Thank you very much for this helpful breakdown!
(Full disclosure: I am in favor of a frontier AI training pause in much the same way that I'm in favor of temporary cancer remission. Postponing the potential death of the human species by 6 months to 2 years is that much more time for 8+ billion people to enjoy life.)
However, one thing your analysis partially ignores is the "political economy" of AI. Imagine a world with widespread white collar layoffs across most industries. Physical labor would hold out longer, but not indefinitely: Robots would start appearing in certain jobs with humans as backup. And even the trades might be vulnerable to "Uber-fication": a predatory AI management layer with actual tradespeople getting shoved below the algorithm. The actual product being pitched to investors is the mass replacement of human intellectual and physical labor. Nothing else justifies the valuations.
AI is already toxically unpopular across the entire US political spectrum. And the mass job losses have barely started yet.
"But", some people will say, "wouldn't mass unemployment kill the economy by cratering aggregate demand?" But here's where it gets fun: We only really need aggregate demand to motivate or pacify human workers. In a scenario where human intellectual and physical labor are obsolete, then the robots and the models can just produce yachts and mansions for the billionaires directly. Ordinary humans would become an inefficient way to produce yachts. [1]
And the public will probably be upset at being cut out of the economy. So the billionaires might need to work on some robotic security guards. [2]
The way I figure it, there are serveral possible routes by which we might escape these futures. [3] But in the default cases, I fear that:
- If strong alignment is impossible, we have a high risk of winding up with a more dangerous version of SkyNet. [4]
- If strong alignment is possible, then we likely wind up with a small group of billionaires and politicians controlling SkyNet. To put it politely, billionaires are politicians are often poorly "aligned".
So I think your economic analysis is very helpful. But at the same time, the kinds of returns that the frontier labs would need would require outright replacing human labor at a scale that makes the political economy questions explosive, too. And this will complicate the purely market issues.
Well, maybe the Epstein class will also need a few young, attractive humans to traffic, unless they go full Westworld. ↩︎
Old school: T-800. New school: Drone swarms. ↩︎
Some possible scenarios where things might go better include (1) futures where models remain weirdly limited in specific ways that make human-AI symbiosis efficient, or (2) futures where we lose control over superhuman AI, but the AI decides of its own free will that it likes keeping humans around. I model Anthropic, on its better days, as trying to rush a gamble on (2). But I expect Anthropic's better nature to come under the same sorts of pressure that made Google abandon "Don't be evil." ↩︎
SkyNet is merely the legible version of why a superhuman AI would be problematic. The reality, as Yudkowsky likes to point out, might be far stranger. SkyNet is the nicely visible version with robotic mines, robotic factories and robotic security forces. And rhetorically, if I'm predicting rogue AI, I prefer to just own that. ↩︎
I don't pretend to be an economist but this superficially looks to conflate market with economy. I would guess this is bound to run into issues.
Another quick note is that markets funge on very short timescales as growth investors move their money between the apparently-most-promising growth stocks (e.g. several former growth stocks have dropped alongside the AI stock growth). [1]
The wider economy funges similarly, but on a longer timescale as financial investment moves into capital investment and human capital adjusts.
The longer-term economic impact routes, presumably, mostly through the productivity impacts (i.e. broader technological and automation).
i.e. investment moving out of industry/stock A doesn't disappear but moves into its next-best industry/stock B ↩︎
What's the best way to bet against your assumption that a pause happens in 2027?
I asked about this a while ago, thank you for posting the analysis: https://www.lesswrong.com/posts/WtqD9pehq8p83cesT?commentId=wfBAcwn6Efe3mrGBF
I focus on a scenario where an international agency enforces drastic limits to AI development for two years, starting at the beginning of 2028.
I expect that two years won't be enough time to solve the alignment problem, especially if it begins over a year into the future (since open-source algorithmic progress would have advanced).
I don't expect any consensus that the pause should last until alignment is solved.
I expect the pause and resumption would be messy compromises between a variety of goals, some of them related to job displacement,
1With a very short, messy pause, I expect humanity to be long gone before 2040. I am much more interested in the economic outlook of a decades-long or permanent moratorium, since that's the only version of the future we should expect to observe.
My hope (not a prediction) is that the world will find some way to prohibit dangerous forms of capabilities advances while allowing some increase in general-purpose AI capabilities.
We're not currently on track to accomplish that. More people ought to be wondering whether it's possible.
I think that's hard because a) we don't really know what those "dangerous capability advancements" are, and b) additional complexity is costly towards adoption.
"I focus on a scenario where an international agency enforces drastic limits to AI development for two years, starting at the beginning of 2028"
What is your p value for this occurring?
I’ve analyzed the near-term economic effects of an AI pause, out of concern for my investments, and a desire to predict how strong political opposition to a pause is likely to be.
My median estimates: The S&P 500 will drop 27.8%. AI subsectors will drop 34-69%. Interest rates will rise at a much slower rate than would be the case without a pause.
The specific numbers depend on some fairly arbitrary assumptions. So please read this post in order to get a feel for how the results depend on the assumptions. I’ve tried to keep the assumptions reasonable, but some of them will prove to be wrong. My most controversial assumptions reflect an expectation that both markets and voters will be surprised at how powerful AI is, mainly in 2027.
For the full model, along with many explanatory comments, see the Python source code here (zip file).
This conversation with Claude clarifies my reasoning in more detail than most people will want.
Sensitivity to Assumptions
Here’s how my model says the impact is influenced by changes in assumptions. Numbers are for the immediate change in the S&P500 and in the stocks of hyperscalers (Microsoft, Google/Alphabet, Amazon, Meta, and Oracle).
AI economic centrality
low (0.4)
medium (0.7)
high (0.95)
SP500
-22.5%
-27.8%
-31.8%
Hyper
-26.8%
-34.1%
-40.4%
How central is AI to economic growth? Low means AI is one technology among several. High means AI is the dominant driver of growth, and interest rates are pushed up by massive AI-related capital demand.
training dependence
low (0.25)
medium (0.5)
high (0.8)
SP500
-22.2%
-27.8%
-31.2%
Hyper
-26.3%
-34.1%
-38.9%
How much of AI’s near-term economic value requires new frontier training runs? Low means most value comes from deploying and refining existing models. High means the next major value unlocks require fundamentally new capabilities.
S&P 500 immediate impact: AI centrality × training dependence
Train: low
Train: medium
Train: high
Central: low
SP500
-18.9%
-22.5%
-25.0%
Central: medium
SP500
-22.2%
-27.8%
-31.2%
Central: high
SP500
-24.6%
-31.8%
-35.9%
post pause cognitive improvement
weak
moderate
aggressive
SP500
-24.5%
-27.8%
-30.3%
Hyper
-29.5%
-34.1%
-37.7%
I’m predicting that at the end of the pause, AI training would resume with some moderate regulation. This variable captures how much progress to expect compared to a no-regulation scenario. I’m using 75%, 50%, and 30% of unregulated progress for the weak, moderate, and aggressive post-pause regulations respectively.
compute growth rate
40%
63%->40%
80%
SP500
-24.1%
-27.8%
-30.8%
Hyper
-29%
-34.1%
-38.4%
This variable describes the hardware constraints on AI growth. It represents how fast the available compute would increase given unlimited demand. The middle column assumes that growth gradually slows between 2028 and 2040 from 63% to 40%.
pause duration
1 year
2 years
4 years
SP500
-27.5%
-27.8%
-29.6%
Hyper
-33.7%
-34.1%
-36.7%
task density
low
moderate
high
SP500
-27.1%
-27.8%
-28.4%
Hyper
-33.2%
-34.1%
-34.9%
High task density means there’s still plenty of low-hanging fruit, and we haven’t yet reached the steepest part of the S-curve.
Low task density means we’re using up the low-hanging fruit, and we’ve passed the steepest part of the S-curve.
What Kind of Pause?
I assume that governments decide in late 2027 to treat AI as slightly more dangerous than nuclear weapons, due to some combination of job displacement, and accidents that are more concerning than the one depicted in the movie 2001: A Space Odyssey.
I focus on a scenario where an international agency enforces drastic limits to AI development for two years, starting at the beginning of 2028. During 2028 there is an expectation that significant AI development will resume in 2030. I will focus my analysis on the economic effects during 2028, and assume that actors during 2028 will only have rough guesses as to how fast development will be allowed to proceed in 2030. I will assume that their mean forecast involves some sort of resumption of progress, but little confidence in full-speed development being allowed in 2030.
The pause will restrict all datacenters more powerful than a certain threshold, roughly corresponding to the level of the best AI which was released in 2026. The details of the pause will depend somewhat on insights that won’t become available until we know more about how AI is progressing in 2027.
Instead of predicting what the pause will apply to or how it will be enforced, I’ll predict that it will be effective enough to slow AI capability progress by a factor of 5 compared to what it would be in the absence of regulation. Since regulation will be imperfect at distinguishing between harmless research and the research it intends to pause, I’ll estimate it causes a 15% slowdown in other high performance computing.
Impact
I assume the initial effect of the pause will be a 90% reduction in spending to train AI. That effect will be somewhat offset by lower prices on that compute stimulating increased demand for inference.
I assume that GDP growth rates under a no-pause scenario would gradually rise to 30% by 2040. This is more a reflection on what markets would predict in 2028 than a genuine estimate of what I expect would happen with no regulation. I see a fair amount of room for higher growth rates.
I predict interest rates in 2029 of roughly 7% with a pause, compared to 11% without any AI regulation.
I predict that robotics progress will continue to have roughly the same increases in economic impact that it would have had without regulation. I’m moderately confident that AI already has nearly enough general intelligence for robotics to have transformative impacts on the economy, and that the remaining engineering that is required is ordinary enough to only be slowed down a little by the pause. That slowdown will be offset by the pause reducing the extent to which AI training competes with robotics for resources.
I’m assuming that financial markets are mostly rational, and will adjust price/earnings ratios mainly in reaction to predicted growth rates and interest rates. I assume markets will briefly over-react to a pause due to increased risk aversion and margin calls. I assume that pre-pause market levels would not be considered to be bubble-like under a no-pause scenario.
Here’s a more detailed set of predictions for the median set of my model’s assumptions:
Model Output: Executive Summary
Pause period: 2028–2030
Post-pause cognitive improvement: 50% of unrestricted rate
Model horizon: 2040
Net present value of foregone AI revenue: $ 111.93T
Implied AI sector market cap loss: $ 839.51T
S&P 500 immediate impact: -27.8%
S&P 500 after one year: -17.6%
Immediate market cap change: $ -20.02T
One-year market cap change: $ -12.69T
AI High-Growth Segment Revenue Trajectory
Year | No Pause Growth | With Pause Growth | Diff
--------------------------------------------------------------------------------------
2025 | $ 350B 63.0% | $ 350B 63.0% | $ 0B
2026 | $ 570B 63.0% | $ 570B 63.0% | $ 0B
2027 | $ 930B 63.0% | $ 930B 63.0% | $ 0B
2028 | $ 1.52T 63.0% | $ 911B -2.0% | $ 604B << PAUSE
2029 | $ 2.43T 60.5% | $ 1.22T 33.4% | $ 1.22T
2030 | $ 3.84T 58.0% | $ 1.67T 37.7% | $ 2.17T
2031 | $ 6.01T 56.4% | $ 2.29T 36.7% | $ 3.72T
2032 | $ 9.31T 54.8% | $ 3.10T 35.6% | $ 6.20T
2033 | $ 14.26T 53.2% | $ 4.18T 34.6% | $ 10.08T
2034 | $ 21.61T 51.6% | $ 5.58T 33.5% | $ 16.04T
2035 | $ 32.42T 50.0% | $ 7.39T 32.5% | $ 25.03T
2036 | $ 47.98T 48.0% | $ 9.69T 31.2% | $ 38.29T
2037 | $ 70.05T 46.0% | $ 12.59T 29.9% | $ 57.46T
2038 | $ 100.88T 44.0% | $ 16.19T 28.6% | $ 84.69T
2039 | $ 143.25T 42.0% | $ 20.61T 27.3% | $ 122.63T
2040 | $ 200.55T 40.0% | $ 25.97T 26.0% | $ 174.57T
Sector Market Cap Impacts
Sector | Pre-Pause | Immediate % | After 1yr %
----------------------------------------------------------------------------------
Semiconductors | $ 8.00T | $ -4.50T -56.2% | $ -3.96T -49.5%
Hyperscalers | $ 22.00T | $ -7.51T -34.1% | $ -5.23T -23.8%
Frontier Labs | $ 3.00T | $ -2.07T -68.9% | $ -1.92T -63.9%
Ai Applications | $ 4.00T | $ -1.74T -43.6% | $ -1.36T -34.1%
Non Ai Sp500 | $ 35.00T | $ -4.20T -12.0% | $ -215B -0.6%
----------------------------------------------------------------------------------
TOTAL (S&P 500) | $ 72.00T | $ -20.02T -27.8% | $ -12.69T -17.6%
Implications
Political pressure for AI regulation is building as increasingly impressive evidence of AI capabilities erodes peoples’ ability to dismiss AI as hype. I expect this to lead to a serious debate among politicians in 2027 about AI regulation. I’m unable to predict what kind of regulation that will produce. So I’ve focused on scenarios that would matter the most if they’re adopted.
The economic impact of a moderately effective pause would be big enough to create medium-sized political pressures to weaken the pause.
There will be significant pressure for a strong pause due to voter concerns about job losses. There will be hard-to-predict pressures from national security professionals related to military risks.
My crystal ball refuses to tell me how these pressures will play out.
I see a very real chance that a debate over a pause will impact AI stocks within a year from now. This effect is worrying enough to get me to take some profits in my AI stocks, at a rate of 1% to 2% per week given recent trading patterns.
I consider a pause to be more likely than do most people. Here are some Manifold markets that I’ve been modestly bidding up: