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*
Miscellaneous things I learned in [econ] grad school:
1. The returns to experience are high(er than I thought)
- Someone who has studied a single topic for a decade or two or three really does know a LOT about that topic
Very happy to share that I'll be joining the University of Virginia as an AP in economics in fall 2025🎉!, after a year as a postdoc at the Stanford Digital Economy Lab (
@DigEconLab
) 🤖. Grateful to many, many folks
I think the Solow/neoclassical model is overrated *for explaining catch-up growth and convergence*
Some thoughts, as a prelude to a thread on Chinese growth accounting:
Related to this
@tylercowen
column, a graph that (frankly) shocked me when I first saw it:
Teen suicide rates drop by 15-20% in summer months when school is out – whereas for older groups, suicides rise then
(A drop in December during winter break, too)
How much of Chinese growth is *TFP growth* vs ‘merely’ Solow catch-up growth?
A mini lit review, for the purpose of attracting real experts to correct my errors 🤠
Tyler Cowen's philosophy papers are, dare I say, underrated – EAs usually seem unaware he
[1] advocated (coauthor w Parfit) a social discount rate of 1, ie """longtermism"""... in 1992
[2] described the wild implications of wild animal suffering before Tomasik/etc (in 2003)
...
People who know the EA world - is Tyler Cohen considered influential therein? I know he wrote a piece with Parfitt but I don't know of much else. Regardless, are there any contemporary economists who are considered active and influential members of the school?
File under "big if true": new Jorda, Singh, and Taylor paper uses trilemma IV and finds that
Tight monetary policy lowers productivity in the (12y) long run (!)
=> Expansionary monetary policy boosts long-run productivity (!)
So money is not neutral?
Happy to see that our piece on "AGI and the EMH"
@ZMazlish
/
@tmychow
won joint first prize in the OpenPhil AI Worldviews contest :)
Working paper version coming very soon -- featuring some new data + new ideas
New post: Recessions are always and everywhere caused by monetary policy
Some macroeconomics, but really some philosophy on *what is causality* and what do we mean by ‘cause’
10. In general, academia/grad school is an oral culture (i.e., not a written culture), and it pays to just talk with people to share information
You can overinvest, but taking the break from research to get coffee with friends and just blab has high EV, in my experience
New post: It was a mistake to switch to sticky price models from sticky wage models
50% niche macro + history of thought
50% an argument for nominal wage targeting
3. All that said though, youth/fresh eyes are still crazy powerful for "thinking about things" ie research
- It's bonkers that a 20-year-old can bring insight to a topic studied for ***decades*** by others
2. Even in the 6-year PhD, you can see slow-growing returns to experience. This feels weird!
Eg only after years of seminars did I begin to feel attending was ever worth the time (compared to just reading the paper). The brain sometimes needs LOTS of training data to rewire
4. It's hard to keep learning new things while doing research
While doing research, it feels like there is a temptation to just grind on the research, and not continually invest in learning new things. Tough balance 🙃
Tyler kindly shares my JMP (thread coming soon!)
2 years ago he wrote of the job market: "fewer papers on macro than 5 or 10 years ago, and very few on monetary economics or crypto"
Half of my research is monetary economics, including a bit of crypto :)
First (credible) evidence that water fluoride is actually good! (afaik)
Geological variation in water sources in Sweden => variation in water fluoride
1. + dental health (big)
2. + labor income
3. No effect on IQ (...mostly?)
I like this paper for sort of formalizing the Eichengreen/Sachs/
@delong
"eyeball diff-in-diff" showing 'countries that left the gold standard earlier had quicker recoveries from the Depression',
by showing that leaving gold => higher inflation expectations + lower real rates
12. Sometimes I think of academia as a royal court:
- clear hierarchy
- reputation/status are omnipresent
- *everyone knows everyone* (not really, but it can feel like it), literal *names* matter a lot
- obviously the exclusivity
- ceremony, patronage, etiquette, court intrigue..
Newcomb’s problem is just a standard time consistency problem
I argue the typical description of Newcomb’s problem is confusing only because it unintentionally conflates *two different possible points in time* from which you could make your choice
One way you can see this: the way experienced faculty have seemingly photographic memories of subsections of papers and interactions in seminars from years ago --
In the same way that experienced NBA/chess players can have photographic memories of games:
@ATabarrok
YIMBYism is a victory for Glaeser and econ research, but also a really astounding victory for *blogging and policy entrepreneurs* --
All hail (MR and) Matt Yglesias and Sonja Truss!
The pressure is to *produce now*, which can lead to myopic underinvestment in human capital
"Fail[ing] to continue to plant the little acorns from which the mighty oak trees grow", as Hamming put it. Extremely real!
@BrianCAlbrecht
@WilliamJLuther
Unstructured time can really screw with people's heads. In that respect, grad school is more like unemployment than high school for a lot of people.
A back of the envelope calculation based on this^:
If US teen suicide rates were as low during the school year, as they are during the summer, then there would be ~2600 fewer teen suicides per year
Some bonus China economy charts from my lit binge:
1) China is spending more on R&D than you would expect based on GDP per capita
(my update of a chart from Wei, Xie, and Zhang 2017)
How much of Chinese growth is *TFP growth* vs ‘merely’ Solow catch-up growth?
A mini lit review, for the purpose of attracting real experts to correct my errors 🤠
Excited to have the one and only
@leopoldasch
at the Digital Economy Lab in conversation with
@erikbryn
*tomorrow* -- Zoom registration below (& in person if you're at Stanford!)
Our fall Seminar Series starts up next Monday 9/30 when Leopold Aschenbrenner joins us for his talk, "Situational Awareness."
As always, our seminars are free, virtual, and open to everyone.
Registration and talk details 👇
Or:
9. At the frontier, it is not clear whether (written) papers or (oral) seminars matter more...
From the outside, you only see the papers -- it looks like academia consists of the papers. On the inside, it's clear there's less reading and the seminars matter a lot
On the experience of grad school:
5. It's more like high school than undergrad: it's a smaller environment, everyone knows each other for years; it can feel like you're less in control of your fate (imho)
UK 10y real rate moving from -0.61% to -0.31% today (and from -2% last month)
Breakeven inflation is basically flat though!
=> the spike in (nominal) gilt yields today is all about about higher real rates, not about higher expected inflation
17. In education, *teaching effort* is that which is scarce
As a student, I would hear education policy recommendation X and think it sounds sensible. As a teacher, I hear X and think about "well that has a benefit, ...but it requires costly effort from the teacher"
7. Academia in general has fractal complexity (this is true of many careers of course): in a lot of ways, the more you learn about what it's like, the more you realize you don't know
(🙃)
💡💡💡NEW PAPER ALERT 💡💡💡
How should firms apologize?
w
@BasilHalperin
, John List, Ian Muir
(forthcoming
@EJ_RES
)
Ans: With caution and with money is what we find after following 1.5 million customers for 12 weeks in a
#fieldexperiment
with Uber 1/n
Strawman: "inflation will go up". Market expectations (TIPS breakevens) obviously say no.
Steelman: "inflation RISK is up. Expected inflation is 2%, but the right tail of the distribution has gotten fatter."
Market expectations: also no, not really!
13. Something that took me longer to learn than I'd like to admit --
Papers are mini textbooks. It's not about having "one big insight" (maybe it used to be). It's about fleshing out *everything* related to your topic (within some bounds)
The new Chetty-Friedman-Hendren-Stepner paper, interestingly, lists "The Opportunity Insights Team" as a full coauthor
They have their own NBER author page too:
More experimentation with authorial credit would be interesting...
Amazing *high-frequency* data on economic effects of 1918 pandemic/lockdowns from
@VeldeFrancois
In short: production, retail ↓ -- but (1) small, especially vs. 1920 depression (first vs. second yellow shading) + (2) no rise in bankruptcy + (3) all entangled with WWI armistice
11. Relatedly: everyone warns you about spending too much time being overly strategic ("just work on your research")
But you can also easily be not-strategic-enough! You see people on both sides of the Laffer peak
Underrated: nominal rigidities are endogenous, not exogenous (sorry Calvo), and making wages/prices/contracts fully state-contingent is desirable
The goal of optimal policy is to make reality look like the RBC model
I’m a total noob in economics, but it looks to me like many problems are caused by the stickiness of wages, yet there seems to be little thinking about how to make wages less sticky. Best guess as to why is that sticky wages are a politically pleasant result for many economists.
The background context here is that US suicide rates have risen substantially in the last 15 years, particularly in relative terms for the <14 group, which makes the study of this – and other “public mental health” issues – especially important and underrated, IMO
it's now been six years since caplan's case against education was published. has anyone written a persuasive counterargument to its core thesis? or has it just been ignored because people don't want to believe it
8. For example -- this was embarrassingly not clear to me before starting the PhD; trigger warning: naivete -- the classes are 90%+ about teching up, not about discussion
Frankly, when I first saw the Hansen-Lang paper, I thought it must be totally noise. But my replication has 15 years of new data on the original paper’s 25, which helped convince me
Also: regressing on a dummy for summer months, the coefficient is stable over time
16. Relatedly, for doing theory research: it was weird switching from my entire previous life where "time" was the biggest constraint to "energy" being the biggest constraint
(though for most empirical/coding work it's still "time" in my experience)
As search costs decline, there are 2 countervailing effects:
(1) mechanically, you have to search less
(2) *you're willing to wait longer to find something good*
This has implications for job search, product variety, but also like -- the dating market!
Higher TFP *causes* capital accumulation:
If you get more productive, you want to (gradually) build more machines
=> AND those machines cause more growth
Should we attribute that higher growth to the machines, or to the productivity?
What about outside the US? Suicide seasonality patterns seem to vary, which seems suggestive of… something, unclear what
Japan – where daily data is available – are particularly depressing: shaded areas are school breaks. Spikes follow each school start
I emphasize that this very much IS a correlation, not a strong argument that “school causes teen suicide”
You could think about identification strategies to tease out causal effects, and hopefully someone will do that (!)
Having fun with some 1970s Fed history today and enjoyed learning that Arthur Burns in 1977 sent the just-inaugurated Jimmy Carter an unsolicited memo on how to fight inflation, with 20 points, and did not mention monetary policy once.
@ZacGross
That the wage Phillips curve is alive while the *inflation* Phillips curve is dead-ish is a direct result of "Friedman's thermostat":
Modern central banks (try to) stabilize inflation, not wage growth -- so there is no equilibrium relationship between inflation and unemployment
That the “summer effect/correlation” seems to have been constant while total suicides have risen, though, is maybe a point against the idea that ‘increasing school-related stress’ is driving teen suicide rates
text has magical properties. it has the built in modularity, reflexivity, compositionality of the natural languages. it is the best latent space evolution ever came up with for inter creature communication
14. On fields. We're still living in the aftermath of the credibility revolution
- We're still picking the low-hanging empirical fruit made available thanks to these tools; and status in the profession is allocated accordingly
plausibly this is efficient ¯\_(ツ)_/¯
My graph for 1980-2019 above replicates/extends
@benconomics
and Lang (2011), who showed this using 1980-2004 data
(My graph above normalizes suicide rates by group average, since younger groups have lower rates)
And though it may be tempting to read this as “school causes mental health issues and teen suicide”, many open questions, especially:
Is this reporting differentials? Are school officials more likely to classify a death as a suicide, compared to family members over the summer?
Minimum wage is fixed in nominal terms
Monopsony lit says *real* min wage is too low
Therefore, Fed should target hyper DEFLATION to raise the *real* minimum wage, since legislators won’t, to raise employment
=> deflation raises employment, QED
The Phillips Curve says that raising unemployment lowers inflation.
According to Krueger/Card, raising the min. wage lowers unemployment.
Therefore, we should lower the min. wage to raise unemployment and lower inflation.
Yes, I go to UChicago. How could you tell?
Is there an easy way to redo this analysis adjusting for the fact that real rates have spiked this year, and TSLA has higher duration than traditional manufacturers? 🤔
Tesla stock seems to have started underperforming other auto companies since its charismatic founder/CEO started spending a large share of his time dealing with highly politicized fights about managing an unrelated company.
[3] had one of the first (?) "impossibility" characterization theorems of the Repugnant Conclusion (1996)
also [4] had an early and very useful analysis of the philosophical trickiness of reparations, way before it was a hot topic (2002)
(and [5] obviously Stubborn Attachments)
(This is a good reminder that interest rates on nominal government bonds are just an absurdly confusing price to interpret, because they mix together a mix of so much... it's unfortunate so much media/policy/academic discourse focuses on them!)
This thread is based on new work with
@ZMazlish
and
@tmychow
, full post here:
We show *real interest rates* would be high if markets were expecting transformative AI
But long-term real rates are low!
Longtermists who want a zero social discount rate should advocate for *subsidies for saving/bequests*
bc capital subsidies cause "shorttermists" to save *as-if* they were longtermists
cf Barrage (2018), Farhi-Werning (2010)
Episode with
@tylercowen
is out.
We discussed how the insights of Hayek, Keynes, Smith, and other great economists help us make sense of AI, growth, risk, human nature, anarchy, central planning, and much more.
Links below.
Enjoy!
18. (A few) things I still want to understand:
(i) The division of labor in different kinds of academic teams. Team production functions vary a lot?
(ii) The life cycle of the academic researcher. How sharply do goals change over time? How much of that is constrained vs. not?
Great scientists like Einstein, Newton, and Darwin had a single year in which they made multiple career defining discoveries.
Was it their youth? Their lack of obligations? The right problem at the right time?
I explore the mystery of the miracle year:
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:
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)
Alwyn Young (1992, 1995) famously argued that the rapid growth of the East Asian Tigers was largely *not* due to productivity growth
Paul Krugman popularized this view in an influential Foreign Affairs article, “The Myth of Asia’s Miracle”, suggesting effectively zero TFP growth
What's the best evidence that """superforecasters""" truly are better at forecasting and not just inputting a larger quantity of time?
Not to be too blunt but -- has anyone replicated the Tetlock-Mellers research agenda?
(this is not a comment on this particular study)
Overrated, since Solow attributes fast catch-up growth entirely to temporarily high capital accumulation – and *not at all* to higher technological growth
Poor countries also have low productivity! Shouldn’t some of the catch-up growth come from simply imitating tech leaders?
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*
So for example, Klenow and Rodriguez-Clare (1997) reanalyze the East Asian Tiger experience: properly accounting for TFP-induced capital accumulation + analyzing Y/L rather than Y
They find that TFP growth accounts for 50-75% of Tiger growth (!), except for Singapore
Something like half of East Asian Tiger growth was from TFP growth – call it 3% per year
As context, US TFP growth over this period was ~1%. So 3% is pretty fast!
=> Tiger growth seems to have been *not* just Solow-style catch-up growth (again, see previous thread)
Convergence is, obviously, a huge topic in the lit and I feel like there must have been more study of this / someone else must have framed the question this way:
“Is catch-up growth due to Solow-style capital accumulation vs. technology diffusion?”
References very welcome!
This is the view of the E. Asian Tiger experience I had accepted, until recently
There are a few issues, with the overriding fact being that: growth accounting is actually just really hard. Empirically tough, but also conceptually
I’ll focus on a conceptual issue (“wonkish”)
From Young’s legendarily careful data work, it was argued Tiger growth was merely capital accumulation: Solow-style catch up growth
The growth accounting from Young (1995) – “the tyranny of numbers” – shows TFP growth explaining at most 30% of Tiger growth
For economists, I wanted to pull out what I think is the most interesting piece from the below post:
Some empirics+thoughts on measuring the relationship between real interest rates and growth
"r vs. g"
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*
Traditional growth accounting gives credit for growth caused by that extra capital *to* capital itself
But it really makes more sense to give (long-run) credit to the higher productivity, which is itself causing the extra capital!
I'm struck by how MMT has actually had relatively little influence on the political system. Bernie Sanders and Elizabeth Warren, for example, both proposed paying for their spending proposals (whether the numbers added up is another question).
Asking "what are the welfare costs of inflation" (as every textbook does!) is a category error: it bakes in an assumption – stable inflation is optimal monetary policy
Instead the relevant question is: "what are the welfare costs of *bad monetary policy*"