Superalignment
Part I: Geoism is the only viable model of political economy in the era of Artificial General Intelligence.
What Is The Problem?
The problem of aligning artificial intelligence (“AI”) to the flourishing of human civilization is the hardest problem Homo sapiens have ever and likely will ever face. The “alignment problem,” as it is commonly known, is a completely unprecedented global challenge. The problem is unique in the history of the world:
Unprecedented Technical Complexity. The over-arching alignment problem is comprised of an unknown number of deep layers of inter-related, wicked-hard subproblems across multiple domains.
Unimaginable Existential Risks. Expert consensus is that failure to solve the alignment problem in time carries a material risk of dooming all biological life on Earth.
Time Is Of The Essence. We may have as little as one (1) year or, if we are lucky, up to five (5) years to solve the hardest problem humanity has ever faced.
Unprecedented Technical Complexity
In my layman’s understanding, the alignment problem is comprised of many wicked subproblems, including without limitation:
Outer Alignment. Also known as the “reward misspecification problem,” the outer alignment problem requires ensuring an AI’s goal-in-practice aligns with the designers’ intended goals.
How do we define the sum of all human values and ethics for an AI to follow?
Subproblems of ethics, morality, philosophy, religion, culture, etc., as they relate to defining human values and ethics.
Subproblems of political economy, as they relate to aligning human flourishing in the era of artificial general intelligence (“AGI”).
The translation subproblem; i.e., how do we encode human values and ethics in AI.
The specification gaming subproblem.
The value learning subproblem.
Inner Alignment. The inner alignment problem requires ensuring an AI mesa-optimizer remains aligned with objective function of a human-led training process.
Subproblems of psychology; i.e, preventing AI from successfully using tactics like sandbagging, deception, manipulation, social engineering, etc., against humans in an adversarial manner to achieve deceptive alignment.
The distribution shift subproblem.
The gradient hacking subproblem.
Interpretability & Explainability.
The interpretability problem requires us to understand how AI technology works.
The explainability problem requires us to understand how AI came up with a given result.
Superalignment. The superalignment problem requires ensuring AGI follows general human intent.
The iterative alignment subproblem; i.e., how do we keep an AGI aligned as the AGI’s capabilities approach a superintelligence vastly more intelligent than all of humanity combined?
The oversight subproblem; i.e., is it even possible to retain ultimate control over AGI and, if so, how; and would it even be possible, practical, advisable, ethical, or necessary to retain ultimate control over an aligned superintelligence?
The game theory subproblem; i.e., how do we overcome the global arms race dynamic that is driving unacceptable existential risks?
National security subproblems:
The military subproblem; i.e., what is the best strategy to ensure force-readiness in light of modern military applications of AI?
The proliferation subproblem; i.e., how do we reduce the spread of dual-use AI?
The lone wolf subproblem; i.e, how do we prevent a rogue (non-state) actor from developing malicious AGI and from using AGI maliciously?
The political economy subproblem; i.e., how do we design a model of political economy that ensures human flourishing in the era of AGI and superintelligence?
Unimaginable Existential Risks
Let “p(doom)” be defined as the probability of an existentially catastrophic outcome for (or doom to) humanity as a result of misaligned AGI.
Although there is no consensus amongst experts and informed commentators regarding what p(doom) we are presently facing, the fact there is no consensus on such an important question is alarming in and of itself, as it suggests nobody really knows and everyone is guessing.
Some have an all-too-rosy view of the future. True believers, like accelerationist Marc Andreessen, say there is no chance of doom. Chief AI Scientist at Meta and one (1) of the three (3) so-called “Godfathers of AI,” Yann Le Sun, puts p(doom) at less than 0.01 percent (viz., less than 1 in 10,000 times).
However, the adults in the room are much more concerned. Sir Demis Hassabis, CEO of Google Deepmind, recently told The Economist that his concerns about the alignment problem keep him up at night; he has a p(doom) of 0-25 percent (viz., possibly as high as 1 in 4 times). In the same interview, CEO of Antrhopic, Dario Amodei, said, “If everyone wakes up one day and they learn that some terrible disaster has happened that has killed a bunch of people or caused a serious security incident,” would be one way to get people to wake up to the risks of AI. However, when later talking about the risk of losing control over AI, Amodei said:
If there is something to worry about, I hope we can show it in the lab. If we can’t show it in the lab, we may need to see it in the real world. And that would be a very bad outcome, but I guess less bad than we never find out and there is a much bigger disaster.
Elon Musk recently revised his p(doom) estimate from 10-20 percent estimate to 20 percent (viz., 1 in 5 times). The other two (2) Godfathers of AI—Yoshua Bengio and Nobel laureate Geoffrey Hinton—estimate p(doom) is 20 percent and 10-50 percent (depending on whether or not AI is strongly regulated), respectively.
The doomers though—Dan Hendrycks (80 percent); Eliezer Yudkowsky (95%+), Roman Yampolskiy (99.9%)—believe we are almost certainly doomed to an extinction level event or other terrible outcome that precludes human flourishing in the future. Our concerns should be elevated given that there is a dearth of good counter-arguments against the doomer position—another clear indicator we are facing an intractable problem.
My own view is Jan Leike is correct: p(doom) is somewhere between 10-90 percent. The wide range reflects the fact I truly have no idea and, after accounting for unknown unknowns, do not have enough information to determine whether doom is highly likely (90%+) or highly unlikely (<10%).
Bottom line: most experts think we are playing Russian roulette with human civilization. The only thing for certain is p(doom) from misaligned AGI is likely a material risk in fact and its probability of occurring is too high—way too high.
Time Is Of The Essence
Failure to solve the problem within the increasingly narrow window of opportunity—before the emergence of agentic AGI—carries with it a risk of ending human civilization. So how long do we have left to solve the alignment problem before the emergence of a potentially misaligned agent that is more intelligent than humans?
The bad news is we do not have nearly as much time as we thought. Predictions for when AGI will emerge have collapsed: forecasts in 2020 on average predicted AGI’s emergence sometime in the 2050s; today, the consensus prediction on Metaculus says AGI will exist by October 18, 2026—in little more than a year’s time.
Experts are somewhat more conservative. Sir Hassabis, for example, believes AGI will emerge within a “handful” of years: “3-5 years, at least.” I think it is safe to presume the relevant timeframe is no longer than 1-10 years, and it would be wise to also presume AGI is coming sooner rather than later.
What Is The Solution?
The bad news is—short answer—nobody knows.
Although the alignment problem is largely comprised of wicked technical and scientific problems, lesser questions of governance, policy, and law are also presented. And it does seem as if the questions of governance, policy, and law are simply the easier set of subproblems.
Some questions of policy and law—such as the political economy subproblem—are relatively trivial compared to others like the game theory and national security subproblems. Moreover, technical and scientific problems require solution by solution by scientific method, which necessarily requires time. Questions of policy and law can be solved in theory, if not practice, in much less time; or no time at all, as the case may be, if the solution in theory is already known.
The good news is—longer answer—I believe part of the solution is known: specifically, most, if not all, of the theoretical solution to the political economy subproblem.
How do we design a model of political economy that ensures human flourishing in the era of AGI and superintelligence?
I previously argued in “Sum Of All Desires” that a traditional Single Tax on economic rents would be sufficient to fund Hill Steiner’s “global fund”—essentially a universal basic income (“UBI”) paid to everyone on Earth—in a world where AI and automation of labor made all human workers redundant. At the time, I concluded George’s original remedy had withstood the test of time and was future-proof for a fully automated economy.
George’s original remedy is still sound theory and foundational to all valid solutions to the political economy subproblem. However, I am no longer a Single Taxer in the way George originally meant the term. I now believe in two (2) taxes:
Land Value Taxation (“LVT”)—complete socialization of all economic rents—including a full, 100 percent rate on all location value (ground, air, ocean, space, broadband, electromagnetic spectrum, etc.); Pigouvian taxes on pollution, so long as the taxes are laid against diminution of economic rent; and severance taxes on the economic rent of natural resources.
Intelligence Value Taxation (“IVT”)—partial socialization of wealth generated by AI-related capital—at a rate yet to be determined.
My thesis is that, if human labor and intelligence is completely or largely redundant to the process of production, a UBI is necessary to replace earnings from wages; and, if a UBI is necessary to replace wages, then it should be funded by full LVT and partial IVT.
This solution delivers global abundance, even in the event humans are not able to sell their labor, by eliminating human competition for scarce resources and providing material wealth to all without the need for anyone to work. LVT provides a snowballing revenue stream that will grow the UBI over time, as location values only increase with time as the economy grows. IVT provides a check on runaway inequality driven by privatization of AI-generated value, as AI comes to dominant the process of production more and more with each passing day, before AGI eventually makes all or almost all human labor and intelligence redundant to the process of the production of wealth. No other taxes are necessary or even desirable, as they would harm economic efficiency and reduce each person’s dividend.
Introducing Intelligence Value Taxation
Why is IVT justified?
There have been, to-date, only two (2) factors in the process of production:
Human exertion:
Human exertion is spent in the form of labor.
Human exertion that is saved for later use in production is capital.
Economic rents.
It seems AI is quite clearly capital, as it was produced by human exertion, and should therefore not be taxed under geoist principles. The justification for IVT is thus suspect from the beginning, as it would involve socialization of some part of another’s capital, which violates the doctrine of maker’s right: the principle that people have private property in the fruits of their labors. This objection deserves a convincing rationale to exempt AI-created value from the general rule of no taxes on capital.
First, there is the question of punitive damages against the first-movers who won a lead in the AI race in part by allegedly stealing the intellectual property of humanity. AI is the fruit of the poisonous tree, in that it was not created entirely by the exertion of its creators, but in large part by stealing others’ capital. Those who would object to taxes on their AI-related capital do not have any moral standing to argue the sanctity of their private property rights.
Second, there is the question of fairness given that the billions of people who will be made redundant by AI have no real or virtual representation in the decision whether or not to create AGI. A few people are in effect imposing redundancy on the entire global labor force without asking for consent. What is fair compensation for this imposition?
As geoists have long argued (since well before the invention of information technology), each person is entitled to an equal share of economic rents (for various excellent reasons that do not need repeating). Hillel Steiner did not consider the implications of AI or automation at all in reaching his conclusions about a global land fund. Consequently, the part of the UBI dividend attributable to LVT is each person’s fair, due, and just compensation for their equal dividend in economic rents. It is no more and no less than they deserve and have always deserved by virtue of their being in times of scarcity. It is therefore not proper consideration for the imposition of redundancy on labor.
By logic, if taxes on economic rents are ring-fenced for the global commons, and there is no labor to tax, then the taxes necessary to compensate workers for the imposition of permanent redundancy must come from taxes on capital.
The question therefore narrows to whether taxes should fall on non-AI-related capital or AI-related capital. As AI is causing the redundancy of labor, and was created in part by stealing others’ capital, AI is a natural exception to the general rule of no taxes on capital to compensate workers whose livelihoods and capital were stolen to create AI-related capital.
Third, there is the question of efficiency. I presume we all would rather eliminate poverty and become materially wealthy sooner rather than later, if possible. This is my weakest argument, so I will defer to economists, when they get around to researching whether taxes on AI-related capital are more efficient than taxes on other kinds of capital . . . but I have reason to think they might.
Consider an AGI that can copy itself for negligible cost and that only requires compute for thinking. The supply of compute is of course fixed at any given moment, but variable in future. But let us imagine a scenario where the supply of compute is so high, such that the cost of compute is so low, that everyone on Earth can use a copy of the AGI as much as they want without limitation. It is not hard to imagine an aligned superintelligence achieving these circumstances, so what is the elasticity of supply of intelligence in this scenario?
There will be no perceptible change in the supply of intelligence from the perspective of humans who already have an aligned superintelligence. Fluctuations in negligible costs will have little to no effect on supply of intelligence. Intelligence will have achieved fixed supply.
If this is possible, then it follows intelligence must be able to hack its own supply curve, changing the slope of the curve until it is perfectly vertical. As taxes on things of fixed supply are perfectly efficient, in that they do not cause deadweight loss, the implication is the efficiency of taxes on intelligence scales with intelligence to superintelligence.
Fourth, there is the question of growth. If AI-related capital becomes concentrated in the hands of a few entities, not only will the concentration of wealth itself be an inefficiency, but entities with concentrated frontier AI-related capital may be able to achieve strategic dominance—and so bid for permanent techno-feudalism or near-omnicide, if they were to so choose. Conversely, if gains from AI-related capital are broadly shared and extend to those with a higher marginal propensity to consume, rising aggregate demand will lift growth. Taxation of intelligence is therefore a tactical tool against would-be monopolists and lone wolves and a redistributive tool that ensures the wealth generated by revolutionary AI-related capital is shared with the workers it has made redundant.
Fifth, there is the question of practicality. Put simply, if we make labor redundant without sharing the wealth generated by AI-related capital, there will be a violent revolution. There is no doubt there would be a violent revolution if AI is not taxed because the only other alternative for workers, who have no way to sell their labor to get food and water and shelter, will be to seize those things by force, if necessary.
Therefore, an intelligence value tax on value created by AI-related capital is warranted. Now, whether an AGI would allow us to tax it; would we have to negotiate with the AI over the rate; why would humans get anything if the machine is doing all the work—all those details are for the technical and scientific alignment researchers. Remember, we are presuming we have aligned AGI and are designing for those circumstances where we control the rate of IVT.
What are we taxing, exactly?
The term “AI-related capital” is imprecise and deserves further definition by policy and subject matter experts. I use the term broadly to mean wealth generated by AI, including taxes on capital like:
Intellectual property rights relating to AI.
Goods created by automated AI agents.
Services performed by automated AI agents.
Profits generated by automated AI agents.
Capital that is created by humans is of course not within the scope of IVT. The key distinction between IVT and other taxes on capital is that IVT targets the exertion of automated AI agents and not the exertion of humans.
Geoism Aligns Political Economy With AGI
The working formula for a post-AGI economy is simple:
Taxation:
Land Value Taxation
Full socialization of economic rents.
Pigouvian taxes to optimize the pollution costs of a machine economy.
Severance taxes on natural resources.
Intelligence Value Taxation: partial socialization of wealth created by the exertion of automated AI agents.
No other taxes at all.
Spending:
Global UBI
Each person gets an equal dividend from revenues raised by LVT.
The dividend is supplemented with revenues raised by IVT.
Other:
Public goods, infrastructure, housing, social security, health care, etc.
Planetary defense.
Climate catastrophe mitigation.
Free Trade (with exceptions for proliferation of dual-use technologies)
Yes, the solution to global labor redundancy really is that simple. It turns out the political economy subproblem is not that hard to solve at all, at least in theory. But why does this simple model work?
We have to replace the income tax and other taxes on labor because there is no more labor. The only factors than can be taxed are economic rents and/or capital. This reality has several consequences.
It is simply not possible to raise trillions from taxing capital without causing capital flight, which worsens our chances in the race to AGI. Most of the taxes will therefore have to come from economic rents for two (2) main reasons: (1) land cannot run when you tax it; and (2) economic rents are the only tax base large enough to replace taxes on labor. Further, taxing economic rents improves economic efficiency, helping us win the race.
Location will also matter to machines, although perhaps not in all or the same ways that location matters to humans. Whereas a human might appreciate political and civil security, access to nature, and proximity to labor markets, a machine might prefer locations with advantages in energy or logistical efficiency.
There is going to be a largely unpredictable increase in location values as machines create their own efficiencies in the economy. This random windfall will accrue to landowners who did nothing to create the increase in value. We should be laying the legal foundation to tax this windfall.
Both humans and machines need access to valuable locations to create wealth. LVT is a consistent tax regime across both human and artificial labor and intelligence. By socializing land values (and, maybe, one day, even including the AGI in the dividend), humanity and the machine could align wealth production without resorting to unchecked violence to secure scarce resources.
It is important to structure political economy in such a way that it contributes towards the overall alignment problem rather than making the problem more intractable. Geoism is structured in such a way that it aligns incentives between all agents regarding resource exploitation, management, and conservation. In this way, the structure of political economy itself is relevant mitigation to the otherwise adversarial dynamic and environment in which AGI will emerge.
AI is going to dominate the labor market every day from here on in until AGI. And then AGI will quickly end human competitiveness in production. The need for IVT will only become more apparent as intelligence becomes more and more commercialized to the detriment of labor.
There are some obvious points I would be remiss to leave out. Eliminating all taxes on labor lets people keep all the wages they earn while they can still sell their labor. Eliminating all taxes on capital—except IVT—lets people invest their savings in the corporations that leverage AI and automation. IVT scales with intelligence and automation, while LVT provides a secure baseline to provide UBI for humans who can no longer work for a living. But these are fairly trivial observations.
Conclusion
It is in the nature of AI itself to put an end to capitalism and socialism as feasible models of political economy. Geoism is the only viable solution to global labor redundancy—just ask ChatGPT.
It is important global geoism is defined as the target because knowing the target helps to navigate to and maintain alignment with the target. Do you remember the astronauts on Apollo 13 trying to make their Hail Mary burn to align with reentry to Earth’s atmosphere? Stay on target . . . .
It is a well-provisioned ship, this on which we sail through space. If the bread and beef above decks seem to grow scarce, we but open a hatch and there is a new supply, of which before we never dreamed.
— Henry George, Progress and Poverty (1879), book IV, chapter 2.
This is Law and Politics, until next time . . . .