Thank you to both evaluators for your comments! We will do our best to incorporate many of the suggestions into any future revisions. A few responses are below.

Seth Benzell

The representative agent of the model is neither a normative goal nor a positive prediction, making implications nebulous.

I agree; perhaps this should be further emphasized. The paper essentially finds that X (a tech acceleration) lowers Y (existential risk) given conditions C (which include [in some sense] optimal policy). Since I certainly find it plausible that C does not hold, the only concrete implication is that one might do well to assume that X lowers Y until one has identified how exactly, say, policy failures have overturned this relationship. But it is a step further removed from a directly applicable result than would be ideal.

Omission of savings from the model eliminates a potential mechanism for intertemporal consumption smoothing, which may impact demand for safety.

Yes, this is an interesting possibility I have thought about a bit but might be worth looking into further. Given that technology is, on most accounts, the only driver of long-run growth, I expect that introducing capital to the model would not much affect the impact of long-term accelerations. But I expect it would worsen the risk implications of short-term accelerations to the extent that risk must be mitigated by specialized capital that must be accumulated in advance (since then an unexpected jump in technology is all the more costly to mitigate), and would improve the risk implications of short-term accelerations to the extent that risk mitigation can be mitigated by general-purpose capital (since then an unexpected jump in technology can be mitigated by capital spend-downs that modestly reduce consumption in the future, instead of sudden large consumption cuts).

The authors argue that historical [anthropogenic]1 existential safety expenditure has been near 0… If I view my enemy as completely alien to my values, wouldn’t military spending count? An even older example could be the… prophets who connected contemporary decadence… to a coming disaster, and recommended shifting resources to ameliorating religious rituals.

i. We make the claim that anthropogenic… existential safety expenditure has risen over time in the context of pointing out that, if one believes existential risk has risen over time, one must attribute it to [a] rise in the baseline hazard rate rather than to a decline in safety expenditures. Unless one believes that military and religious expenditures in the distant past are in fact the reason why anthropogenic existential risk was unlikely in the distant past, I believe the argument stands. Perhaps instead of using the term “safety expenditures”, it would be more helpful to refer to expenditures that seem today in fact to have lowered risk in the past (though this has its own complications).

ii. If one does believe that safety expenditures in the relevant sense have fallen, to the extent that the baseline hazard rate posed by today’s technology is lower than that posed by the technology of the past (and any risk-increase can be attributed to the fall in safety expenditures), the headline results are only strengthened, as noted in a reply to Dr. Bournakis below.

Ioannis Bournakis

I think there is a misleading point in not modeling any benefit function along with the hazard function. Without considering both benefits and potential risks of technology within the same framework, it is impossible to identify whether technology induces existential catastrophes for society…. Technological advancements have contributed to environmental degradation, yet… we can create technologies that are more environmentally friendly.

i. I fully agree that technological advances may lower the hazard rate. I believe we allow for this in Section 2 and in Appendix A.3 (which generalizes Section 3): there is only a hazard function 𝛿(A), but 𝛿 may be locally or indeed globally decreasing in A.

ii. The exception is Section 3, where for simplicity we focus on a particular 𝛿(A) that is increasing in A. The result of that section is that, even here, speeding growth in A lowers 𝛿 after a risk-mitigating policy response, so I do not see how allowing A to lower 𝛿 directly would add anything.

Have the authors considered the case where 𝐴 induces a constant hazard, regardless of how quickly the increase occurs?

Yes. The case where each technology level A is associated with a constant hazard rate, regardless of how quickly the technology developed, is the case we consider throughout Sections 2 and 3. It is only in Section 4 that we consider the case that the speed of technological progress affects the hazard rate. Alternatively, if you are referring to the possibility that the development of technology A induces a fixed amount of risk (not a constant hazard rate per unit of time we linger at technology state A), that is covered by the model of Section 4 with 𝜁 = 1, as discussed at the end of Section 4.2.

Does this statement describe the case where the risks posed by two events occurring simultaneously are associated with certainty, while the sequence of events might induce uncertainty, as we do not know the exact intensity of the risk? “One valid criticism of this functional form is that it overemphasises a channel through which the risks posed by a series of technological developments can be cheaper to mitigate if they occur at once than if they occur in sequence.”

No. This statement is about the fact that in the model of Section 4, a given amount of safety spending this year cuts the risk in (say) half this year, so if you cram in a lot of technological developments this year instead of spreading it out over the years, you lower the risks posed by all of them with the single expenditure.