AI economics brief

This is a very brief summary of economic and social implications of continued development of artificial general intelligence. Contrary to most other articles on this topic, I am focusing on the big picture, long-term outcomes, game rules, and endgame, all in as few words as possible.

First off, don't believe the skeptics. Language models like ChatGPT are indeed AGI (Artificial General Intelligence). AGI is not magic. It's not (necessarily) conscious or emotional. It's simply a machine that can perform an unbounded variety of tasks at a level comparable to humans. Language models undoubtedly satisfy this definition of AGI.

Growth constraints

Looking at AI benchmarks done over the last 10 years, we can infer there are only two rules for developing artificial intelligence. Firstly, neural networks (almost) always beat other AI algorithms. Secondly, larger networks (almost) always beat smaller networks, regardless of network architecture. There is no silver bullet. Consequently, there will be no revolutions, only gradual improvement.

Although the AGI problem is solved in principle, there's still a mountain of little problems that must be solved before AI replaces majority of human labor. These problems are diverse, but only three classes of problems fundamentally limit maturation of AI in the long term: hardware, training data, and adoption.

• Hardware: GPT-4, including the integrated DALL-E 3, has memory capacity that is equivalent to brain of a single mouse. It runs on large, expensive, and hot servers only because contemporary hardware is very inefficient compared to biological brains. Compute power available to AI will grow somewhat with the money flowing into AI business, but catching up with (and possibly surpassing) biological brains requires development of new hardware. Hardware improves slowly (a few percents per year), so change by several orders of magnitude will take decades.
• Training data: Current top machine learning models are already trained on everything, on the entire Internet. This poses a challenge, because ML models are essentially compiled training datasets. They cannot improve without being fed bigger and better datasets. Intellectual property owners took notice and they are now locking down their databases using both legal and technical means. Corporate rules, privacy concerns, and government regulation prevent access to most data in the world. Once acquired, datasets still need cleaning, filtering, and transformation, all of which are labor-intensive. Dataset management will be the main source of employment and probably the largest expense in AI business. Even once artificial intelligence itself becomes a viable way to manage and expand datasets, data acquisition will be limited by the cost of real-world validation.
• Adoption: Integration of AI in industry has to account for long obsolescence cycles. Capital is expensive. A factory will keep old machinery for 50 years and upgrade only after continued operation becomes economically unsustainable. Automation, including AI deployments, requires workflow changes. This is hard even in business environment where employees can be ordered to use a new tool. Consumers learn to use their favorite products when young and then ideally stick to them for the rest of their lives, so about 50 years. Initial wave of adoption will take decades, perhaps a century. Full exploitation of AI potential may require several adoption cycles spanning centuries.

Using AI for industrial and home automation additionally depends on robotics, which has its own performance, cost, and growth constraints. Even though machines have outperformed humans in a lot of specialized tasks, human body is still far from obsolete when it comes to general labor.

Hoarding

Natural intelligence (brains) is a valuable natural resource, which is fairly evenly distributed among humans and it cannot be removed from the individual. This gives people bargaining power, which is probably behind past bargaining successes like freedom, democracy, and worker rights. Artificial intelligence upsets the power balance, because it can be hoarded and controlled by the rich and powerful while at the same time devaluing natural intelligence of ordinary people.

This dynamic will produce a number of surprising winners:

• Land owners: As intelligence becomes a commodity and its price drops, natural resources will gain in relative value. Up until now, land was useless without workers to exploit it. But with AI and robots, land owners will be able to capture full value of the land without having to share anything with workers.
• Governments: While artificial intelligence will start as an unregulated mess like computers and the Internet before it, governments will eventually have the upper hand, because the nature of the technology is to mediate interactions between people via something centralized and thus easy to control.
• Dictators: Artificial intelligence, like any other innovation, will first favor democracies, which will develop and exploit it faster. The benefit will however gradually shift toward autocracies, which will use it to better control the population, to run an efficient economy without the threat of unrest from the workers, and to wage wars that are limited only by available natural resources.

Hoarding of artificial intelligence and its benefits will happen simultaneously with increases in overall economic efficiency, which will mask deteriorating position of ordinary people for a while.

Endgame

When machines were first introduced into the economy, lost jobs were regained via economic growth and shift to skilled labor. While this process will repeat to some degree with artificial intelligence, it will not fully compensate for lost jobs. Growth will be limited by scarcity of natural resources and there is no skill beyond general intelligence. Cheap and universal artificial intelligence and robotics will inevitably render most people uncompetitive.

Without intervention, the most likely course of events is that of evolution and natural selection. Humans will be forced to compete with machines and artificial intelligence. As artificial intelligence and robots improve, people will be forced to work for less every time. The technology will eventually improve so much that people will be unable to compete and they will be resource-starved, pushed to the fringes of the society, and eventually driven extinct. Standard evolution.

Evolution nevertheless assumes pure competition. People can use power to stay afloat in environment, in which they wouldn't be otherwise competitive. In a world of abundant artificial intelligence, natural intelligence is worthless and the only source of power is in ownership of natural resources. The ideal way to go about it is for government to own all natural resources on its territory, rent them to the highest bidder, and then divide revenue evenly among all citizens, thus closing the cycle of a utopian fully automated economy. This will disadvantage resource-poor countries, but in the long run, population size will match country's resources, so living standards should equalize everywhere even without any global government coordinating the process.

Collective resource ownership however does sound unrealistically ideal. Most people don't even realize there is a problem, let alone take action to manage it. Taking any action assumes functioning democracy that predominantly cares about well-being of citizens. Since citizens now have much smaller bargaining power, other mechanisms must exist to keep the country democratic and they must be stable in the long run. For government to be able to afford the scheme, it must not be in a resource-draining total war, but total war is exactly what artificial intelligence enables and encourages. That's a lot of assumptions, which is why my default prediction is evolution with eventual extinction of mankind.