Note: I should probably add the usual disclaimer that this is not financial or investing advice etc.
This week, Tyler Cowen raised a sharp critique of the doomer arguments, and even many of the claims amongst us within the broader AI-safety community1. Despite the fairly large pre-existing literature on how to measure cybersecurity costs2, many (though not all) doomers have neglected this. It's not “isolated rigour” to care about basic standards of epistemology3, it's science. Nonetheless, Tyler’s first and second points have already been extensively covered (many long before he posed the questions), so I can offer my best contributions to answering his point 3, which is what I'll be doing for the rest of this post.
First of all, let's reiterate the point as it's central to this piece:
“3. A list of what stocks or other assets you have shorted, now. Obviously if your answer to #2 is sufficiently low, you could answer here zero, as I would do. I expect costs, but not so high that we cannot muddle through and have expected positive stock returns.”
Now this point has by far generated the strongest amount of confusion within a discourse dominated by non-economists. In doing so, in an Straussian manner, the greatest epistemological failure of the doomers in this debate is revealed; their neglect of domain-expertise4 in favour of almost entirely first-principles based reasoning5.
Indeed, the prevailing response is the correct one. The value of a short position where by definition consumption is impossible is zero. So even with a high p(doom), expected utility tends to zero. However, this is the wrong mechanism via which we measure the “doom trade”. Instead, we should be looking at real interest rates.
This is the intuition (simplified, for non-economists6). If you believe you will die soon, you want to consume more now, and save less. Borrowing surges, and this is financed by the assets of non-doomers. Ceteris paribus, interest rates will rise. So if doomers outweigh optimists or the ambivalent, we should see a rise in real interest rates. Indeed, this has been the case for the last three years or so, yet inflation expectations, fiscal and monetary policy7, geopolitical dynamics, and so on confound. However, if the case for imminent extinction was overwhelming, then surely the doomer effect on rates would be obvious prima-facie?
There is also, again, a pre-existing literature on this, which this paper refers to. Note that rates will rise in the most optimistic superabundance scenarios too, yet again we cannot decipher from rates alone which one of the tail-events it is. Hence, if current market valuations imply a doomer scenario, then there's no obvious mechanism to cash in on this.
Yet surely today's doomers should at least change their portfolios? Some have said it cannot reasonably be regarded as hypocritical, or at least irrational, to act in manners contrary to stated beliefs. Again, we're creeping into another debate I have already written about extensively, on whether behavioural economics is a better model than homo economicus8. So let's assume our doomers are rational. Should they change their allocation of assets?
Let f be the fraction of the portfolio allocated to an asset, p the probability that the investment increases in value, g the fraction gained if p, and l the fraction lost if q=1-p. The Kelly criterion states that one allocates according to the following:
f = p/l - q/g
and asymptotically this approaches zero in the doom scenario, when q → 1. Therefore, this actually implies that doomers will not engage in shorting assets correlated to doom positions.
However, to compute this, you need probabilities. Known probabilities. No one knows what value p(doom) takes. Technically, p(doom) should be more accurately described as a likelihood rather than a probability. Therefore applying this criterion is invalid, and a refusal to invest in accordance with your beliefs and expectations cannot save you…
So here are my questions to complement Tyler’s: why is the EMH wrong (if you think markets are being irrational)? Why are the markets wrong (if you disagree with their implied forecasts from valuations)? Why are you not betting against the market (and Kelly's criteron doesn't save you here)?
If you're a doomer and not betting against the market, why are you losing out on money (conditional on you being correct)? Are you correct? To answer these questions, this is exactly why Cowen raised the points he did.
A rigorous method of producing quantitative data to calibrate a logical model, of multiple (possibly offsetting) interactions robust to changes in specification and functional form, is not being petty. It's economics, and science. All other approaches are just mere conjecture, that we should take seriously as extreme tail-risks, yet extreme tails they are. Of course, we should be minimising these risks subject to the marginal costs of doing so (hence why scientific rigour is a baseline necessity), but it doesn't change the fact we'll probably be fine, and the world thinks so too.
UPDATE (13/06/2026): I notice that Tyler's remarks on the quality of this discussion apply to my last post too. What is the marginal cost of preventing cybersecurity incidents, and the net welfare gains or losses from such measures used? How do the total cybersecurity costs compete with the costs of a pandemic as a share of output9? On the internet, I discuss thoughts that are still in the preliminary stage, yet I should be showing my methods and numbers more.
As for biosecurity, the problem with bioweapons is that you can never reliably avoid harm to your own side. This likely limits their adoption, and constrains the existential risks associated. Even if you believe the more preposterous Covid lab-leak conspiracies, China was badly affected too, so any malicious group will be hesitant to use them, and indeed this is why weapons of mass destruction (biological and chemical weapons) are rarely used, despite the malleability of international law enforcement. A realistic base rate is probably around two cases per year, yet nothing on the scale as even Covid. Hence my conditional forecast on this remains that their total costs will not exceed the Covid pandemic.
UPDATE II (16/08/2026): One reason we don't see large doom futures or doom insurance markets (including in catastrophe bonds of the sort discussed in Carson's work on cyber cost forecasting) is that a large proportion of the risk is uninsurable, due to uncertainty on enforcability. Collateralised instruments underprice p(doom), and prices cannot adjust to the values that collaterisation participants would accept.
All these contracts and securities are reliant on courts to enforce them, alongside arbitration mechanisms (e.g. in the event of defaults or payment disputes). However in a doom state, these courts don't exist. Therefore these contracts are typically unenforceable, unless there's a sequential timing decision involved in their sale (where you can hopefully clear before unenforceability occurs). Such risks cannot be precisely measured however, so the rates required to justify such contracts exceed those participants are willing to accept, so the market is almost non-existent. In other words, these risks are uninsurable, so futures and uninsurance market values imply a lower p(doom) than is actually the case. However, the consequence of this is that such markets are not incentive-compatible so (to my knowledge?) don't exist.
In this sense, again the doomers are right. However, this reinforces my earlier notion of the "doom trade" showing up in real interest rates via discounting or intertemporal consumption choices, and so the idea that p(doom) is mispriced in aggregate doesn't really follow.
You see this same pattern in the nonexistence of a futures market in galaxies too. Such optimistic capabilities forecasts are also consistent with much wider variance and larger tail-risks, so again enforcability concerns (in terms of the contracts and the property rights over galaxies) make this market incomplete. Moreover, collateral values would need to be incredibly high to justify such trades on a somewhat esoteric and outlandish bet, and not all participants are willing to provide such amounts.
On a broader note, Imas has made an important point on the division of labour between AI-safety researchers and forecasters vs the economics profession. The two camps have rather different approaches to forecasting counterfactuals. The former (as with most superforecasters) maximise the number of parameters used to inform a reduced-form extrapolation from base rates or trends reliant on historical aggregate series. The latter are explicitly trained (via the microfoundations agenda in response to the Lucas Critique) to avoid this, and specify their models and assumptions clearly.
The former is valid only insofar as we condition on no regime change. However, I think for capabilities, this is surprisingly plausible, as it does seem like a domain where reduced-form works. Scaling trends like Moore's Law have been remarkably stable, and such forecasts can easily be resolved against benchmarks whilst macroeconomic forecasts attempt to predict inherently noisy variables (see the recent debates on GDP and welfare mismeasurement). So stationarity might be a more robust assumption in this case than we may think. Yet it does not hold for modelling the equilibrium economic outcomes with its adoption lags, bottlenecks, comparative advantage and relational dynamics. There you need a structural approach.
This predicts the comparative advantage I see, yet I do wonder if the two domains might be talking over each other? This has become somewhat more of a concern for me this week given the reactions to Tyler Cowen earlier this week on modelling the cybersecurity impacts.
Increasingly, I see it as important to synthesise the two approaches and to bridge the two schools. Let's use division of labour to our epistemic advantage in creating a prosperous world for humanity!
Including myself, albeit I explicitly acknowledged that these rough point estimates should be seen as starting points to more developed forecasts.
Yes there are limitations to these approaches. They are important as these are necessary to model reliable counterfactuals, to reveal isolating assumptions ex-ante whilst doing so, and to formalise the confidence over these results. Fundamentally, science relies on falsifiability, and this is only possible here with a numerical estimate. Please show me better methods for computing probability or likelihood distributions? Nonetheless, if high values of p(doom) are uninsurable, then it follows we lack robust methods of using capabilities forecasts to predict the expected future existential-risk costs as a share of output, contingent on the tail-events occuring.
At least if they're not asymmetrically applied, which was the original point of Scott Alexander's thesis. Tyler holds everyone (including his own side of a debate!) to the highest standards he can, as arguably his entire career is focused on how to bring out the best in people. It's a stark signal on social-media driven affective polarisation that every point not in favour of A is taken as an argument for point B. I mentioned previously that these doomers could never get published in a top economics journal. I think if they find Tyler dismissive, wait until they meet the median economist…
Remember, this was cited as an important correlative variable with predictive accuracy in the famous Tetlock study on superforecasting.
Extrapolation from an aggregate trend of historical data violates the Lucas Critique. The superforecasters have an excellent track record despite this though. In general, the architecture of LLMs is also a challenge to the Lucas Critique, and this is already a debate I have been following on this Substack. I think using MFGs to overcome many of the dimensionality constraints of structural approaches is a promising research direction. One of the main empirical questions is whether this can outperform (Bayesian) VAR identification, at acceptable tractability and costs of compute. This expands the set of structural models you can solve, so is a test of whether computation is the binding constraint to the microfoundations agenda.
I've written an awful lot within my blogging career on issues like the Cambridge capital controversies, whether r* exists, if S=I, heterogeneity, endogenous rates, etc. To abstract from this is therefore necessary or else this post would need a whole data centre, and therefore be rendered useless to readers!
If rates are endogenous, then QT should increase them.
One thing I will add is that, again, current market reactions to monetary policy (in the form of aggregates) are more likely to be observed if the rational expectations hypothesis was correct relative to other approaches to modelling expectations.
This is highly correlated with welfare, yet not identical.