My Feast-or-Famine AI Bet with Holden Karnofsky

Over the last month or so, I’ve been been hammering out a new AI wager with Holden Karnofsky, co-founder of GiveWell and Open Philanthropy. I think of it as the “Feast-or-Famine” bet. Holden suspects that AI will dramatically change the global economy, but he’s undecided about whether the changes will be dramatically good or dramatically bad. In contrast, I’m skeptical that AI will do anything dramatic anytime soon. I think AI will be like electricity or the internet: another new technology that will takes decades to work its ultimate magic.

As usual, though, this characterization of our disagreement is vague poetry.

The literal wording of the Caplan-Karnofsky bet lies below.

---

Holden wins $5000 if, by the end of 2044, there has been at least one year for which it is plausible — according to Bryan Caplan — that at least one of the following was true:

• Real Gross World Product was (conceptually, as opposed to officially) at least 4x the real GWP of 2024.

• Or: real GWP was (conceptually, as opposed to officially) at most 1/2 the real GWP of 2024.

Otherwise, Bryan wins $1000.

This is an attempt by Holden to operationalize "Stuff is going to get extremely crazy (due to AI)" in a way that captures lots of possibilities, from a catastrophe big enough to cut GWP in half to a productivity explosion big enough to bring it well above trend. Based on historical data, Holden believes that his side of the bet would have lost for all prior 20-year periods in human history; he is expecting something unprecedented to happen in the next 20 years.

Some additional details:

• Bryan will strive to count "weird" production that might not be captured in official statistics — for example, if autonomous AIs are building lots of stuff in outer space, and this is being excluded from official GWP for some reason (for example, because the stuff is neither made/sold by nor consumed/purchased by any humans), Bryan would still take this production into account when adjudicating the bet.*

• The amount due (from the loser of the bet to the winner) will be inflation-adjusted using CPI.

• As soon as the bet's resolution is unambiguous, payment and public concession are due within 3 months.

• If either party dies or becomes incapacitated before the bet is resolved, the financial stakes are void, but the bet will still have an official winner as determined by a backup adjudicators named by Bryan.

• Bryan’s choice of backup adjudicators, in order of priority, are: Alex Tabarrok, Maxwell Tabarrok, and Zixuan Ma.

* An example Holden has in mind would be if misaligned AI systems were producing huge amounts of stuff for purposes of their own or even just for mysterious reasons, and neither using it to enrich humans nor attacking humans. This might not get counted in official GWP if the AI systems weren't considered to be part of any particular country, but it would still be production.

---

P.S. Notice that the terms make me the judge in my own case. I’m humbled by the trust Holden is putting in me, and I will strive to deserve it!