The AI Industrial Explosion — Part 5: Given AGI, automating physical production is probably not that hard

In this series so far we have investigated the consequences of automating labor with advanced AI systems and robotics. Part 1 found that physical growth would be unprecedentedly fast even if production technologies otherwise remain stagnant, and Parts 3 and 4 showed that even faster growth is possible with process and technological changes.

But all of this took for granted that labor could be automated. In this part I ask whether, given AGI, we would know what robots and machines to build in place of human labor. Manufacturing them is not the obstacle. In Part 2 I found that once we know what to build, making them at scale takes only a few years. The open question is whether we would know what to build at all, and I think that, given AGI, we probably would. Our "geniuses in a data center" could thus do more than sit around contemplating cancer cures and algebraic topology; they could also automate physical labor, once the necessary and not-too-hard to build actuators are constructed. That would kick off the kind of rapid physical growth discussed in previous posts.

An AGI that can do remote work can probably operate robots too

The term "artificial general intelligence" gets bandied about a lot, to the point it's at risk of becoming utterly useless. But there is, and always has been, a clear core concept that AGI points toward: an AI capable of doing the work humans can do. Authors have operationalized this in more precise ways for various purposes, but a system that can automate essentially all remote tasks is good enough for us. Astute readers may have noted that remote workers still exist (the author among them) and that current LLMs, smart as they may seem, cannot perform remote work (I've tried). AGI is also a much lower bar than superintelligence, which Bostrom (2014) defines as an intellect that "greatly exceeds the cognitive performance of humans in virtually all domains", or than Amodei's "powerful AI", which is "smarter than a Nobel Prize winner across most relevant fields".

Discussions of AGI focus on what the system can accomplish cognitively rather than on bodies or actuators. While sensible for isolating the core concept, this should not be confused with the claim that AGI cannot or will not act in the physical world. Humans can operate factories, drive cars, teleoperate robots, direct air traffic, perform surgery, and do numerous other "physical" tasks while being physically remote. If provided with suitable actuators, AGI would be similarly capable.

One could gerrymander these tasks out of the definition, so that AGI covers programming and paperwork but not machine operation. Such a system is not a logical impossibility, but I don't think it reflects the likely trajectory of technological progress. Remote cognitive work is hard. Doing all of it means mastering real-time control, spatial reasoning, physical prediction, and continual learning, and a system with those skills has what it needs to learn to operate most machines, much as a human handed the controls becomes passable within a few hours and proficient within months.

We can already see current LLMs developing rudimentary physical competency. They can provide high-level control over robots, watching the scene and choosing the next action, for example running a robot arm or flying a drone.[1] They do the same in analogous virtual settings,playing Pokémon slowly. And they show broader physical reasoning, from "vibe CADding" parts for 3D printing to programming robot dogs.

The most challenging case is work that requires dexterity and quick, fine-grained control of motion and force. It needs a fast control loop, since the large models supplying the high-level judgment are too slow to run it, and running that loop well takes a detailed understanding of how the actuator behaves. But neither is clearly a hard limit — a model running many times faster than a human need not find the speed a problem — and both are just what current systems are getting good at, with Physical Intelligence's π0, Gemini Robotics, and Figure's Helix integrating high-level reasoning with fast low-level control. Low-level hand dexterity, moreover, is mostly learned from very short interactions — a grasp or a fine adjustment, over in a second or two — which makes it a natural target for brute force, with a fleet of arms or a simulator running through vast numbers of them.

One might worry that the tacit, hard-to-verbalize know-how of a skilled tradesman would be hard for an AI to acquire. But acquiring tacit knowledge is not peculiar to physical work, and an AGI able to do remote cognitive work will almost certainly have to be competent at it regardless.[2] Indeed, an AGI could acquire a vast amount of tacit knowledge, much as it has already acquired vast, even superhuman, knowledge by ingesting the internet. This could be facilitated by parallel training across many instances, if that proves feasible, and what it learns can be specialized and copied as many times as needed.

For these reasons, I expect that an AGI able to automate most remote work will also be able to learn to operate robots, potentially to a high degree of sophistication. And operated robots would be valuable enough that there would be every incentive to push further, whether by specializing the models themselves or by giving them access to specialized tools for fine motor control.

Automating physical production mostly requires automating physical industry

What is required to automate physical production? At the very least, you need the robots and computer chips that can substitute for human labor, and you need to be able to make those work automatically. The computer chips require silicon fabs, which in turn require refining the silicon and producing a range of machine tools, dies, and so forth. Likewise, the robots themselves require all sorts of parts that must themselves be made from more materials. We obviously need a lot of energy to power this whole economy, and a good deal of infrastructure like roads and bridges to move things around.

Using the US input-output tables and the estimated compute and robotics costs from Part 1 of this series, we can actually calculate what a very naive version of this self-replicating economy would look like. We do exactly what we do today, but replace the human labor with the equivalent chips and robots. What we find is an economy heavily structured toward production — toward construction, production and other machinery, electrical equipment, computers, and other heavy industry:

Sector

Labor %

Output %

Construction

35.1

24.0

Machinery

18.5

20.2

Fabricated metal products

9.6

7.3

Electrical equipment and appliances

6.4

5.3

Computers and electronic products

5.9

6.5

Primary metals

3.5

7.9

Motor vehicles and parts

2.3

5.0

Professional and technical services

1.9

1.3

Nonmetallic mineral products

1.8

2.0

Plastics and rubber products

1.6

1.7

Aerospace and other transport equipment

1.5

2.2

Miscellaneous manufacturing

1.3

1.2

All other sectors

10.6

15.4

The same data lets us look at which jobs this economy needs to automate, occupation by occupation. Again it is dominated by the heavy-industry trades:

Occupation

%

Occupation

%

Occupation

%

Construction laborers

5.0

CNC machine operators

1.5

Customer service reps

0.9

Assemblers & fabricators

4.7

Secretaries & admin assistants

1.4

Production helpers

0.8

Machinists

3.7

Wholesale & mfg sales reps

1.4

Accountants & auditors

0.8

Construction-trade supervisors

2.4

Plumbers & pipefitters

1.4

Molding & casting operators

0.8

Welders

2.4

Mechanical engineers

1.3

Maintenance supervisors

0.7

Electrical & electronic assemblers

2.2

Construction managers

1.3

Cost estimators

0.7

Production supervisors

2.2

Heavy truck drivers

1.2

Industrial production managers

0.7

General & operations managers

2.2

Bookkeeping & accounting clerks

1.1

Buyers & purchasing agents

0.7

Electrici…