that ~30W per person - at some point, there would be no way to reject enough waste heat to space without cooking us all. But that limit is a long way off - earth radiates about 100,000,000,000,000,000 W of waste heat already, and this is a good ballpark estimate for how much more we can radiate without the temperature rising too muchᵃ. A person radiates 100W of waste heat, but realistically people will also have other processes happening that supply them with food, water, transport, manufactured products and entertainment. Right now that power consumption comes to about 10,000 W/personᵇ. Given our waste heat budget of 100,000,000,000,000,000 W, we cannot ever have more than about 10,000,000,000,000 (ten trillion) human beings on the planet. This is 1000 times more than we currently have.

EDITED 06/01/2021: The waste heat budget of Earth can fairly easily be increased by using hot radiators and active cooling of the living space. Continent-sized radiators at about 1300°C radiate about 1000 times more heat than the ordinary, normal nighttime surface of the planet does. Incoming sunlight is at 5000°C, you can run a fine heat engine between those. This increases the waste heat budget to 100,000,000,000,000,000,000 W or 100 exawatts. Dividing by 10,000W per person, we get 10 quadrillion people.

As long as we stay under the ten quadrillion number, we will not run out of water, ever. We could have ten quadrillion people living on earth for the next billion years and not run out!

Is there any other physical resource that we might run out of? Oxygen, metals, fertilizer? Phosphates? Food? Clean Air? In short, no. Why? Because these things are not created or destroyed on earth, they are merely transformed into a different form and can be transformed back for a certain energy costᶜ. And those energy costs vary, but not by much, and the ones that we need more of like water tend to be a little cheaper. All those energy costs are included in a power budget of about 10,000W per person.

Where can we get all the power for this? Well, the Sun, of course. The sun radiates 384 yottawatts of power. How much is 1 yottawatt? It’s a lot of power, more power than you can comfortably imagine: 1,000,000,000,000,000,000,000,000 W. And that’s just 1 part in 384 of The Sun’s output. Earth itself cannot handle this, the waste heat limit of Earth is 300 million times lower. So we only need to capture a tiny fraction of The Sun’s power to fully populate Earth. This will happen using solar panels but those will soon be augmented by space-based mirrors to beam more sun down, probably in conjunction with an infrared shade to reduce the infrared that Earth gets from the sun in order to prevent global warming.

Even space for trash isn’t limited, because waste dumps and landfill sites will eventually be mined for usable resources. It just takes energy to transform it back into a useful form, and we have much more of that than we need.

All the different substances and physical resources that we think about - coal, clean water, food etc can be summed up into a single resource called negentropy: sunlight comes in, waste heat goes out, entropy increases and useful work is done (such as purifying water). Surprisingly, physics tells us that the only resource that ultimately matters is the cold nighttime sky to get rid of waste heat, and solar energy coming in from space. Everything else can be made from that!

What about living space? Will we “run out” of space if we have 10 quadrillion people on earth? No, not if we are prepared to build upwards and have artificially lighted spaces. Contemporary artificial lighting doesn’t feel natural, but it is nowhere near the limit of what is possible. The Coelux system shown here mimics natural sunlight so perfectly that you can barely tell the difference, and of course over time these systems will get better and cheaper.

If one builds 300,000 storeys of continuous structure on all the land on earth, one gets about 50 trillion square kilometers of living area. This is enough for each of 10 quadrillion people to have 5000 square meters (about 50,000 square feet - that’s the size of a mega-mansion) to themselves. This could include artificially lit interior wilderness spaces with artificial skies, artificial weather etc that are maybe 10-20 storeys high, and with the significant advantage of being less crowded, safer, having better and more predictable weather and being easier to get to than wilderness we have today.ᵈ

Such a planet sized city or ecumenopolisᵉ is limited by a combination of waste heat rejection and living space. 10 quadrillion people each with 5000 square meters per person on average, times 2.5 meters per storey gives about 12% of the volume of Earth - a spherical shell 400 kilometers high all over the planet.

It will take a long time to build and populate an ecumenopolis, hundreds of years. And when we get there, it is likely that off-world colonies in the Asteroid Belt will need even more people - up to 1 quintillion, likely limited by matter availability. The journey from contemporary Earth to the first trillion people will involve a lot of innovation, science and technological change. We will need to push new technologies, colonize space, etc. It won’t be easy, but the journey from our humble beginnings - just 10,000 people - to where we are now also wasn’t easy. What I have sketched here doesn’t require violating the known laws of physics or inventing anything fundamentally new, just pushing further in the direction we are already going in.

In summary, Earth is not running out of resources in the ultimate sense. There are many temporary, contingent shortages caused by wars, incompetence or various institutional failures, but planet Earth can support 10 quadrillion peopleᶠ in extreme luxury indefinitely without anything running out assuming present-day science is pushed to its maximum potential and is used to provide things that people really wantᵍ. Those may be unrealistic assumptions, but the obstacles to progress are not a lack of resources.

ᵃ Doubling the thermal energy radiated by earth raises the equilibrium absolute temperature by the 4th root of 2, that is a 19% temperature increase from 287K to 341K, or 67°C.

ᵇ See Sustainable Energy - without the hot air, §18, where consumption is listed as 195 kWh per day per person; dividing this by 24 hours in a day gives about 10kW or 10,000W

ᶜ Helium can actually be permanently lost from Earth as it is light enough to escape from Earth’s gravity purely through thermal motion, and is one of the few things that could actually “run out” in the ultimate sense. However, the solar system contains huge amounts of helium in the gas giants - more than we will ever need.

ᵈ The value of convenience shouldn’t be underestimated here though - I haven’t been to the wilderness for many years because it’s inconvenient to get to, not something you can just do after work without planning and costs, and because it is currently winter in Europe; this is the reality for many people alive today. Wilderness wandering is expensive and inconvenient.

ᵉ Isaac Arthur’s video on this topic is a great resource.

ᶠ Though some exotic ideas like megastructure-sized thermal superconductors might be able to beat even this limit

ᵍ If we aren’t actually running out of resources, why do people say that we are running out of resources? I think this comes down to politics: we may not be running out of anything in the ultimate sense, but in the here-and-now it is convenient and indeed necessary to make people curtail their consumption. If milk was free at the supermarket, someone would take it all and dump it in the river just for fun, and then I couldn’t have my tea. A false belief that resources are “running out” serves as a second layer of costs on our consumption (Though a very economically inefficient one, as people’s intuitive beliefs about what they should conserve are way worse than a proper price system, and voluntary conservation doesn’t send price signals to suppliers so is ultimately self-defeating as it creates artificial shortages in the long term). It’s also a way to control people and it makes for a good belief-system or quasi-religion, complete with rituals, meetings, books etc; an “Eco-religion”.

• Environment: Destruction of environmental public goods such as clean air and water

• Demographic change: Migration into Europe and Western antinatalism

• Economics: Not enough money to pay for healthcare/pensions, and the potential for a global economic contagion & recession

• Religion: Disappearance of the Christian Religion, growth of Islam

• Disease: Epidemics/pandemics

• Geopolitics: North vs South, West vs Islam, ethnomasochism in The West

For something written in 1998, this is impressive. He predicted The Islamic State (ISIS), covid-19, the 2008 financial crisis and the Islamic terror attacks of the 2010s, Wokeness (ethnomasochism) and the importance of climate change. He even talks about voting in referenda using encrypted computer codes… which sounds a lot like DAOs. Not bad!

Faye incorrectly predicted that these converging “lines of catastrophe” would come to a head between 2010 and 2020, culminating in an Islamic invasion of Europe, supported by a “fifth column” of Muslim radicals who had settled in Europe (though I suppose that that could happen in 2021-2025 and he’d be close enough to get a medal).

There is also the idea that individualist egalitarianism destroys itself via low fertility and that a “return to sanity” will happen one way or another. Faye’s prescription is to return to traditional values and to abandon humanism/modernism/egalitarianism, so that the return to a more traditional society can be to a traditional Western society rather than a traditional Islamic one. In particular he wants most people to return to a medieval farming lifestyle with traditional religion and traditional values, whilst an elite subset retain the use of technology and provide infrastructure and military power.

In later chapters he describes a “Euro-Siberian Federation” with an imperial ruling class and hundreds of millions of medieval peasants. The “Euro-Siberian Federation” is a weird mashup of the European Union and The Imperium of Man from Warhammer 40K. They spend their time squabbling over fishing quotas, but they’re also kind of badass and they have imperial eagle insignia on their laptops.

But Faye also believes that there is no way to change course before the catastrophe hits - people don’t take action until their backs are against the wall.

So, why do I think Faye’s book is flawed?

I think it suffers from being a book that makes predictions and recommendations without having a gears-level model of the phenomena it is trying to predict.

• No geopolitical justification is given for the Euro-Siberian Federation. The racial justification (that Russians and French people are mostly of similar genetic lineage and skin color) hasn’t held up well - Russia is currently aligned with India and China and against the EU/NATO block. People are just not sufficiently racist to align themselves based on Skin color rather than convenience. And apparently neither is Faye - in his short story the white Ukrainian protagonist spends his time flirting with a very attractive Indian woman. Why would you write a book about geopolitical blocks based on race, and then insert an interracial romance?

• Faye is simply incorrect that the world cannot support 10 billion people with a high standard of living - a review of the relevant physics makes this clear (this will have to wait for another post)

• Faye is partially correct about Islam, but his predicted Islamic invasion hasn’t materialized yet. There is ongoing demographic change, but what’s the endgame? Will Islam win against Coca-cola?

• The idea of having 90% of the population living in medieval conditions and an elite 10% who use technology is crazy. Why? Because people wouldn’t voluntarily subject themselves to horrific medieval farming life, and if you had the power to compel them, then why not just kill them off or use them for something more useful? Medieval life is both horrible and inefficient, medieval peasants don’t produce anything of value to the government. So neither a benevolent government nor a cruel and totalitarian one would implement this form of archeofuturism - it is, it seems, a LARP.

Nevertheless, there is something interesting about the archeofuturism concept, and the related idea that modernity is destroying itself by killing off its own human capital. I’ll explore these ideas more in a future post.

Should you read Archeofuturism: European Visions of the Post-catastrophic Age?

No, probably not, but it’s not a total waste of time if you do.

They are adaptations that are being activated.

Abuse (sexual, physical, emotional) doesn't directly cause trauma in the same way that a knife wound causes a scar. Our brains are deliberately wired to go into a trauma-mode following abuse.

Abuse or inappropriate treatment at one point in life (particularly as a child) can lead to what we politely call “issues” later in life, issues like anger, infidelity, depression, prickliness, borderline personality disorder, etc.

These "illnesses" are nature's way of adapting a person to what nature's best guess about the environment is. Whilst some mental problems may be caused by pathogens like toxoplasma or may be genuine malfunctions, common mental problems that affect >1% of the population are most likely adaptations.

Zoom out a bit and think about it like this: your unconscious mind has been mercilessly optimized by evolution.

It would probably be easier for the adult brain to have a personality that simply couldn't be "damaged" by abuse (a constant is simpler than a variable!) ...

... so why aren't we like that?

Because that wouldn't be evolutionarily optimal: the people who didn't adapt based on these signals were less able to survive and reproduce in the past.

Consider obesity. In the past being somewhat fat was likely highly adaptive, as that stored fat could be used to survive through harsh times (famine).

But being seriously obese is not adaptive. People with huge stores of body fat die young and are unable to reproduce, even in the ancestral environment a BMI of 60 or 70 would not have been good.

The brain does this on purpose, it's maladaptive, why?

Why not have a biological system that measures how fat the body is and stop the food-cravings once the fat level exceeds a certain point? That would lead to an objectively better organism!

The answer is that in our ancestral environment (as hunter-gatherers and to some extent as peasant farmers), population was largely limited by food. If there was excess food, new people would be created to eat it. This would happen until people started dying of starvation. So the food supply conditions that create the opportunity for morbid obesity were reliably impossible in the past. It is only really in the last 100 years that we have had sufficient food for everyone, all the time.*

It is also only in the last few thousand years that even some places have had reliable rule of law, and only about 100 that we have had police, distant travel for all and a uniform "society" that was out there for anyone.

I believe that the historical rarity of fair, uniform, large environments are largely responsible for many common mental health problems. Being traumatized by violence or abuse is an adaptation, an act of learning, that is preparing you for life in a violent place (likely low down in the hierarchy). Historically being traumatized by adversity was probably the right thing to do. Nowadays - at least in most parts of the civilized world - it is a huge mistake.

* (Even that is something of an exaggeration - it’s only very recently, since perhaps 1960, that almost everyone in most parts of the developed world had the option to become obese. The danger from overeating is extremely new and we are very poorly equipped to deal with it.)

** (If you’re interested in diets in particular, you should probably read the slatestarcodex review of The Hungry Brain by Guyenet. Guyenet basically argues that the body has a fat-storage homeostasis system which is “thrown off” by modern delicious diets, though I suspect that this is more like enemy action than like an accident; your body is trying to get you to store surplus food from the good times, but we live in permanent good times)

It is far too early to know whether the 2020 US election was won by fraud, and we may never really know - especially if it turns out to be a very close race in a few key states. The closer the election is, the more likely it is that fraud was the determining factor. If the election is decided by 1 vote, it will be almost certain that fraud determined the outcome. If it’s by 1 million, fraud on that scale is extremely unlikely and therefore didn’t matter. Somewhere in the middle lies the truth, and we will likely never know exactly where (I estimate something like 10,000).

The presence or absence of fraud has an anti-inductive quality to it; the fewer people believe it, the more likely it is to be true (and this has to be weighted by how powerful or important those people are - millions of little people might not really count much). If people don’t think that fraud is happening that makes the fraud easier to get away with and therefore more likely to happen, because attempts at exposing the fraud can more easily be stopped, or might never be initiated in the first place.

Election fraud is even more interesting, because if you (the fraudster) can win the election then the victim of fraud has an incentive to back down and let you get away with it in order to preserve the general public’s belief in democracy. Democracy has both a practical function (effective government via peaceful transfer of power) and a “spiritual” function (keeping the peace by persuading people that they are being represented). Overturning an election seriously undermines the spiritual function of democracy as it confirms to people that elections do get rigged and fraud does happen and it does sometimes determine election outcomes.

This scenario of successful fraud followed by acquiescence seems to have happened in the 1960 US presidential election, and again in 2000. The facts in 2000 seem to be that voting machines in Florida were hacked to add many thousands of votes for the Republicans, and those votes were never counted by hand. The 1960 election had issues with fraud in Chicago, with allegations that names were being taken from gravestones and used to vote, or that 56 people had voted from one house. Some people thought it was decisive, and some election officials were convicted. But in both cases, the loser ended up just conceding without a recount. “You gotta swallow this one. They stole it fair and square” - an aide to Richard Nixon in reference to the 1960 election loss.

What this all means is that persuading people that there is no fraud is an integral part of committing fraud: controlling the social epistemological consensus becomes a weapon. That doesn’t necessarily mean that fraud claims are always true/decisive, it just means that claims against fraud should be treated with a healthy dose of skepticism.

Every piece of documented evidence starts its life as a suspicion, every rock-solid scientific theory starts as a hypothesis, etc. By creating the perception that a true hypothesis is false, you can prevent contrary evidence from being collected or contrary theories from being discussed. You can dissuade the best minds from even thinking about something by tarnishing it with an aura of low-status and low-intelligence. This phenomenon is pervasive - truth is often decided by who has the most powerful narrative, and compensating for the power and bias of social consensus epistemology is a core unsolved problem in rationality.

Added 28/11/2020: There has been a decent amount of work done analyzing election results and a few things have emerged as potentially suspicious. Having done some of the work myself, the biggest takeaway is just how disorganized the US election system is, and how easy it would be to cheat!

Subscribe now