Quote of the Issue

“Radioactive monsters, utopian atom-powered cities, weird ray devices, and many other images have crept into the way everyone thinks about nuclear weapons and power plants. The images, connecting with major social and psychological forces, have exerted a strange and powerful pressure on our history. This is no story of things locked away safely in the past: the images are as strong today as ever.” – Spencer R. Weart, The Rise of Nuclear Fear

The Essay

🤖📈 How financial markets could predict the arrival of artificial general intelligence

When will artificial intelligence/machine learning, especially the new generative AI models, have a massively transformative effect on the American economy — if ever?

Well, don’t expect a big impact this year, according to MIT economist Daron Acemoglu, someone whose research efforts center around the economic impact of new technology. In a new essay for Wired, “Get Ready for the Great AI Disappointment,” Acemoglu argues that after so much hype in 2023, most notably with ChatGPT, this year will see a reality check. Nothing particularly groundbreaking here from the noted economist, but certainly worth noting. He thinks efforts at mitigating the tendency for large language models to generate “hallucinations” will prove a harder task than many optimists think, undermining the technology’s impact on boosting business productivity. It will take longer than expected for companies to figure out “which human tasks can be augmented by these models, and what types of additional training workers need to make this a reality.”

Not surprisingly, then, Acemogul doesn’t see human-level artificial general intelligence as happening anytime soon. From the essay:

Anticipation that there will be exponential improvements in productivity across the economy, or the much-vaunted first steps towards “artificial general intelligence”, or AGI, will fare no better. … Some people will start recognizing that it was always a pipe dream to reach anything resembling complex human cognition on the basis of predicting words.

Hints of our AI future from Wall Street

Of course, there are plenty of folks who see things differently, both in terms of near-term productivity improvements and the possibility of recent advances leading to AGI within a decade or so. I’ve been compiling such forecasts over the past year, including those by technologists, economists, and prediction markets.

And speaking of markets, what about looking to financial markets — vast, dynamic calculating machines that aggregate and analyze information — for signs of an emerging Age of AI?

That’s the intriguing question explored in the recent preliminary paper “Transformative AI, Existential Risk, and Asset Pricing” by Trevor Chow (Stanford), Basil Halperin (MIT), and J. Zachary Mazlish (Oxford and GPI). The paper’s abstract:

We study the implications of transformative artificial intelligence for asset prices, and in particular, how financial market prices can be used to forecast the arrival of such technology. We take into account the double-edged nature of transformative AI: while advanced AI could lead to a rapid acceleration in economic growth, some researchers are concerned that building a superintelligence misaligned with human values could create an existential risk for humanity. We show that under standard asset pricing theory, either possibility – aligned AI accelerating growth or unaligned AI risking extinction – would predict a large increase in real interest rates, due to consumption smoothing. The simple logic is that, under expectations of either rapid future growth or future extinction, agents will save less, increasing real interest rates. We contribute a variety of new empirical evidence confirming that, contrary to some recent work, higher growth expectations cause higher long-term real interest rates, as measured using inflation-linked bonds and rich cross-country survey data on inflation expectations. We conclude that monitoring real interest rates is a promising framework for forecasting AI timelines.

A few explanatory notes:

• By “transformative,” the researchers mean AI that has a profound impact on humanity, similar to the industrial or agricultural revolutions. They consider two scenarios: “aligned transformative AI,” where AI technology leads to a global GDP growth exceeding 30 percent per year, and “unaligned AI,” where AI causes the extinction of humanity.
• “Standard asset pricing theory” is a set of basic financial principles that help determine the fair price of an investment. It considers factors like future cash flows, time value of money, and the risk associated with the investment to understand its true value.
• Econ 101: A glorious world of faster productivity and economic growth, propelled by AI, would be associated with new investment opportunities that increase the demand for capital, driving up real interest rates. (The opposite of secular stagnation.) If people or businesses anticipate significant economic growth due to AI advancements, they might choose to spend or invest more now, expecting higher future income, reducing savings. (The paper examines historical interest rate data, specifically looking at the yield on inflation-protected securities, to analyze how changes in AI development expectations impact these rates.” Similarly, if there’s a fear of existential risks, or even extinction, from AI, individuals might prioritize current consumption over saving for a future that seems uncertain or bleak. Both scenarios would lead to a decrease in overall savings. As savings decrease, the supply of loanable funds drops, causing real interest rates to rise due to increased demand for these now scarcer funds.

As the paper concludes:

… we do not use any detailed inside knowledge of artificial intelligence technology to provide a forecast of the likely timeline for the development of transformative AI. That is, we do not present an ‘inside view’ on AI timelines. Instead, we argue that market efficiency provides an ‘outside view’ for forecasting AI timelines.

(And what about looking at stock prices, both of AI-related companies or the overall equity market, as a forward-looking indicator? Not so much, apparently. While transformative AI may increase future profits, the economists note, it might also raise the interest rate used to discount them, leading to an ambiguous net effect on stock prices.)

But predicting market reaction is … complicated

Readers might recall a JPMorgan report on this same topic that I wrote about last year. While the economists at JPM also thought the productivity-enhancing effects of AI might put upward pressure on real interest rates, they offered two caveats:

First, by facilitating progress in health research that extends life expectancy, or by automating cognitive work that allows older workers to retire earlier, an AI might lengthen the expected retirement period for workers. This could encourage workers to increase their savings to finance a longer time in retirement, resulting in higher aggregate savings in the economy. Higher savings would increase the supply of loanable funds and place downward pressure on real interest rates. (This effect may be amplified if workers have less confidence in future Social Security benefits, by the way.)

Second, if an AI boom concentrates more income into the hands of capital owners versus labor, it may further boost aggregate savings in the economy if capital owners have a higher propensity to save compared to lower-income groups. The increased savings from this channel would also place downward pressure on real interest rates.

All that said, I’ll be looking for clues wherever I can reasonably find them. Keep your eyes on the bond market!