On Charkhas, AI & Economic Growth
Almost everyone gets the question of ‘AI economic impact’ completely wrong. a few leaps of imagination is all that it takes.
How does economic disruption work? A new abstraction comes along that renders an existing profession redundant. So in the 1800s, the handloom rendered the textile workers of the South redundant. Cars rendered horse drawn carriages redundant.
The way technological progress works is that a new technology/method that is 100x better replaces scores of slightly mediocre solutions implemented in parallel. The charkha isnt just an inferior solution, both serially and in parallel, the handloom is just a completely new thing altogether.
Todays discourse is solely in terms of unit-wise replacement. Humanoid robots replacing other blue collar workers. One programming agent replacing another an SWE at work. One AI call center employee replacing 5 call center employees.
Or its about quantity. 100 billion new agentic programmers working at your beck and call! 50000 AI researchers! What would a company do with 500 new employees that worked 24 hours a day, never tired. The Charkha mindset, again.
We have these incremental ways of thinking about the effects of AI because we dont really have any AI technology that can invent new things. We dont have anything that can look at a charkha, causally model all the parts that work together, evaluate the inputs and outputs in the supply chain, identify bottlenecks which require the most human intervention, and painfully iterate on hardware prototypes until it invents the handloom. If you asked charkha operators in 1700s what machine would replace them, they’d say something that like a machine that spins the wheels faster, or auto-loads the yarn. They would never think of the motorized handloom.
In todays terms this looks like : ‘AI writes 80% of my code’. This is meaningless for any economic miracle. What you want is the kind of AI builds something new that replaces the need for vast swathes of programmers to ever be required to write that code in the first place.
Infact, AIs that can invent and discover are entirely possible to build and will change the way we think about the economic impact of AI. If we assume these machines can invent new things, and plan out entire technology trees of progress by causally chaining parts of economic clockwork in ways where the sum is exponentially better than any the parts, then the way to think about AI is like the role of an inventor.
The role of an inventor is not just an incremental improvement, it is an entirely different phase change from all other types of economic activity. There is no precedent, no previous training data, no sign that the impossible is within grasp apart from an irrational nudge that says keep going on. 99% of outcomes are failure, and each instance failure has to be modelled uniquely, and mined for just the right lessons to try the next attempt with. One mistake in this modelling, or mistake in measurement and you misattribute failure to the wrong changing variable, or overindex on noise, leading to vicious cycle of failure that is hard to recover from (this cycle is why inventing new things is so rare, and we dont have any set ways to reason about it). Insight and inspiration comes from entirely unexpected places. It is impossible to teach or train for the method of invention because by definition no curriculum includes anything new. Often it involves finding and standing on the shoulders of obscure giants forgotten in time due to epitemological reasons. Often, leaps of faith in the face of heaps of counterevidence. The discovery of fire was an irrational act. We are descendants of the most irrational ape.
But when it works, the messy process of invention deals wonders. It results in an ecological change - a forest without a tiger is just a bunch of trees, a forest with a tiger isnt just the same forest with the addition of a tiger, its a completely different forest altogether.
So what does economically disruptive AI look like? It doesnt look like 5000 spun up AWS instances running 500 instances of the same agent loop. It doesnt look like 5000 1-person companies catering to the same consumer demand with fewer people. It looks like a team of 4-5 AIs working in solitude, mostly away from noise of daily trends, building a library of tools and methods that build on top of and complement each other, like Newton inventing calculus to explain gravity, iterating quickly and mining the correct signal from failure, incrementally making and validating the subcomponents of their hypotheses, adapting to change. They work to make one expressive abstraction/tool - a new vertically integrated factory that produces 5 electric cars per second during the day, and can be repurposed to manufacture appliances at night. A new type of cooking appliance that takes in ingredients and outputs cooked packed meals, that fits in a room/kitchen top, rendering a lot of restaurants useless. new virtual reality technology that makes in person work feel like co locating, shooting down property prices in city suburbs. A new fusion method that is so ubiquitous and easy to harvest, that it renders most oil geopolitics irrelevant overnight. A medicine that makes it 1000x less energy expensive to produce a female ovum, upending every scarcity-based gender gynamic in society. A way of programming matter so physical goods can be 3d printed in house, reducing the need to ship things all over the world.
In common discourse, there is a distinction drawn between AIs that can just automate existing tasks vs AIs that can invent new things. I differ here. All economic value comes from inventing/creating distinct abstractions to manipulate the environment - the world puts a higher value on you depending on how unique your abstraction library is. New recruits are valued based on how quickly they acquire new skills (not based on how many things they can do). New companies are based on the tech tree they can own and manipulate, the unique abstraction tree that only they have access to. Abstractions cease to become valuable when they are commoditized - airlines were valuable services when only a few knew how to make planes. But today in aviation only boeing and airbus make money by selling planes to everyone. They have figured a proprietary abstraction ladder that only they have tools, personnel, institutional scarring and experience with to maintain and use. AI disruption will work similarly.
Think obscure strange unique abstraction ladders spun up ad hoc over weeks (not years), that allow an AI to build and offer a 1000x better service/product for a 100x lesser price. What does a few cycles of this iterative oneuppance look like? The world looks very different when you have AIs rapidly building on each others abstraction libraries, exchanged via precise context free grammar (and not encumbered by stagnating high school and college curricula). This is the real disruption to worry about because it’s non-linear and hard to predict. The handloom replaced an entire manual textile industry, what happens when you have a machine that can climb the abstraction ladder to create arbitrary handlooms for arbitrary industries at arbitrarily short time scales?