My point holds across all types of spending. OP's spending on expanding their team should be optimised to ensure the marginal cost-effectiveness of their grants matches that of their internal spending, and that both do not vary across time. I do not know whether OP is striking the right balance. However, I think one is implicitly claiming that OP is making some wrong decisions if one expects the marginal cost-effectiveness of OP's AI safety grants to decrease across time.

I think it is more likely that people do not take my bet because they do not actually believe in short AI timelines.

Thanks for the comemnt, Oscar! Right, I am assuming the cost-effectiveness of donations does not vary much over time. Donors have an incentive to equalise the marginal cost-effectiveness of donations across time. If Open Philanthropy (OP) thought their marginal spending on AI safety in 2025 was more cost-effective than that in 2029, they should decrease their planned spending in 2029 to increase that in 2025. More broadly, money should be moved from the worst to the best years.

Hi Yanni.

Hi Vasco! I'm keen for you to paint me a persona. Specifically; who is the kind of person that thinks sinking 10k into a bet with an EA (i.e. you) is a better use of money than all the other ways to help make AI go better (by making it as a donation)?

If the winner donates the profits, the bet has the effect in expectation of moving donations from the organisations preferred by the loser to the ones preferred by the winner. So the bet would increase total social impact (not just the winner's social impact) under the view of someone who thinks their preferred organisations (e.g. in AI safety) are more cost-effective than the organisations in animal welfare I would donate my profits to.

Even if you were big on bets for signalling purposes, I think its easy to argue that making one of this size with an EA on a niche forum isn't the way to do it (i.e. find someone more prominent and influential on X or similar).

I have been messaging some prominent people who are worried about AI about similar bets, but no success so far.

Do you know about decent estimates for the standard deviation of meat consumption in kg over a given period (like the median delay of 2 weeks among the studies you reviewed) in a given country? One could multiply it by the meat consumption per capita to get the standard deviation in kg, and then multiply this by your effect size to get a rough estimate for the reduction in meat consumption in kg.

Thanks, Seth! I thought you may not have that data easily available because I did not find it in your moderators' analysis in Table 2. Do the effect sizes refer to the whole period of the delay (e.g. 14 days), or just to the last part of it (e.g. last day of the 14 days)? It does not seem clear from section 3.2. I would expect a greater decay if the effect sizes refer to the last part of the delay.

Thanks, Seth. I assume it is also difficult to know at which rate the effect size decays across time. A 3 % reduction in consumption over 1 year would be more impressive than a 3 % reduction over 1 week. Do you have a sense of whether the pooled effect size of 0.07 you estimate should be interpreted as referring more to 1 month than 1 week?

Our overall pooled estimate is SMD = 0.07 (95% CI: [0.02, 0.12]). 

Do you happen to have an overall pooled estimate in percentage points?

Great point, Gregor! Tom Adamczewski has done an analysis which combines the answers to the questions about tasks and occupations. Here is the mainline graph.

Tom aggregates the results from the different questions in the most agnostic way possible, which I think is the best one can do.

I achieve this by simply including answers to both questions prior to aggregation, i.e. no special form of aggregation is used for aggregating tasks (HLMI) and occupations (FAOL). Since more respondents were asked about tasks than occupations, I achieve equal weight by resampling from the occupations (FAOL) responses.

Here is how Tom suggests people describe the results.

Experts were asked when it will be feasible to automate all tasks or occupations. The median expert thinks this is 20% likely by 2048, and 80% likely by 2103. There was substantial disagreement among experts. For automation by 2048, the middle half of experts assigned it a probability between 1% and a 60% (meaning ¼ assigned it a chance lower than 1%, and ¼ gave a chance higher than 60%). For automation by 2103, the central half of experts forecasts ranged from a 25% chance to a 100% chance.²

Hi Rebecca.

I would not agree with that - because time is limited and there are no other options. You could be trying to lobby for outlawing meat consumption instead of advocating for killing people in third world countries.

Advocating for prioritising more animal welfare and mental health over global health and development due to the meat-eating problem, relative to a situation where there was not this problem, is different from advocating for killing people.

I would point out, though, that by your logic it actually would be good to save the life of the suicide bomber - by killing people they'd be saving animal lives.

I think saving the suicide bomber may well decrease nearterm suffering (1st few years afterwards), but I do not know about the overall effect (1st few decades).

If this is really your concern, then would you be equally willing to advocate for denying healthcare to your loved ones? For example, if someone you loved (who was not vegetarian/vegan) was having a heart attack, would you forgo calling 911 in order to save the animals they might eat? Based on your answer to NickLaing's response, it seems like the answer might be yes.

I estimate neutralising the harms caused to poultry birds and farmed aquatic animals per person in 2022 only requires donating 0.0214 $ to SWP. For a remaining life expectancy of 50 years, and constant harm to farmed animals over time, that would imply neutralising the negative impacts linked to saving a life with just 1.07 $ (= 0.0214*50). Saving the life of a person who is close to us can easily increase my donations by way more. Moreover, I would feel quite bad due to not saving them, which would also make me less productive, and therefore have a lower social impact, as I think this is driven by my work and donations.

Additionally, from a utilitarian standpoint it's a bit hard to see the difference between letting someone die and killing them. Do you intend on killing anyone or committing any acts of terrorism?

No. I strongly endorse utilitarianism, and I think there is a huge difference between not saving someone in the other side of the world and killing someone. Utilitarism is all about assessing the consequences, and these are totally different. If I killed someone, I could easily go to prison, and therefore my donations and direct work would decrease a lot, thus majorly reducing my social impact. According to my estimate above, killing someone would only have to decrease my donations in expectation by 1.07 $ for it not to be worth it. In contrast, letting someone die in the other side of the world via not donating has no clear negative consequences for the potential donors. Lots of people buy expensive houses, whereas they could save tens of lives by buying cheaper houses, and donating the difference to GiveWell, which saves a life for 5 k$. However, such people are not arrested or considered anything close to evil.

Will policymakers like to implement a tax based on 'suffering units' with quite some uncertainty? I wonder if we can find a decent proxy that is easier for taxation.

I think it is worth trying. Higher welfare animals tend to grow slower and reach smaller slaughter weights, so taxing animal-years and animals slaughtered will tend to desincentivise welfare reforms, and, if there is a carbon tax, doubly so because they tendentially imply greater GHG emissions.

I agree the suffering of farmed animals is more uncertain than GHG emissions, but I think it may well be less uncertain than the suffering of humans caused by GHG emissions, which I would say is a more relevant comparison. Bressler 2021 estimates "the 2020 MCC [mortality cost of carbon] is 2.26 × 10⁻⁴ [low to high estimate −1.71×10⁻⁴ to 6.78 × 10⁻⁴] excess deaths per metric ton of 2020 emissions". So there is significant uncertainty with respect to whether GHG increase or decrease human mortality. For comparison, below are the time hens and broilers spend in pain as estimated by WFP. I think the bars are supposed to be the 95 % confidence intervals, although I did not find information about this in the page.

The uncertainty respecting the suffering of animals is much larger than suggested by the above due to uncertainty in welfare ranges and pain intensities. Yet, one can at least be pretty sure eating animals causes them pain, and this happens very soon after consumption. In contrast, Bressler 2021 estimates roughly no impacts until 2055.

I still think you have a point because people may overestimate the uncertainty of the impacts on animals, and underestimate that of the impacts on humans. In addition, people may care about other types of uncertainty beyond uncertainty about the direction of the effect, which is the one I think is lower for animal suffering.