Jevons Paradox: The Most Important Idea in AI

Thoughts on AI and "The End of Software"

Chris Paik recently wrote “The End of Software”, where he predicts LLMs will “drive the cost of creating software to zero.” Paik expects a “Cambrian explosion of software” in the same way we had a Cambrian explosion of content due to the introduction of the internet.  As a result, Paik predicts that: 

1. Software companies will be like newspapers: good businesses in the past but bad to invest in going forward given the new abundance of competition.

2. “Majoring in computer science today will be like majoring in journalism in the late 90’s” and software engineering will go from a prestigious highly paid role to a lower paid one that shrinks in importance and number. 

Paik makes some interesting points but I believe he’s too focused on the supply side impact of AI.  Software engineers will get more productive, but demand for software is going to increase by even more, a phenomenon known as “Jevons Paradox.”¹ As a result, I expect software engineering will remain a great job to get and the dominance of software companies will continue even if the day to day of software engineers changes materially. 

Jevons Paradox in Software

AI is likely to make software engineers (as well as almost any other knowledge work profession) way more productive over time.  This demo from Devin² shows the future promise of AI automating software engineering.  In the future, maybe a software engineer will be able to do a task in 1 day (or less!) which previously took them 2 weeks. On paper, it seems inevitable that we'd need fewer software engineers.

But this is where Jevons Paradox comes into play. This paradox describes when technological improvements that increase the efficiency of resource use actually lead to an overall increase in resource consumption, rather than a decrease. In Econ 101 terms, when demand for a good/service is very elastic, a reduction in price can cause such an increase in quantity demanded that the total consumption increases.

In this example, quantity demanded is rising from 1000 (100*10) to 1120 (80*14)

Economist William Stanley Jevons first made this observation with coal consumption in the 1800s, but it also applies to software.  The lesson of the last few decades is that demand for software is extremely elastic and when we get more productive with software, we just demand more of it.  And this makes sense - software is valuable because it can replace labor and save customers and businesses money. That’s not changing anytime soon.

If you think about it, we’ve already seen a huge productivity increase in software engineering in recent decades similar to the one Paik is predicting going forward. We now have the internet, Github, cloud, better software packages, open source libraries, etc. Some things which took days to do in 1990 can take minutes today: setting up new projects, deploying web applications, adding user authentication, scaling applications, etc.  Plenty of things we do today were impossible back then and many types of tasks can now be done by non-technical workers (ex: creating your website on Squarespace with no code). 

And despite these productivity improvements, we have more software engineers employed than ever before and being paid record amounts³! The huge demand for software has meant that companies in aggregate did not fire software engineers - they instead hired them in droves to produce more of it. As simpler tasks were automated/took less time, software engineers now work on more complex and valuable use cases: AI and machine learning, realtime data management, security, APIs, etc. Given there are endless software problems to solve and the value of software is only increasing in today’s economy, I’d expect the same dynamic to continue with AI.   

Source: BLS

Jevons Paradox in other industries

There are plenty of examples in history where we’ve seen technological improvements cause Jevons Paradox:

1. Data Storage: As the costs of data storage have declined by several orders of magnitude, we’ve just ended up consuming way more of it.  The original iPhone in 2007 had a max of 16gb of storage.  Even though the unit cost per GB of storage has come way down since then, people demand so much more storage that new iPhones start at 256GB.  

2. Data Analysis:  The average data analyst today can probably do the work of a whole team 20 years ago given improvements in productivity.  But as we’ve gotten faster and better at data analysis, we’ve just demanded more and more analysis.  As a result, we have way more data analysts and scientists than we did back then despite each individual analysis taking way less time. 

3. Movie production: Marc Andreessen’s favorite example is CGI. Instead of making the cost of movie production fall, CGI actually increased the costs because audience expectations rose so much.

4. Healthcare: While this is less of an automation or productivity story, the slew of healthcare innovation across devices (MRIs), drugs, surgeries, and information (the internet) has only increased demand for healthcare.  One of the best explanations for why the US spends more on healthcare than other countries is simply that we are richer and demand for healthcare is highly elastic.  Even if healthcare prices were to come down, we would likely consume even more of it so total spending may not fall. 

There are plenty more examples like this.  Going forward, Jevons Paradox is the reason total spending on compute is likely to go up drastically in the next 10 years, not down, even if per token costs come down by multiple orders of magnitude. 

When does demand not offset?

Jevons Paradox doesn’t always apply.  There are plenty of examples where technology improvements do cause employment to fall in industries when demand is not elastic enough to offset supply shocks.  

As farming productivity has skyrocketed, farming employment has fallen from 1/3rd of US workers in the early 1900s to 1.3% today.  Unlike the some examples above, our demand for food can only rise by so much.  If we can produce 10x more food per laborer than we did previously, but only eat at most 2x more food than we used to, it’s hard for farming employment to not decline.  For some industries, AI may have this type of impact (ex: call centers) but I don’t believe software engineering is in that category. 

We’ve also seen some technologies fully substitute certain workers as opposed to complementing.  Switchboard operators, milkmen, telegraph operators, and human alarm clocks have all gone to zero as technology removed the need for them completely.  

If true Artificial General/Super Intelligence comes along as soon as some predict, then it’s plausible to see this same thing happen to software engineers.  The strongest argument I’ve heard here is that software engineering output is fundamentally about code, which AI models are better at creating today than the output of other white collar jobs (ex: an operations manager needs to deal with people). But I think this underrates the multi-faceted nature of the job and software is not different enough from other white collar jobs where it would be uniquely at risk in an AGI future.  

In the absence of true AGI in the near-term,⁴ I don’t see a strong argument for full substitution of software engineers and as François Chollet predicted on Dwarkesh’s podcast last week, I think there’s a good chance we have more software engineers in five years than there are today, not fewer.

The future of software 

What does this all mean for the future of software engineers and software companies?

Even if Jevons Paradox holds, I think we’ll still see massive changes to software engineering.  The day to day is likely to move further down the complexity spectrum but at a faster rate than ever before. There will be far less time writing coding and far more time evaluating code and thinking about the more complex tasks which require judgment like “what to build” and “how to build it.”  Because of this, as my friend Rob Dearborn notes, Product Manager/Design/Software engineering roles could converge into a “software producer” role where the same person who comes up with the product ideas is also prompting AI to produce the code and then evaluate it.  

Unlike Paik, I think software as a category remains a strong place to invest too.   The difference between software and media is that there will always be more complex software to build which requires more time/resourcing and enterprises are willing to pay more for.  In media, there really isn’t a premium tier that is worth paying for so the flood of content creators on social media provide stiff competition for traditional media.  And while it’s easy to build a simple app today and share it out, there isn’t as much competition for more complex software used by enterprises, especially in highly technical and R&D heavy categories like GPU chip design software, cloud, aerospace, etc.  

Another reason why software companies are so valuable, and why it’s hard for your average person to compete, is the distribution they control. Paik predicts new platforms will control distribution: 

“Vogue wasn’t replaced by another fashion media company, it was replaced by 10,000 influencers. Salesforce will not be replaced by another monolithic CRM. It will be replaced by a constellation of things that dynamically serve the same intent and pain points. Software companies will be replaced the same way media companies were, giving rise to a new set of platforms that control distribution.”

But will AI really have this impact? The internet and social media proliferating hasn’t done a lot to help smaller software providers compete with incumbents.  Even if AI lowers the cost of creation, it’s not doing a lot to change distribution.  As software gets more complex with AI, I’d expect even more value to come from large companies who are able to serve increasingly complex enterprise use cases.  

Anytime you have a big shift like AI, there will obviously be some disruption of incumbents.  But I’m betting that software companies as a category remain at the top of the list of most valuable companies in the world.

1. Credit to Marc Andreessen for first flagging this concept in recent months.

2. Unclear how good it really is today, but the idea is promising.

3. I only see data until 2019 so it’s possible things are flat/down since.

4. Answering this question would take a much longer piece so I’ll leave it there for this article.