Thanks for reading and commenting on my post! I have now updated it to more accurately reflect your views. 

Regarding your counterarguments:

1. Great to see we agree on using cost-effectiveness when you can make a credibly good estimate. However, the way I see it, is that the model, as I understand it, takes cost-effectiveness (of sorts) as an input. Cost-effectiveness can be estimated trough essentially 3 techniques: (a) having a think and taking an educated guess, (b) measurement and (c) a Fermi estimate. You can reasonably make the argument that sometimes these options are either impossible or just not rigorous. That much is often true. And from this you would be tempted to dismiss these techniques (or maybe just Fermi estimates). But when applying formal ITN you need a number for solvability. And the ways to get that number are also (a) (b) and (c). So any flaws in Fermi estimates must also be present in formal ITN (a type of Fermi estimate itself).^[1][2]

2. Let me now acknowledge your strongest argument and admittedly one I kinda brushed off without proper care in my post: log returns. I admit I am not familiar with the literature on (log) return functions, so I might as well concede the point that log returns apply rather broadly. However, I don't think this assumption does the model a favour. Assuming log returns allows for formulation the model as an ordinary multiplication of independent factors, sure. And in doing this we go from having to estimate cost-effectiveness directly, to having to estimate the effects of doublings of our resources. But I have yet to see an argument as to why estimating the effects of resource-doubling should improve the predictive accuracy of the model (this might be ignorance on my side). My intuition instead tells me: going from BOTECS to formal ITN you must take into account an additional factor (neglectedness) => another source for error => less accuracy.

   It is true that models are often made better by adding additional informative assumptions. I have just not been convinced this is the case here.

1. ^{^}

   There is an argument you could make that ITN might be preferible to other CE estimates because it is standardized. This would imply that making garbage models is harder because the pitfalls are well known. But this would only give us a reason to have a standardized model, not specifically formal ITN. As I kinda pointed out in the post, there seems to be a whole class of models like formal ITN, that may however have less problems.

2. ^{^}

   I think your arguments here are better positioned to defend the heuristic rather than the formal model.

Hi so I'm definitely one of the people that is in the Lizardmans constant in this case [edit:I just read actually I'm not, only the first attempt counts]. I first took the test honestly and got totalism (as I already expected). But then, since the test allowed it, I took the liberty of retaking it. This time less honestly. I was mostly curious how the website would react to not intended answers. Among the thins I tried are: averagism, the opposite of totalism, the opposite of averagism, maximizing the number of lives (natalism), responding at random and trying to get as many inconsistencies and bitten bullets as possible. I was pleasantly surprised you did not label consistent but absurd views as inconsistent and liked the quiz a quite a bit (if you can tell). I hope you don't mind me contaminating the data a bit.

Hi so two remarks:

1. I would say that talking about a fraction here is too akward. Like awkward enough that I think it would be better to avoid it.

2. Given this, let me outline a ITN inspired model:

   Cost-Effectiveness = Importance x Cost-Efficiency

   or

   B = I * CE

   where:

   B (for benefit) = moral goodness achieved per resource

   I = moral goodness of helping an extra being

   CE = extra beings helped per resource

   With some EA-style ethical assumptions this model becomes informative/true.

   Then your post can be summarized as EAs focusing too much on having a reasonable estimate of CE, while making half-arsed assumptions about "I". I wholeheartedly agree with this criticism. I think we should normalize and standardize focusing a lot more of our effort on carefully assessing "I", just like we have with CE. And I like that you have outlined just how many considerations go into "I". It can include indirect effects, like the meat eater problem, but also moral considerations. Importantly, it should also take into account moral uncertaintiy (not doing so is epistemic hubris). So a full analysis of an intervention/cause area would be:

   cost-efficiency analysis -> indirect effect analysis -> moral analyisis -> profit!

   and in the moral analyisis you have to jointly determine:

   a) Which normative theories you give creedence to

   b) How you'll deal with moral uncertainty

   c) The specific numeric assumptions you'll rely on (like welfare weights for example).

   Or at least this is what my brain tells me should happen. I have been thinking about this a lot and there is a very nasty tradeoff between making a model more likely to be true and making said model more likely to guide action. There are just too many unknowns. Even then, I think we could profit from adapting this structure for reporting our estimates. In short, if somebody's case for an intervention's cost-effectiveness looks like this:

    

   Cost-efficiency:

   CE = a gazillion chickens saved from torture per dollar

   determined by an RCT

    

   Indirect effects = 0

   I made it up

    

   Moral analysis:

   A) Normative Theories = Utilitarianism

   B) Strategy against moral uncertainty = none lmao

   C) Specific assumptions = chickens have a moral weight of exactly 2 because I feel like it

    

   then, even if we agree that their strategy is lacklustre, their assumptions are transparent and it's easier to criticize. And showing off all of the steps that were skipped could incentivize people to skip these steps less (or get serious about defending why they can be skipped). I've been thinking about writing a post about this in the future (I have a post on the formal ITN model coming soon. As I am thinking about writing another one about making new models, I struggle to come up with something that is both rigorous and usable)

3. You seem to think ignoring "I" ("the denominator") drives us to expand our moral circle too much. That might be true, but the effect might be smaller than you think. If we start taking "the denominator" more seriously, maybe the biggest change won't be the moral patients we care about, but rather which interventions we deem more effective at helping them

While some people might notice the fact the companies are making losses, others will notice the fact that they are valued in the billions! I am not sure if a shutdown would actually be the right move but it would definitely send a message.

Thank you for writing this up! I was trying to use the ITN framework to decide in which cause area I personally want to go into. And as an economist with ADHD I couldn't resist the temptation of trying to use 80K's quantitative version. But, while trying to understand it, ...it seemed to make less sense than it should. Now I wonder if I should add my own math to the literature of ITN critiques. Which is actually what made me wonder if such a literature existed in the first place and find your 7 year old post

We are now going towards the back of the platform. Train is delayed (11:06)

We are waiting in front of platform 17 at the main station

As a former trial lawyer, I can say there is only one correct answer to this question. It is: No. A dentist appointment is not a good enough reason to not show up to court (unless it’s somehow a life-or-death emergency, but c’mon). A letter will make you look bad. Failing to show up could get you arrested. Let’s call the dentist.

This is a flawless experimental setup, I love it!

I would have liked n>>1 though. 

How can we donate towards this? Via Effektiv Spenden's Defend Democracy Fund? Or are you working on a US midterms specific fund?

Apart from the gruesome ratios, I found the fact that smarter LLMs are closer to having a well defined utility function quite surprising and maybe a little concerning