Heretical thoughts on AI

What if AI changes everything except the economy?

26
  • Support Eli Dourado

    Since you liked this post, why not support Eli Dourado with a subscription?
    Annual$80/year
    Monthly$8/month
    Founding Member$240/year

Programming note: Last week when Twitter cut off access to third-party apps, I got a bit spooked. I realized I need to cultivate my other socials. Consequently, I spent Friday morning migrating my personal blog to Substack. In the process, I imported my disused Tinyletter email list—if you’re receiving this, you at one point, possibly long ago, signed up to receive emails from me. If you’re no longer interested, please feel free to unsubscribe; I won’t be offended.

Fool me once. In 1987, Robert Solow quipped, "You can see the computer age everywhere but in the productivity statistics." Incredibly, this observation happened before the introduction of the commercial Internet and smartphones, and yet it holds to this day. Despite a brief spasm of total factor productivity growth from 1995 to 2005 (arguably due to the economic opening of China, not to digital technology), growth since then has been dismal. In productivity terms, for the United States, the smartphone era has been the most economically stagnant period of the last century. In some European countries, total factor productivity is actually declining.

What should we make of this? Are computers and the Internet economically unimportant inventions? I find it hard to make this case. The way I work would be completely different without computers. I’m old enough to remember the 80s, and yet I can’t imagine doing research without Google and Wikipedia. My org runs on Slack and Google Docs—I shudder to think what it would be like to type up a draft on a typewriter and circulate it among colleagues for feedback. How did anything get done in olden times?

Intuitively, introspectively, the productivity statistics should be through the roof. But they aren’t. That’s why we have economic data, so that we aren’t fooled by our intuitions. The numbers constrain the stories we can tell about the world.

The next iteration of the computers, Internet, smartphone sequence appears to be machine learning. "AI" is having a moment. Github Copilot (my favorite AI tool) is already here—it is not speculative technology. Large language models have definitively passed the Turing test. Diffusion is transforming digital art and soon video. DeepMind’s AlphaZero and AlphaFold models have mastered StarCraft and protein folding; Meta’s CICERO plays Diplomacy at a human level. Today’s models are the worst that they will ever be—they will only get better from here.

Consequently, it seems staggeringly obvious that some form of AI should, real soon now, create unprecedented economic abundance.

But what if that’s wrong? Chow, Halperin, and Mazlish note that markets are not yet expecting any productivity gains from AI. What if AI ends up like the Internet—transformative to our daily lives while somehow not actually delivering major productivity gains? It’s worth considering.

An industrial perspective. I like to reason about the economy sector by sector because it imposes a bit of intellectual discipline. Stating what you expect or expected a particular technology to do to a given sector, and then summing across all the sectors is more concrete and rigorous than just stating what you expect the effect on the economy will be. In addition, you can initially focus on a few big sectors because it’s only the big sectors that can really move the needle on aggregate productivity.

Housing. The biggest driver of housing costs in major urban areas is land use policy. Computers don’t really help us here. Local policy is firmly in the realm of flesh and blood, of trust and human relationships and beating the other side. Perhaps the Internet has been helpful in spreading information about land use liberalization and organizing the YIMBY movement. But it’s not clear to me, even granting that, how AI could revolutionize land use policy.

One area where I wrongly expected the Internet to play a bigger role is in breaking the real estate agent cartel. In the Year of Our Lord 2023, real estate agents still reap nearly six percent of the sale price of most homes. Prices have gone up, so in dollar terms, real estate commissions are at all time highs. If the Internet could not break the real estate cartel, I am not sure why machine learning models would.

What else could AI do to make our housing sector more productive? Maybe something in construction? Boston Dynamics’s Atlas robot can do impressive construction-like things, but I am skeptical of such capital-intensive approaches to construction. For now, I am not seeing it and am writing off housing as a big chunk of the economy that the AI revolution will not really affect. Go prove me wrong.

Energy. The biggest challenge in energy is deployment. If you want to build a solar farm or a wind farm or a nuclear power plant or a geothermal plant on federal land or a natural gas pipeline or an electric transmission line or a mineral mine, you are in for years of delay, regulatory headaches, public opposition, and lawsuits. Computers and the Internet did not really affect this deployment challenge, nor do I know of any reason to believe AI will solve it.

Other aspects of the energy industry have benefited from information technology. Resource exploration and extraction is more effective with digital tools, like the sensor packages that go downhole in the shale fields. Computer-based tools have helped optimize the design of energy-producing equipment like turbines. They could also help us discover new materials and battery chemistries.

AI tools, then, are promising for some aspects of the energy industry, but unless we fix the deployment obstacles associated with energy projects, the economic productivity gains could be bottlenecked.

Transportation. We are bad at transportation in the US for reasons that don’t have anything to do with digital technology. We could automate our vehicles, but so far the Biden administration has opposed reducing even cargo train crew size by one for basically make-work reasons. Supersonic flight has been illegal over the United States for 50 years. We seem constitutionally incapable of building high-speed rail service in this country. The Anglosphere has uniquely high subway and urban transit costs. Don’t get me started on the Jones Act.

Of course, there have been some gains from digital tools. Ride sharing is an innovation that stems directly from smartphones and the Internet. As in all engineering disciplines, digital tools have made automobile and aircraft design better. Yet again, the bigger issues are non-digital. It’s not clear to me how digital tools or AI are going to dramatically increase our productivity here unless we address those issues first.

Health. I can think of all kinds of ways that digital technology, whether it’s simply computers and the Internet or new AI algorithms, should be able to improve medical productivity. But the reality is that the field moves very slowly; it took a global pandemic to get us modest telemedicine service.

The biggest opportunity to improve health productivity is to get humans out of the loop entirely. In this regard, drugs are unreasonably effective—if you can replace a surgical procedure performed by highly trained specialists with a few pills, that is a clear productivity gain.

The biggest gain from AI in medicine would be if it could help us get drugs to market at lower cost. The cost of clinical trials is out of control—up from $10,000 per patient to $500,000 per patient, according to STAT. The majority of this increase is due to industry dysfunction.

Milky Eggs, an ML-focused engineer who has seen clinical trials close up, despairs of ever fixing it. "The problems seem nearly intractable in their scope and magnitude," says Milky. "Even if a return to positive real rates leads to a renewed focus on the ‘world of atoms,’ it could take decades for the industry to become more efficient!"

I’ll stop there. OK, so that’s only four industries, but they are big ones. They are industries whose biggest bottlenecks weren’t addressed by computers, the Internet, and mobile devices. That is why broad-based economic stagnation has occurred in spite of impressive gains in IT.

If we don’t improve land use regulation, or remove the obstacles to deploying energy and transportation projects, or make clinical trials more cost-effective—if we don’t do the grueling, messy, human work of national, local, or internal politics—then no matter how good AI models get, the Great Stagnation will continue. We will see the machine learning age, to paraphrase Solow, everywhere but in the productivity statistics.

Swimming in content. The one industry that AI is sure to disrupt is media. In media, there are no environmental impact statements, and the First Amendment ensures the industry is open to all. I expect we’ll soon have AI-authored newsletters, virtual celebrities, algorithmically generated movies, and more. We will be swimming in content.

There are those who think that more content is a bad thing. We will waste more time. We will be more distracted. But even putting those issues aside, we may be reaching diminishing marginal returns to media production. When I lived in Portugal as a child in the late 1980s, we had no Internet and two TV channels. I don’t know how much more content I have access to today, but it is perhaps a million times more (Ten million? More? I’m not even sure of the order of magnitude.)

That increase in content is life changing, but if the amount of content increased by another factor of a million because of AI, it’s not clear my life would change at all. Already, my marginal decision is about what content not to consume, what tweeter to unfollow, and more generally how to better curate my content stream.

And what should we say about the fact, as Ezra Klein notes, that the cost of bullshit will go to zero? In my opinion, the most trenchant prediction about the near-future media landscape comes from Neal Stephenson in his otherwise forgettable

Fall, or Dodge in Hell. To cope with a wave of online public shaming, one of the characters floods the Internet with AI-created bullshit. After all, if nothing on the Internet is believable, you don’t have to worry about what the Internet might say about you.

In response to an Internet full of bullshit, users require the use of an "editor," a service that edits the news feed for you. It weeds out the stuff that you don’t want and gives you the good stuff. Many users can only afford an algorithmic editor, like a spam filter for news. The well-to-do have human editors who monitor the news feed and artisanally curate it for their customers.

A funny thing happens in the book. Some customers, it turns out, prefer the fake news. They prefer bullshit. With "editors" existing as a competitive market, services spring up to give the people what they want. The most uproariously funny consequence in the novel is that a new cult called the Leviticans forms—naturally, they only believe in the book of Leviticus.

Stephenson’s use of the word "editor" is apt. It gets at an asymmetry between content creation and content curation. Writing is easy; editing is hard. Large language models like ChatGPT can write, but they are unable to curate their own output. They are unable to make it consistently good. To be sure, LLMs will only get better from here. But in a world of unlimited content, I only want the works of staggering genius. It’s not clear that a model trained in some sense to give the average next token can ever produce something so far above average.

Even if AI dramatically increases media output and it’s all high quality and there are no negative consequences, the effect on aggregate productivity is limited by the size of the media market, which is perhaps 2 percent of global GDP. If we want to really end the Great Stagnation, we need to disrupt some bigger industries.

I could be wrong. I remember the first time I watched what could be called an online video. As I recall, the first video-capable version of RealPlayer shipped with Windows 98. People said that online video streaming was the future.

Teenage Eli fired up Windows 98 to evaluate this claim. I opened RealPlayer and streamed a demo clip over my dial-up modem. The quality was abysmal. It was a clip of a guy surfing, and over the modem and with a struggling CPU I got about 1 frame per second.

"This is never going to work," I thought. "There is no way that online video is ever going to be a thing."

Exponential growth not only in processors but also in Internet speeds made short work of my denunciation of video streaming. I think of this experience often when I make predictions about the future of technology.

And yet, if I had only said, "there is no way that online video will meaningfully contribute to economic growth," I would have been right.

Elsewhere. Over at the CGO, I have a piece arguing that Congress, not FAA, should take the next steps on legalizing supersonic flight over the US. If you’ve gotten to the end of this and can stand more from me, you should read that one as well.

1

Forgive me, Neal! I love you! Long live The Baroque Cycle!

Subscribe to Eli Dourado

Launched 3 years ago

Trying to end the Great Stagnation

26
  • Support Eli Dourado

    Since you liked this post, why not support Eli Dourado with a subscription?
    Annual$80/year
    Monthly$8/month
    Founding Member$240/year

26 likes
Comments

Ready for more?

© 2023 Eli Dourado
Substack is the home for great writing