Pleb Economist #9: Comparative Advantage, AI, and You

• Introduction
• Trade before LLMs
• Trade after LLMs
• Conclusion

IntroductionIntroduction

Couple weeks ago, @Undisciplined made a nice post explaining the Theory of Comparative Advantage, which itself was motivated by a cartoon posted by @Car (#1408254):

As @Undisciplined explained, the Theory of Comparative Advantage says that people can mutually benefit from trading with each other, even if one person is better at everything. Michael Jordan might have been a better basketball player and better baseball player than the worst MLB scrub at the time; but his time was still better spent playing basketball than baseball, whereas the MLB scrub's time was better spent playing baseball. Similarly, Ray Dalio is probably a better investor and a better economics teacher than me; but his time is better spent investing whereas my time is probably better spent teaching since I suck even more at investing.

How does it apply to the cartoon? The idea is that even if LLMs make everyone equally good at coding, that does not mean the coder loses his comparative advantage. Since the coder is still relatively better at coding than at other tasks, it still makes sense for them to keep doing the coding while people better at social skills keep doing the talking and managing (joke about antisocial engineers).

This post, however, is going to add a bit more nuance to the idea. Basically I'm going to argue that although the coder still has a comparative advantage at coding, in the age of LLMs the bargaining power that their comparative advantage affords them dwindles. They actually end up worse off in the world of LLMs because they can't offer their services at as favorable of an exchange rate as they could before.

Trade before LLMsTrade before LLMs

Ok, so let's set up a simple model. There are two people in this economy: HandyWorker and DevWorker, and two goods Stuff and Code.

• HandyWorker can either produce 20 stuffs per day or 10 codes per day. (Or he can split his time between the two in a continuous manner.)
• DevWorker can produce either 10 stuffs per day or 20 codes per day. (He can also split his time continuously).

The Theory of Comparative Advantage can be illustrated with a very simple trading relationship:

• HandyWorker will spend all his time making 20 stuffs per day
• DevWorker will spend all his time making 20 codes per day
• The total daily production is therefore 20 stuffs and 20 codes
• They then trade equally at a 1:1 exchange rate, so that each consumes 10 stuffs and 10 codes per day.

Note that 10 of each was not possible for either of them without trade, because if HandyWorker wanted 10 codes without trading for it, he would have had to work all day for that; and if DevWorker wanted 10 stuffs, he likewise would have had to work all day for it. The simple example shows why trade is beneficial.

Trade after LLMsTrade after LLMs

Now let's introduce an LLM that makes HandyWorker equally good at coding as DevWorker:

• Now HandyWorker can produce either 20 stuffs per day or 20 codes per day
• DevWorker isn't affected by the LLM since he was already a master coder: he still produces either 10 stuffs per day or 20 codes per day

If we now try to work out the possible mutually beneficial trade scenarios, we realize that the 1:1 exchange rate between stuff and code no longer works. This is because HandyWorker can make that exchange rate internally. If DevWorker offered to trade 10 code for 10 stuff as in the previous scenario, HandyWorker would say "No" because he can just do that himself without trading.

In fact, if we work out the math, we can show that the set of mutually beneficial trades has actually shrunk for DevWorker as shown in the below graphs:

The first two graphs shows the set of possible consumption states for the two workers, with and without trade, before and after LLMs. The LLM expands HandyWorker's consumption possibility set but shrinks DevWorker's, because it erodes the bargaining power of DevWorker's comparative advantage.

However, comparative advantage still exists, as we can see that trade is still mutually beneficial both with and without LLMs. It's just now less beneficial to DevWorker than it used to be, prior to LLMs.

At the same time, we see that the LLM does improve the overall economy as a whole. The third graph shows that the set of all possible production states has actually expanded.

To bring this into real world terms: it suggests that LLMs will epxand total GDP, but lower the relative wages of those whose advantage is replaceable by LLMs, while raising the relative wages of those who are less replaceable by LLMs.

ConclusionConclusion

Ok, so the final result isn't that surprising or deep. Probably anyone who sat down for a few minutes to think about it would have come to the same conclusion (it's not rocket science).

But it's still useful to see how it plays out in a model, and the model helps highlight the chain of reasoning that leads to the result. The primary reason for the erosion of DevWorker's wages is because HandyWorker's acceptable exchange rate for code has changed, due to his own better options. In a sense, the LLM actually shrinks the set of acceptable trades, because HandyWorker is now more picky.

And I think this is exactly how it's going to play out in reality. The young entrepreneur who has an idea but few coding skills no longer has to pay his first engineer to build a prototype; he can do it himself. If an engineer wants to work for him, the engineer will have to offer a lower wage. The small to medium sized business who wants to scale can do so with a smaller staff than before, and the staff may have less bargaining power over their wages. We know these stories, and the stories are confirmed by economic theory. Those who want to maintain a high degree of bargaining power in the post-LLM economy would do well to invest in skills that are not easily replaceable by the LLM.

As to what those skills are... well, don't ask me. It's a brave new world for all of us.