The modern economy is quite integrated, so lots of things ultimately require quite a diverse range of raw and transformed materials. How diverse? What’s the minimally-reachable (not at tech maturity, but starting where we are now) in-principle fully closed manufacturing loop? How big can it scale, and how soon do other loops get closed?
I believe there’s designs that vastly reduce the diversity and/or quantity of expensive inputs. As an example, Boston Dynamics Atlas humanoid robots use servohydraulic joints. A small control motor similar in size to ones found in tiny quadrotors can control a lot of power. This vastly reduces power electronics and motor size/weight. There’s some compelling reasons to use servohydraulics and patents related to that are important to really optimising for growth.
If something is in-principle self-sustaining, under what conditions does that get actualised? And how voraciously?One naive perspective is that if it’s all fully grounded in the human demand economy, then it’d expand only to the point where the marginal utility of extra supply of goods output per raw materials ingested is matched by alternative uses of those raw materials.
One naive perspective is that if it’s all fully grounded in the human demand economy, then it’d expand only to the point where the marginal utility of extra supply of goods output per raw materials ingested is matched by alternative uses of those raw materials. Since marginal utility usually diminishes quite fast with increased supply, that might mean not very much expansion in practice.
There are military implications to being able to produce arbitrary quantities of stuff.
There would be immediate demand to replace/upgrade existing production equipment with things that are fully automatable.
To formalise this model a bit more:
in the legacy economy, producing a widget requires paying a bunch of people and obtaining natural resources at some cost X
after automating, most of the labor hours are gone. Amortised capital costs are lower since capex is mostly manufactured goods. The cost for a given good goes down to the cost of obtaining raw materials and remaining fraction of the labor hours.
You can imagine some fraction of the dependencies (weighted recursively) are now close to their material costs.
There’s pressure to route around dependencies that aren’t cheap yet or make them cheap.
I could see 50%-90% drop in costs of manufactured goods.
As to actual demand drivers, a lot of the economy is bottle necked on manufacturing. How many self-driving Waymos does alphabet want to build? Is there a business case for building enough to replace 10% of the first world vehicle fleet this year if production could be scaled that far?.
What are the timelines and bottlenecks to the AI components for something like this? What about the machinery and robotics components?
Manufacturing mostly doesn’t work because the people involved often don’t know if what they’re doing was good or bad. A/B testing is not universal, data isn’t being collected. There are some industries that get this right because they have to to function. TSMC couldn’t make the latest logic node if they weren’t collecting gigabytes of data from each machine, Doing automated machine learning-like tuning of control variables and running sophisticated statistical models to track correlates of good outcomes. Achieving process stability in semiconductor is just that hard.
In the rest of the economy, that almost never happens. Engineers often don’t know if a change they made was good or bad unless there is an immediate and obvious failure. They often don’t have the domain expertise to properly root cause problems. Processes accumulate mutational load as people tweak things until the stability margin is gone. Most processes operate with no extra stability margin because that’s the point where you can tell obviously that what you did broke things. Then something hard to troubleshoot goes wrong, the stability margin disappears completely and things are very very much on fire. That’s my experience in industry.
Timelines on the AI components are hard to pinpoint. In the best case scenario, AI would be able to design machinery. Learn from experience building/operating it, and then build stuff that works.
My belief is that actually working big data is enough. What we have today, properly diffused into industry will collect all the statistics/data necessary to telling when things are going right/wrong and be able to bring processes under control. Automation will start to work reliably.
Keep in mind, what we have today (claude code/codex) is enough to fix a lot of issues. Buggy pallet loading/unloading code written in python has locked up automated machining cells and caused significant downtime. That’s the sort of thing claude code can one-shot. Entire projects I’ve done involving data acquisition and interfacing with machines (EG:automating a setup process involving setting machine tool offsets) are trivial with agentic coding tools. Some things that didn’t feel worth it related to monitoring certain machine data to detect failures early are trivial now.
I’m in the process of writing this all out, but IMO, what we have right now might be enough to get us there. There are already vendors in place with boxes plugged into manufacturing cells who will start doing this.
I believe there’s designs that vastly reduce the diversity and/or quantity of expensive inputs. As an example, Boston Dynamics Atlas humanoid robots use servohydraulic joints. A small control motor similar in size to ones found in tiny quadrotors can control a lot of power. This vastly reduces power electronics and motor size/weight. There’s some compelling reasons to use servohydraulics and patents related to that are important to really optimising for growth.
There are military implications to being able to produce arbitrary quantities of stuff.
There would be immediate demand to replace/upgrade existing production equipment with things that are fully automatable.
To formalise this model a bit more:
in the legacy economy, producing a widget requires paying a bunch of people and obtaining natural resources at some cost X
after automating, most of the labor hours are gone. Amortised capital costs are lower since capex is mostly manufactured goods. The cost for a given good goes down to the cost of obtaining raw materials and remaining fraction of the labor hours.
You can imagine some fraction of the dependencies (weighted recursively) are now close to their material costs.
There’s pressure to route around dependencies that aren’t cheap yet or make them cheap.
I could see 50%-90% drop in costs of manufactured goods.
As to actual demand drivers, a lot of the economy is bottle necked on manufacturing. How many self-driving Waymos does alphabet want to build? Is there a business case for building enough to replace 10% of the first world vehicle fleet this year if production could be scaled that far?.
Manufacturing mostly doesn’t work because the people involved often don’t know if what they’re doing was good or bad. A/B testing is not universal, data isn’t being collected. There are some industries that get this right because they have to to function. TSMC couldn’t make the latest logic node if they weren’t collecting gigabytes of data from each machine, Doing automated machine learning-like tuning of control variables and running sophisticated statistical models to track correlates of good outcomes. Achieving process stability in semiconductor is just that hard.
In the rest of the economy, that almost never happens. Engineers often don’t know if a change they made was good or bad unless there is an immediate and obvious failure. They often don’t have the domain expertise to properly root cause problems. Processes accumulate mutational load as people tweak things until the stability margin is gone. Most processes operate with no extra stability margin because that’s the point where you can tell obviously that what you did broke things. Then something hard to troubleshoot goes wrong, the stability margin disappears completely and things are very very much on fire. That’s my experience in industry.
Timelines on the AI components are hard to pinpoint. In the best case scenario, AI would be able to design machinery. Learn from experience building/operating it, and then build stuff that works.
My belief is that actually working big data is enough. What we have today, properly diffused into industry will collect all the statistics/data necessary to telling when things are going right/wrong and be able to bring processes under control. Automation will start to work reliably.
Keep in mind, what we have today (claude code/codex) is enough to fix a lot of issues. Buggy pallet loading/unloading code written in python has locked up automated machining cells and caused significant downtime. That’s the sort of thing claude code can one-shot. Entire projects I’ve done involving data acquisition and interfacing with machines (EG:automating a setup process involving setting machine tool offsets) are trivial with agentic coding tools. Some things that didn’t feel worth it related to monitoring certain machine data to detect failures early are trivial now.
I’m in the process of writing this all out, but IMO, what we have right now might be enough to get us there. There are already vendors in place with boxes plugged into manufacturing cells who will start doing this.