AGI and the EMH: markets are not expecting aligned or unaligned AI in the next 30 years

Basil Halperin@BasilHalperin

AGI and the EMH: what does the market tell us about AI timelines? TLDR: Markets are not putting much probability on the development of transformative AI (aligned or unaligned!) in the next 30-50 years – as evidenced by low *real interest rates*

forum.effectivealtruism.org

AGI and the EMH: markets are not expecting aligned or unaligned AI in the next 30 years - EA Forum

Basil Halperin@BasilHalperin

(By “transformative AI”, we mean the truly radical changes envisioned in the EA/AI safety communities: - a 10x increase in GDP growth — as Davidson (2021) notes, on the scale of the industrial revolution – or - AI-induced human extinction, ie the “unaligned AI” of Yudkowsky/etc)

Basil Halperin@BasilHalperin

Real interest rates are interesting here because -- Real rates are high when (1) growth is high, or (2) probability of death is high SO: real rates would be high if markets were expecting (1) aligned AI (explosive growth), or (2) unaligned AI (existential risk)

Basil Halperin@BasilHalperin

(1) Why high expected growth pushes up real rates: If I expect to be astoundingly rich in 2040, there’s no reason to save today, because I’m going to be rich => lower supply of saving pushes up the real rate

Basil Halperin@BasilHalperin

You can see this even in a simple cut of the data: strong correlation between real interest rates and future growth [AFAIK this simple stylized fact is novel – unlike us, existing lit doesn’t use real rates from inflation-linked bonds (!), and so measures real rates very badly]

Basil Halperin@BasilHalperin

(2) Why high existential risk pushes up real interest rates: If I expect a high probability of being dead in 2040, I'm not going to be willing to lend much today, because I probably won't be around to enjoy the payoff of the loan => lower supply of lending pushes up real rate

Basil Halperin@BasilHalperin

Again: either of these – future GDP explosion or extinction risk – would be much more likely if the market was expecting transformative AI => real rates at (say) a 30y horizon would be high, if market was expecting TAI in the next 30 years But 30y real rates are low!

Basil Halperin@BasilHalperin

- US 30y real rate on TIPS 1.4% - UK 30y real rate is 0.7% - even UK 50y real rate is 0.7% (!!) - other countries with real bonds are similar => real rates are low at 30-50 year time horizons => markets are not forecasting transformative AI on a 30-50 year time horizon (!)

Basil Halperin@BasilHalperin

In particular, using the simplest possible model, we show that markets are decisively rejecting the shortest possible timelines of transformative AI in 0-10 years: real rates would be absurdly high if the singularity were *that* near

Basil Halperin@BasilHalperin

Okay, maybe you think markets are not that efficient, or at least wrong here – maybe put your money where your mouth is then 🤗 There’s easily a trillion dollars on the table, just from shorting US treasuries alone:

Basil Halperin@BasilHalperin

You can EASILY make this trade with a number of ETFs 🤗 we give some specific examples in the post 🤗 This is not *not* financial advice 🤗 (Paul Christiano has said publicly he’s short treasuries) “Get rich or hopefully don’t die tryin”, to paraphrase

Basil Halperin@BasilHalperin

How I interpret this post: a useful OUTSIDE VIEW on forecasting timelines, which complements but does not completely substitute inside view forecasts (e.g. Cotra 2020) => only *one* model among a mixture of models that you should consider when thinking about AI timelines (!!)

Basil Halperin@BasilHalperin

There is a lot of evidence that financial markets are the best information aggregators produced by the universe (so far?) => the (dead simple!) logic here, showing markets are not predicting TAI soon, should be taken pretty seriously

Eliezer Yudkowsky @ESYudkowsky

The only serious debate is whether the Efficient Markets Hypothesis is true about *all* asset prices *all* of the time for you, or merely 99.9999% of asset prices. If EMH wasn't true about almost-all prices relative to you, you could trade arbitrary assets and go up 10% per day.

Basil Halperin@BasilHalperin

I’ll also note that the argument here is strongest re: “transformative” and/or world-ending AI – but this logic does not rule out “smaller” changes, like AI which has an impact equal to that of the internet – still a very large impact!