I’ve met and seen several people that spend time with LLMs and think this is going to fundamentally change the economy, to the point we’re going to get massive increases in productivity and massive dislocation. And there is therefore a justification in the intense build-out of AI infrastructure. If you spend a trillion dollars today, you will get many more trillions in benefit in not the distant future, but the near-term future. And while they might say the power problem needs to be solved, or we need social policy to manage the economic dislocation, the don’t question the primacy of building more data centers, faster.

Ed Zitron has done a fantastic job of taking apart the holes in some of the financing. And they are substantive. But I’m going to make a slightly different argument. One based on how realistic it is to believe LLMs will change the world around us. And for this, I will assume that the products of LLMs do not face a quality issue. That they write code, screen plays, or generate images very well.

First, take a look around and look at all the things you’ve bought that have nothing to do with computers. Like your sofa, the lamp next to you, or the dishes in your cupboards. The production of many of these things, where possible, is already automated. There is already a built in limit as to how much LLMs will impact the world because the vast majority of the world is not computerized. Would new robots based on LLMs improve the world by making good cheaper through automation?

Take your the IKEA Billy Bookcase. IKEA sells about 7 million of those every year. The process of getting that Billy from a plant in Indonesia to your house is automated. From the production of the sheet goods, to their cutting, their packaging, in the warehouses, on containers, and so on. Maybe a human had to take a fork-lift to put it on a shelf at IKEA, and you had to lug it up the stairs from your car, but it is already highly automated. Your car is the same. Most of what is done is already automated. On a car like my Honda Accord, its is probably less than 10% of the cost and may be as low as 5%. We are already highly automated and becoming more automated, even in the absence of LLMs.

What drives the development of better robots is not just software, it’s also better materials, better design, better motors, better sensors, better ball-bearings, and so on. We’ve been making and using industrial robots for decades, and are now more cost effective than not just skilled workers like welders, but unskilled workers, like people packing boxes with stuff to mail. That is already true today and done mostly without LLMs (although other types of AI have helped). LLMs will only partly help us with parts of the discovery of new materials, designs, or meta-materials to make these robots more cheaply or with greater capabilities. AI technologies in use since the 1970s, like evolutionary algorithms, already help design and build lighter and better structures (for example).

That leaves the impact of LLMs to the service economy. And in the service economy there are limits on what is actually automated. A policeman is a service worker, and cannot be automated. Firemen cannot be automated with a chatbot. Your doctor can be augmented with an LLM, but still has to see you in the office. The nurse inserting an IV cannot be automated away with an LLM. An LLM can’t clean your teeth. Or wash your car (something that has been almost fully automated for decades). Can an LLM translated “The Brothers Karamozov” for you? Yes, but maybe miss some of the nuance of Russian to English translation. But it would work fine for news stories, business emails, or technical documents.

Let’s be generous and assume that 25% of the economy is computerized or computer adjacent, such that developments in computers will directly translate into improvements in that field. And by direct, I mean that a computer is used in the process, like accounting, engineering, controlling machinery, and so on. That means a 100% improvement in productivity will have a maximum impact of 25% on GDP. (Long term real growth in an economy is roughly productivity + population growth¹).

But of that 25% of the economy that could be improved by LLMs, because comptuers are very involved, not everything needs or uses what LLMs bring to the table, Of that, how much can be improved with an LLM? Let’s say half of that is either simple automation, like a point of sale system, industrial robots, accounting software, or other systems that are already automated with normal code. And that, at best, only part of their function is improved, so we cut the total impact in half. Now, the largest impact is 12.5%.

If we take a look at a job like writing software, we realize that’s only part of an engineer’s time. Maybe 1/3 is spent in meetings, 1/3 spent figuring out what to build, and 1/3 actually building it. Some weeks are spent 80% coding, but some weeks are spent 0% coding. LLMs can’t really tell you want to build, so let’s assume that doctors, lawyers, and all other professions will be improved as much as coding, and cut that 12.5% down to about 4%. A 4% bump in productivity is still HUGE! But we all know LLMs are not 100% effective and require some amount of re-work, prompt tuning, and we need to actually look at the outputs to make sure it’s going well. So maybe that 4% is actually closer to 2%. I say this because there is much more code-churn on AI generated code than human code (as my example). So if we do it once, we’re likely to have to do it over in the future.

Two things can be true at the same time. That the impact of LLMs on the over-all economy can be so small it’s lost in the month to month noise, and in specific fields they are transformative. Take, for example, LLMs that have come up with proofs of open math questions. At some point someone still has to verify the proof. (And it has happened that an existing proof was later found to be incorrect – sometimes decades after). An LLM proposing a new meta-material based on being fed hundreds of studies and papers still requires someone to figure out how to produce it and test it. It still must be verified in the lab and in real world use. And, like when humans read the same papers, the LLM could have gotten it wrong because the underlying papers suggested the wrong conclusion.

There are about 160 million workers in the United States. All engineering disciplines (including civil, electrical, mechanical, aeronautical, etc) account for 6 million workers. That’s close to 4% of the labor pool. It’s easy for people to think every other worker is a software engineer, but that’s far from true. There are 8.3 million construction workers, about 5% of the labor pool. And there are about 17 million in hospitality, 10% of the labor pool. How many people are going to want a steak made by an LLM invented recipe (knowing it could change any time), cooked by a robot? It might be a delicious steak, but would consumers want it? And knowing that, would they spend $40 to $50 on that steak (a price at a nice steak house)?

But the costs of building out LLMs isn’t just paid by software developers. And the systemic risk to the economy of driving this much investment, this quickly, into one technology is borne by everyone. It’s at the point where it chokes off the investment in other technologies by absorbing vast amounts of capital. Companies running LLMs are pushing to burn more fossil fuels faster, to power their data centers. The belief is that it is so transformative, it will change … everything, even if people can’t really explain how. And the transformative impact on the economy just isn’t there. But, when the bubble bursts, and companies find themselves with billions of worthless debt used to finance empty data centers, everyone will be tapped when the government has to step in an ensure credit markets continue to work.

1. Note that I said real growth – meaning outside of inflation. Nominal GDP should increase (long term) as inflation + productivity increases + population growth. Real growth is when we factor out inflation and just focus on the improvement in living standards. However, if you don’t control inflation, and your population isn’t growing, you will wipe out the effect of productivity by swings in the price level. ↩︎