I'd like to add that I think there are ways in which safety work gets done without people working on 'AI safety'. This isn't in conflict with what you said, but it does mean that people who want to work on safety could not go on to work on it but there are still people doing the jobs of AI safety researchers.

It seems plausible to me that a person could end up working on AI and economic incentives push them to work on a topic related to safety (e.g. google want to build TAI -> they want to understand what is going on in their deep neural nets better -> they get some ai researcher to work on interpretability -> [AI becomes a bit interpretable and possibly more safe]). I guess that in this case, the people may not be thinking about safety but if they are doing the jobs of the people that would then I don't think it really matters.

I do think that people should work on AI safety on net, but this seems like a reasonable counterargument.

I haven't read the original paper but I like this post and thought that your video was excellent. I thought it struck a good balance between being nuanced/informative and being engaging/sharable and I'd love to see more material like this from the community.

I thought this was a joke at first but I assume this is the org, thanks for the suggestion!

Thanks for this, I'll be sure to take a look!

Related question:

Given how elusive a vaccine against malaria has proved to be how excited are you about the self-amplifying RNA platform (saRNA) vaccine developments (Yale School of Medicine)?

Thanks for writing this, it was really enjoyable to read and I don't see very many short posts like this on the forum.

I agree with most of the points made in this article and I have come across people who have recently come across EA who have had similar reactions to the first paragraph.

I do think that there is a set of people for which the 'vast differences in the goodness of different choices' insight (which can lead to this problem of not feeling like you are doing the maximally good thing) is still very valuable. I don't think this post argues against this but there is probably a trade-off between stressing this point and not making people bad that the thing they are doing may not be maximally good.

I think this post already has a link to that excellent post in the disclaimer in the third section.

Given this bias, this convenient convergence may not be as strong as my intuition tells me.

Thank you so much for giving really actionable feedback. I'll add those notes but just address them here too

Percentage of the consumption decrease due to decreased smoking > prevalence" (cell A9) and "Percentage of lives saved due to quitting" (A14). Not sure exactly what those are referring to at the moment.

Percentage consumption decrease ... is referring to the fact that some people will reduce their intensity and some people will quit (which we call a reduction in smoking). In our model quitters will experience improved health outcomes and people that reduce won't (this is partly us trying to be conservative as we struggled to find good studies looking at the link between reduced intensity and health outcomes).

Percentage of lives saved ... is because not all quitters will have improved health outcomes. Some will only quit after the damage has been done and there will be marginal differences to their exposure to risk. In the model this just discounts the effect of quitting.

I'll have a think about how to go about coming up with more pessimistic guesses for costs/attribution etc. In any case, the feedback is useful. Our model admittedly just goes for a low number guided a by the opinion of some experts in the tobacco control space that we spoke to. Maybe in future we should look at surveying experts a wider variety of experts or using some kind of prediction market as this is a pretty key part of the model.

Apologies for not replying to this earlier but thank you for posing your question.

I had a few thoughts on the points that you made about why consumer surplus might not be a small factor.

Many smokers report enjoying the experience of smoking.

This may be true but I think these people are by far in the minority. Tobacco taxes do not target people who enjoy smoking but are of course applied to the whole population (assuming full compliance).

Many people choose to smoke despite knowing about the health effects.

I am not entirely sure what this claim is based on. I think that this may well not be true and may be less relevant to this project for a few reasons.

1. The health effects can be hard to understand, partly because of the tobacco industry pushing pseudo-science to prevent tobacco control laws from being passed.

2. Even if people do fully understand the effect of addiction means that people keep choosing to smoke even if they want to quit. My understanding is that the majority of smoker do want to quit.

3. In many LMICs, I think this is an unreasonable expectation particularly when the tobacco industry pedals misinformation around the harmful effects of smoking and that many countries don’t have warnings on packages (unlike in the UK/USA).

Newer forms of tobacco consumption, like vaping, have significantly lower health side-effects.

So far the evidence does seem to point that way. There is still considerable debate around whether nicotine delivery systems are helpful or not (particularly in countries which aren't equipped to regulate them) but generally, e-cig adoption is lesser in LMICs and tax structures can penalise more harmful products over lesser ones to encourage switching which would mitigate this risk.

Rational choice is still possible in the presence of addiction - see for example Becker and Murphy (1988).

I have not read through this paper fully but I have a few remarks for it’s applicability to tobacco control (please do let me know if you think I misunderstand the paper).

1. Tobacco taxes reduce the number of people who smoke partly by reducing the incentive for people to take up smoking. This paper seems to examine only the case where the person is already addicted and is choosing whether to go ‘cold turkey’ to quit not whether a person should become addicted. I think given the above points it is sensible to say that the higher expected value option in most cases is not to start smoking.

2. The paper states that "present-oriented individuals are potentially more addicted to harmful goods than future-oriented individuals" but does not seem to say that it is rational to give in to addiction, more that given people’s individual time preferences some people rationally become addicted to harmful products (e.g. people with a more myopic view who discount the future) and continue to use them. I don’t think this really makes a statement about whether this behaviour is good or bad in itself, more an explanation of how some but not necessarily all rational agents can exhibit addictive properties (under their definition of rationality).

It is probably also useful to mention the large number of people that experience problems due to second-hand smoke (where the consumer surplus argument probably would not apply).

I don't believe that the killing of Eric Garner is really attributable to increased tobacco taxes. Even if it were I am not sure that similar cases are common enough to outweigh the huge number of lives saved by this tax.

I have given some thought to the consumer surplus but have not formally modelled this mostly because I think it would take a long time and I trying to focus my attention on our pilot. If you were to build a quantitative model in guesstimate or similar I would love to take a look!