Before moving on to an A.I. question I did get for the Mailbag, I want to answer one I didn’t. Tyler Cowen has been challenging A.I. risk worriers to make some investable prognostications about the looming clouds of doom.
I forecast that he will not find this answer satisfactory, but I think the problem here is that the real-world investment advice I have received from safetyists is mostly bullish. Advice I have taken that has paid off well was stuff like “buy Nvidia.” Advice I have not taken that would have paid off well is stuff like “go get a job at Anthropic and enjoy some stock-based compensation.” I know big safety-heads who advise making large investments in out-of-the-money calls on index funds — i.e., bets that the stock market will increase rapidly. The basic safetyist thesis, after all, is:
A.I. capabilities will accelerate in a transformative way.
This has a nontrivial chance of leading to human extinction.
Item (2) is the basis for policy recommendations but in financial market terms (1) is the dominant consideration. Now of course, mathematically whether you think (2) amounts to a “5 percent chance” or an “85 percent chance” makes a difference. But unless you’re a financial professional, the safetyists and the super-optimists are saying the exact same thing — that A.I. is going to be an incredibly powerful, incredibly transformative technology and not just a boost to coders’ productivity or a new way for kids to cheat on their homework. From an investment viewpoint, the opposite of the safetyist view is “it’s all a bubble,” not “A.I. is going to be really good.”
I think it is a little embarrassing for the safetyist view that it is not more clearly distinguished from A.I. optimism. At the same time, I think the optimists have done the world a disservice by insisting on labeling everyone who has safety concerns a “doomer” and then, rather than understanding what the actual belief is, wondering why the “doomers” aren’t behaving like people who believe the world is doomed.
M_McQ: A lot of people are getting concerned about accelerating AI, including me. I think slowing things down with an international agreement is the only way to avoid a race between companies and nations, leading to a variety of bad scenarios. But even if we could convince the administration that this is a problem, won’t an AI slowdown just cause the AI bubble to burst, tanking the economy? Will political incentives just force us to keep racing and hope for the best?
Ignoring the substantive question about A.I. governance, something that I keep seeing in discussions of A.I. and the macroeconomy is a lot of partial equilibrium thinking, where people look at one change while holding the rest of the economy fixed.
Someone will observe that a very large share of recent G.D.P. growth has been data-center construction, do a little basic subtraction, and then say, “Without the data-center construction, G.D.P. growth would be way lower!” This appears to me to have been the logic behind the Trump administration’s politically bizarre decision to come out swinging as the industry’s greatest friend rather than offer people any kind of reassurances around electricity prices or anything else.
But this is not how a dynamic market economy works.