AI and the Future Economy

As LLMs continue to advance, it is becoming increasingly certain that AI will transform not only individual professions but the economy writ large. However, the extent and nature of this change is the subject of much disagreement. Most disagreements I have seen boil down to different assumptions about how far AI will progress, and people often end up talking past each other when arguing about the downstream effects of AI on the economy. In this post I would like to make these assumptions explicit and explore possible futures under each.

Broadly speaking, I think there are 5 potential scenarios for medium-term AI progress:

1. Low Cognitive - AI progress is mostly constrained to pure information processing and does not surpass the most skilled humans at many important tasks
2. Low Cognitive/Low Robotic - Robotics become significantly cheaper and more versatile, but skilled humans remain superior for complex manual tasks
3. High Cognitive - AI progress in cognitive tasks exceeds human cognitive capabilities, but robotics lag behind somewhat
4. High Cognitive/High Robotic - AI surpasses humans in virtually every task
5. ASI - The much feared/anticipated "Superintelligence", which would render all of this outside of our control

Some people believe that progress in cognitive tasks will necessarily lead to improvements in robotics, and in a limited sense this is true: for tasks that are already performed by machines, we can expect cognitive capabilities to translate to strong performance on these tasks. The most obvious example is self-driving cars, where humans are already limited to providing control inputs. However, most tasks are not like this, including virtually all blue-collar work (even to an extent trucking and delivery, which involves a lot of loading, unloading and hand-delivery work) and much service work (basically anything besides customer service and cashiers).

The challenge in automating these tasks is that human bodies are actually extremely versatile and robust, to a degree we've been unable to replicate in robotics to date. Even our best robots do not have nearly the number of joints that we do, and they require much cleaner environments or much more frequent maintenance to keep working. These issues can be mitigated by building robots for specific tasks and environments, but this is only economical for the most repetitive of tasks such as assembly line work.

Low Cognitive

This scenario is basically what I described in AI and the Future of Work. I still believe this is the most likely scenario, for the reasons I discuss in that post. However, as LLM progress has not clearly slowed down yet^[0] I am gradually adjusting my estimates of this scenario down.

Going beyond the impact on specific fields, the "Low Cognitive" scenario would exacerbate a trend we've already seen in advanced economies, accelerating current issues around inequality and socioeconomic precarity:

• Increased income inequality
• Increased returns to capital vs labor
• More education/training required for many jobs
• Real prices for information-based goods and services would fall drastically
• Real prices for raw materials would rise somewhat
• Real prices for manufactured goods would fall modestly on average, but with high variance
• Real prices for scarce or positional goods such as land would rise
• Real prices for physical services would rise

The biggest change to current trends you would see is a large shift from white-collar to blue-collar and service work, which would introduce an additional degree of instability as many people would need to restart their careers.

Low Cognitive/Low Robotic

In this scenario, progress in robotics reaches the point where a significant fraction of previously human manual labor can be replaced with robots at significantly lower cost than current wages, but there are still a significant number of tasks where humans are superior. Tasks that are likely to be taken over by robots are ones that feature constant or similar environments and a constrained set of required motions. Examples of this include:

• Restaurant work
• Long-haul trucking (not including loading/unloading)

There are also tasks that would be hard for robots to perform as currently performed, but where similar tasks that are more amenable to robotic labor could serve the same purpose — such as performing deliveries of highly standardized packages using drones to side-step navigating obstacles on the ground.

Most likely the tasks that humans outperform robots in will be tasks that take place in a variety of environments and require significant dexterity (in ways that change from instance to instance). Examples of this include most repair and maintenance work such as plumbing.

So what does the economy look like in this world? Compared to the previous scenario, we would see many of the blue-collar and especially service jobs disappear, and for many currently-working people there may be very few jobs in which they can outcompete machines. Adapting this would require governments to:

• Provide more education and training opportunities in order to maximize the % of employable people
• Implement massive wealth redistribution to avoid a permanent indigent underclass

They will also face downstream social challenges that will require political adaptation:

• Foster social engagement, meaning and fulfillment for the unemployable
• Curb the power of the hyper-wealthy in order to maintain democracy and resist external pressures from countries that do not

I believe these are surmountable challenges, and could lead to significant improvements in the average person's life, but they are daunting.

High Cognitive

In this scenario, progress in robotics stalls out around the same level as the previous scenario but progress in cognitive tasks reaches a point where almost no human can make useful contributions.

Before discussing the implications of this scenario, it is worth examining whether it is even possible. Many people believe that AI with superhuman cognitive abilities will necessarily lead to either the "High Robotic" or even "Eudomonic" scenarios. However, this belief (usually implicitly) relies on several assumptions:

• Novel research is not significantly harder than other human activities
• Massive advances in robotics or bioengineering are not vastly beyond the current research frontier

Alternatively, some people believe that the "High Cognitive" scenario will lead to the more exotic scenarios by way of recursively self-improving AI, instead assuming:

• Novel research in AI is not significantly harder than other human activities
• We are not near the long-term practical limit of human-driven AI research
• There is either no practical limit or a very high limit on "intelligence" as we understand it

I don't have a strong view on any of these assumptions except that they are not as obvious as AI maximalists assume. Everyone agrees that novel research is hard and that the greatest mathematicians and scientists of the last few centuries were very smart. Some people take a very reductionist, 1-dimensional view of intelligence and believe that, while usually much smarter than average, these people were not truly outliers and consequently an AI that outperforms 99.99% of all humans on all cognitive benchmarks could have replaced them. But we do not know this, and it is unclear how we ever could, short of inventing a time machine and killing all of our history's great scientists. Even if it is true, it is possible that the 0.01% of tasks humans are best at relative to AI are precisely "coming up with truly novel ideas".

In this scenario, we end up with essentially separate human and AI economies, with the human economy dominated by the physical services that AI cannot provide. AIs might still be owned by humans, but growth in the AI economy would be almost entirely self-sufficient. Humans would consume goods and services produced by AI and perform services for each other, but would contribute very little back to the AI economy.

High Cognitive/High Robotic

In this scenario, humans are effectively no longer part of the economy. Economic growth likely accelerates massively, but the amount allocated to humans will depend on political structures and the nature of the AIs rather than having any meaningful relationship to the total size of the economy.

From a purely economic perspective, it would be similar the creation of a superintelligence, just slower and less fantastical.

0. ^

   I think there's an argument to be made that progress has slowed down since the o1 release, as most subsequent releases have seemed like smaller improvements and come with much higher costs.