Froolow

Karma: 906

Froolow 29 Jan 2026 9:11 UTC
3 points
0 ∶ 0
in reply to: Brendan Phillips’s comment on: AI Red Teaming at GiveWell: What We’ve Learned (and Where We’d Welcome Your Input)
We did a conversion from Excel to R in 2025, although not LLM assisted. It took about three months, which I would say was a disappointing but realistic downside-case timeline (we basically budgeted for two months and something complex came up). So if you’re finding you can get it done in weeks rather than months with an LLM you’re actually a long way ahead of the state of the art.

Froolow 27 Jan 2026 21:01 UTC
2 points
0 ∶ 0
in reply to: david_reinstein’s comment on: AI Red Teaming at GiveWell: What We’ve Learned (and Where We’d Welcome Your Input)
One of the big recurring arguments in pharma cost effectiveness modelling is whether to use databases and scripts with a proper statistical programming languages like R, or whether to stick with Excel / Sheets for our models. The advantages of proper languages are manifold, including—as you’ve pointed out—that you can probably use LLMs on them more successfully to audit and augment your code. However the advantage of spreadsheets is that they are extremely portable, meaning almost anyone can run them natively on their laptop and anyone can understand how they work if they want to change parameters. This matters a lot if transparency is a goal, and transparency is often such an overriding goal that we pick Excel over technically stronger languages. So I’d caution that in converting the existing models to databases / scripts you’re actually making implicit decisions about the nature and audience of the model far beyond just whether they are legible to LLMs or not.
I mention this because I’m about to get nerd-sniped by the problem and want to make sure the limitations of the approach are well understood before I get sucked into it!

Froolow 22 Oct 2023 11:42 UTC
4 points
0 ∶ 0
in reply to: GiveWell’s comment on: A critical review of GiveWell’s 2022 cost-effectiveness model
Really happy you found the critique valuable. I agree with everywhere you’ve said the issue is too small to be worth addressing, and also with the correction that because AMF and SMC used different data sources it was legitimate not to explicitly model indirect deaths for the AMF (I’m also pleased you’ve standardised your methodology in the updated model, because this was quite a big discrepancy between approaches).
I don’t totally agree with the other areas where you’ve said you’re confident there is no error at all. I think a better interpretation might be that you’ve got a reasonable simplifying assumption that isn’t worth adjusting. For example, you say that it is appropriate to use different methodologies to account for the spillover effects of cash transfer vs income increase from deworming, which results in larger effective household sizes for cash transfer than for deworming. This is because deworming increases income via improving the earning potential of one member of the household, so there needs to be an adjustment for multi-earner households. Obviously you’re the topic area experts so I’d defer to you on whether or not this is reasonable, but the way you’ve implemented this seems contestable to me in the abstract; around 40% of the benefits of GiveDirectly are assumed to be accrued in the form of dividends on a long-term capital investment, but those benefits are distributed across the entire family of 4.7 for the entire ten-year period of the intervention, whereas in the deworming models you explicitly downweight family size to account for dependents leaving the household over time and the possibility of having two earners in a household. It seems more defensible to me to use a constant family size over the two interventions, and instead adjust the number of consumption doublings the deworming intervention provides to account for some families having two wage earners (so a doubling in individual income from the intervention doesn’t translate to a doubling in household income in the model). That is to say; either approach seems completely reasonable in isolation, but using a different approach in the two interventions favours GiveDirectly (because a given transfer spread across more people results in more utility overall, due to utility being a function of log income increase). In hindsight it is probably a bit harsh to describe this as an ‘error’ by GiveWell because it is a good-faith effort to capture the most important part of the model dynamics, but I also don’t think it is unambiguously correct as you’ve indicated in your review document; there’s a genuine inconsistency here. Overall I feel the same about most of the other areas you’ve highlighted—I was probably a bit over-harsh describing them as ‘errors’, but nevertheless they are potentially risky methodological steps which seem to have simpler and less risky alternative implementations.
Also on a small point of clarification—your comments in the linked document suggest that you are not worried about the hard-coded cells because they don’t have any impact on calculations. Sorry if I wasn’t clear in the text, but the main impact of hard coded cells is that the model becomes really hard to maintain—you won’t have an intuitive sense about whether you can make certain changes or not, because certain calculations you might expect to take place actually won’t happen because of the hard coding.

Froolow 24 Jun 2023 19:59 UTC
4 points
1 ∶ 0
in reply to: mako yass’s comment on: The Common Sense View on Meat Implies Some Pretty Radical Conclusions
That’s really interesting, and honestly pretty surprising—I’d really have to quite radically change my view if it turns out Freedom Farms have found a way to raise >5000 animals on one farm in conditions which are broadly acceptable. If I understand you correctly you’re saying that each individual farm could plausibly be much smaller than 5000 animals though, which I would still find interesting that there’s a way for the system to produce meat in aggregate without atrocity-level cruelty, but less challenging to my existing worldview because I think it is the ‘factory’ element of factory farms which is what drives them to be especially cruel.
I’d be very interested in anything you can find on the distribution of farm sizes—or if you can wait a week or two for me to get some work deadlines out the way I’d also be happy to investigate myself and get back to you.

Froolow 17 Jun 2023 10:03 UTC
4 points
0 ∶ 0
in reply to: mako yass’s comment on: The Common Sense View on Meat Implies Some Pretty Radical Conclusions
I’m a bit confused. It answers your question unless you believe there are farms with more than half a million chickens / 5000 pigs under farm at a time which are not ‘Factory’ farms. Do you believe such farms exist? Do you have any evidence they exist? If not, in what way has your question not been answered?

Froolow 15 Jun 2023 5:46 UTC
4 points
1 ∶ 0
in reply to: mako yass’s comment on: The Common Sense View on Meat Implies Some Pretty Radical Conclusions
Yes; the Environmental Protection Agency uses various criteria to distinguish between ‘Animal Feeding Operations’ (AFOs) and ‘Concentrated Animal Feeding Operations’ (CAFOs, aka Factory Farms). Within this, there is further subdivision between small, medium and large CAFOs. The definition of a ‘large CAFO’ relies exclusively on the number of animals in that AFO so you can confidently identify a ‘large CAFO’ using the Sentience Institute methodology. This will undercount the true number of factory farms, since it will miss eg some ‘Medium CAFOs’ which need to be a certain size AND meet some other criteria about how they handle sewage, but since most factory farmed animals are farmed in Large CAFOs it doesn’t make much difference.

Froolow 14 Jun 2023 17:44 UTC
13 points
7 ∶ 0
in reply to: Larks’s comment on: The Common Sense View on Meat Implies Some Pretty Radical Conclusions
To be clear the Sentience Institute itself is beyond reproach, describing their approach as being “We estimate that 99% of US farmed animals are living in factory farms at present”, which is totally unambiguous.

I’m quite sympathetic to the idea of moral arguments treating the basic unit of ‘meat’ as being the animal—that seems to be the morally relevant unit

Froolow 14 Jun 2023 11:37 UTC
52 points
10 ∶ 1
in reply to: Arturo Macias’s comment on: The Common Sense View on Meat Implies Some Pretty Radical Conclusions
I researched this fairly extensively a few years ago, and it is a true (but maybe misleading, depending on context) claim.
The usual source for this claim is the Sentience Institute, although if you go to a huge amount of effort to check government records by hand you get basically the same number so I’m not worried that the source is somewhat biased. They get the 99% number by using USDA data on the size of farms, and then defining any farm over a certain size as a ‘factory’ farm. This makes sense to me, and is how I’d approach the definitional problem unless I was shown extremely compelling evidence of a farm processing eg 5000 pigs a year using traditional ‘mom and pop’ techniques.
The reason the claim might be misleading is that it is using ‘meat’ as a shorthand for ‘meat animals’ rather than eg ‘carcass weight’. Because the vast majority of farmed animals are chickens, and chickens are overwhelmingly factory farmed when farmed, the result of the Sentience Institute methodology is that it appears the overwhelming percentage of farmed animals are factory farmed. In fact, by carcass weight it is ‘only’ about 90% of meat which is factory farmed.
This could in theory drop a bit lower if you say that the process for factory farming cows is not all that morally relevant for the ²⁄₃ of their life they spend in pastures and hence were very exacting with your definitions (ie maybe for the sake of argument we would say something like “85% of meat-by-weight is farmed in a way that would be extremely distressing for the animal” rather than “99% of meat is factory farmed”), but it is hard to get very much lower than this because pigs and chickens are almost exclusively raised in cramped factory conditions and also make up a great deal of the meat we eat.

Froolow 14 Jun 2023 11:10 UTC
5 points
0 ∶ 0
in reply to: Elizabeth’s comment on: Change my mind: Veganism entails trade-offs, and health is one of the axes
Ah sorry, I seem to have slightly misled you. The quote which you attribute to Scott is actually written by me and the co-author of an adversarial collaboration hosted on Scott’s old blog. I’m not the author of the Adventist Health Study linked, much that I wish I was!
If you have questions about the statistics in the adversarial collaboration I’d be more than happy to talk through the approach we used. If you have questions about AHS2, by all means let’s share the work of finding the answer but I can’t promise to be any more help than any other random person you’d pick off the street

Froolow 5 Jun 2023 9:55 UTC
24 points
3 ∶ 0
in reply to: Elizabeth’s comment on: Change my mind: Veganism entails trade-offs, and health is one of the axes
Hi Elizabeth, I’m the co-author of the piece linked above. You’re absolutely right we chose to focus on the omnivore-vs-vegetarian comparison, for a variety of different reasons. However, AHS-2 does have some comparisons between omnivores to vegans. From the abstract: “the adjusted hazard ratio (HR) for all-cause mortality in all vegetarians combined vs non-vegetarians was 0.88 (95% CI, 0.80–0.97). The adjusted HR for all-cause mortality in vegans was 0.85 (95% CI, 0.73–1.01)”. So depending on how strict you are being with statistical significance there’s somewhere between a small signal and no signal that veganism is better with respect to all-cause mortality than omnivorism.
I think AHS is the best data we’ve got on this topic, but I’d be cautious about over-interpreting it. In my mind the biggest criticism is that Adventists are generally more healthy than the typical American (they do a lot more exercise, avoid alcohol and tobacco etc), which leads to extremely pernicious selection bias. For example, it could be that a vegan diet is much healthier than an omnivorous diet if you are the kind of person who spends a lot of time worrying about your health generally, but the risk of getting the wrong nutrients is so high with a vegan diet that it is harmful to people who are not otherwise concerned about their health. So I’m not confident the very slight improvement in overall mortality from switching from a vegetarian diet to a vegan diet can be judged to be a real effect from AHS alone.
On the other hand, I think I would be confident enough in the AHS data to say that it shows that veg*nism does not entail a tradeoff on the ‘years of life lived’ axis. The most conservative reading of the data possible would be that a veg*n diet has no effect on years of life lived, and I think it is probably more reasonable to read the AHS study as likely underestimating the benefits a veg*n diet would give the average person. Obviously ‘years of life lived’ is not the same thing as ‘health’ so I’m not saying this is a knock-down argument against your main point—just wanted to contextualise how we were using the data in the linked piece.
What links here?

Froolow 20 May 2023 8:51 UTC
3 points
0 ∶ 0
in reply to: quinn’s comment on: “The Race to the End of Humanity” – Structural Uncertainty Analysis in AI Risk Models
Thank you, really interesting comment which clarifies a confusion I had when writing the essay!

Froolow 10 Mar 2023 11:02 UTC
4 points
1 ∶ 0
on: How many QALYs equate to a typical human life?
There’s no single right answer here, and several good approaches depending on the application to which you want to put the information. A reasonable estimate is that a baby born today and then given the best life modern society can give it will accrue around 24 QALYs (at a 3.5% discount rate). This is equivalent to 70 undiscounted QALYs, but you absolutely must discount to some extent in this case, because a QALY now is clearly preferred to a QALY in 80 years time.
This value is found by multiplying the life expectancy of a baby born in the UK in 2020 by the typical quality of life that baby will experience each year of their life. Life expectancy is pretty straightforward to calculate (I take it from the Office for National Statistics), quality of life is much more complicated—the standard in the field is Ara, R. and Brazier, J.E. (2010), but this is getting a bit out of date now. https://eprints.whiterose.ac.uk/11177/1/HEDS_DP_10-11.pdf
Obvious problems with the translatability of this approach is that the Ara & Brazier paper only applies to the UK. Different countries will have different profiles of population health (which is obvious) and also different ways of interpreting how health applies to QALYs (which is less obvious). For example, if old age affects your ability to walk easily / comfortably then this might matter more to your QALYs in dense walkable European cities than car-focussed American suburbs (I have no idea if this mechanism is true, just giving it as an example). This will be particularly challenging if you’re trying to calculate the number of QALYs a person accrues in a global health context, because there is limited research in the area.

Froolow 6 Mar 2023 10:54 UTC
3 points
0 ∶ 0
on: Scoring forecasts from the 2016 “Expert Survey on Progress in AI”
Could anyone explain why the experts think Angry Birds will be so hard? It seems like absolutely ideal conditions for reinforcement learning, in the sense that inputs are very simple and there is a very straightforward way of determining how successful each shot is. Is the limitation that it has to be an Artificial Intelligence which succeeds at the problem, rather than a dumb reinforcement algorithm which happens to be really well suited for the task?

Froolow 6 Feb 2023 19:48 UTC
23 points
2 ∶ 0
on: Unjournal’s 1st eval is up: Resilient foods paper (Denkenberger et al) & AMA ~48 hours
I’m one of the evaluators involved in the project (Alex Bates). I wanted to mention that it was an absolute pleasure to work with Unjournal, and a qualitative step above any other journal I’ve ever reviewed for. I’d definitely encourage people to get involved if they are on the fence about it!

Froolow 10 Jan 2023 10:43 UTC
3 points
0 ∶ 0
on: Re: QALYs and DALYs
When you say ‘adding QALYs’ you mean interventions which generate QALYs right? Not taking two interventions which generate QALYs and trying to estimate the combined effect of doing both interventions together?

If the former, there’s a very interesting paper giving an overview of many different methods of generating QALYs here : https://pubmed.ncbi.nlm.nih.gov/7604170/
It is quite outdated now, and some of the assumptions about what does and doesn’t count are controversial, however I still think it is an excellent way of thinking about the variety of different ways in which you might generate QALYs.
The general answer to your first question is, I think, that the most robust methods of generating QALYs are exactly what you would expect—laws mandating seat belt usage, fire alarm installations, vaccinations etc
With respect to air pollution, you’ll see that there are significantly varying estimates for cost-effectiveness depending on exactly the intervention used to control pollution. For example, ‘Coal-fired power plants emission control through high stacks’ is approximately as cost-effective as wearing a seatbelt whereas ‘Acrylonitde emission control via best available technology’ is one of the least cost-effective interventions studied. The references attached to each intervention will give more details on how these estimates were arrived at.

Froolow 23 Nov 2022 10:57 UTC
12 points
0 ∶ 0
on: How bad is life exactly for animals in factory farms?
You might be interested in an Adversarial Collaboration I wrote on this topic a few years ago. My collaborator was a meat-eater who was very strong on finding representative statistics (in fact he wrote the first draft of Section 2.2 to keep me extra-honest)
https://slatestarcodex.com/2019/12/11/acc-is-eating-meat-a-net-harm/

Froolow 23 Oct 2022 9:16 UTC
1 point
0 ∶ 0
in reply to: Dan_Keys’s comment on: ‘Dissolving’ AI Risk – Parameter Uncertainty in AI Future Forecasting
Yes I will do, although some respondents asked to remain anonymous / not have their data publicly accessible and so I need to make some slight alterations before I share. I’d guess a couple of weeks for this

Froolow 22 Oct 2022 21:05 UTC
2 points
0 ∶ 0
in reply to: Ross Rheingans-Yoo🔸’s comment on: ‘Dissolving’ AI Risk – Parameter Uncertainty in AI Future Forecasting
I agree that the arith-vs-geo question is basically the crux when it comes to whether this essay should move FF’s ‘fair betting probabilities’ - it sounds like everyone is pretty happy with the point about distributions and I’m really pleased about that because it was the main point I was trying to get across. I’m even more pleased that there is background work going on in the analysis of uncertainty space, because that’s an area where public statements by AI Risk organisations have sometimes lagged behind the state of the art in other risk management applications.
With respect to the crux, I hate to say it—because I’d love to be able to make as robust a claim for the prize as possible—but I’m not sure there is a principled reason for using geomean over arithmean for this application (or vice versa). The way I view it, they are both just snapshots of what is ‘really’ going on, which is the full distribution of possible outcomes given in the graphs / model. By analogy, I would be very suspicious of someone who always argued the arithmean would be a better estimate of central tendency than the median for every dataset / use case! I agree with you the problem of which is best for this particular dataset / use case is subtle, and I think I would characterise it as being a question of whether my manipulations of people’s forecasts have retained some essential ‘forecast-y’ characteristic which means geomean is more appropriate for various features it has, or whether they have been processed into having some sort of ‘outcome-y’ characteristic in which case arithmean is more appropriate. I take your point below in the coin example and the obvious superiority of arithmeans for that application, but my interpretation is that the FF didn’t intend for the ‘fair betting odds’ position to limit discussion about alternate ways to think about probabilities (“Applicants need not agree with or use our same conception of probability”)
However, to be absolutely clear, even if geomean was the right measure of central tendency I wouldn’t expect the judges to pay that particular attention—if all I had done was find a novel way of averaging results then my argument would basically be mathematical sophistry, perhaps only one step better than simply redefining ‘AI Risk’ until I got a result I liked. I think the distribution point is the actually valuable part of the essay, and I’m quite explicit in the essay that neither geomean nor arithmean is a good substitute for the full distribution. While I would obviously be delighted if I could also convince you my weak preference for geomean as a summary statistic was actually robust and considered, I’m actually not especially wedded to the argument for one summary statistic over the other. I did realise after I got my results that the crux for moving probabilities was going to be a very dry debate about different measures of central tendency, but I figured since the Fund was interested in essays on the theme of “a bunch of this AI stuff is basically right, but we should be focusing on entirely different aspects of the problem” (even if they aren’t being strictly solicited for the prize) the distribution bit of the essay might find a readership there anyway.
By the way, I know your four-step argument is intended just as a sketch of why you prefer arithmean for this application, but I do want to just flag up that I think it goes wrong on step 4, because acting according to arithmean probability (or geomean, for that matter) throws away information about distributions. As I mention here and elsewhere, I think the distribution issue is far more important than the geo-vs-arith issue, so while I don’t really feel strongly if I lose the prize because the judges don’t share my intuition that geomean is a slightly better measure of central tendency I would be sad to miss out because the distribution point was misunderstood! I describe in Section 5.2.2 how the distribution implied by my model would quite radically change some funding decisions, probably by more than an argument taking the arithmean to 3% (of course, if you’re already working on distribution issues then you’ve probably already reached those conclusions and so I won’t be changing your mind by making them—but in terms of publicly available arguments about AI Risk I’d defend the case that the distribution issue implies more radical redistribution of funds than changing the arithmean to 1.6%). So I think “act according to that mean probability” is wrong for many important decisions you might want to take—analogous to buying a lot of trousers with 1.97 legs in my example in the essay. No additional comment if that is what you meant though and were just using shorthand for that position.

Froolow 22 Oct 2022 19:27 UTC
7 points
1 ∶ 0
in reply to: Stephen McAleese’s comment on: ‘Dissolving’ AI Risk – Parameter Uncertainty in AI Future Forecasting
Thanks, this is really interesting—in hindsight I should have included something like this when describing the SDO mechanism, because it illustrates it really nicely. Just to follow up on a comment I made somewhere else, the concept of a ‘conjunctive model’ is something I’ve not seen before and implies a sort of ontology of models which I haven’t seen in the literature. A reasonable definition of a model is that it is supposed to reflect an underlying reality, and this will sometimes involve multiplying probabilities and sometimes involve adding two different sources of probabilities.
I’m not an expert in AI Risk so I don’t have much of a horse in this race, but I do note that if the one published model of AI Risk is highly ‘conjunctive’ / describes a reality where many things need to occur in order for AI Catastrophe to occur then the correct response from the ‘disjunctive’ side is to publish their own model, not argue that conjunctive models are inherently biased—in a sense ‘bias’ is the wrong term to use here because the case for the disjunctive side is that the conjunctive model accurately describes a reality which is not our own.
(I’m not suggesting you don’t know this, just that your comment assumes a bit of background knowledge from the reader I thought could potentially be misinterpreted!)

Froolow 22 Oct 2022 19:21 UTC
7 points
2 ∶ 0
in reply to: Dan_Keys’s comment on: ‘Dissolving’ AI Risk – Parameter Uncertainty in AI Future Forecasting
I completely agree that the survey demographic will make a big difference to the headline results figure. Since I surveyed people interested in existential risk (Astral Codex Ten, LessWrong, EA Forum) I would expect the results to bias upwards though. (Almost) every participant in my survey agreed the headline risk was greater than the 1.6% figure from this essay, and generally my results line up with the Bensinger survey.
However, this is structurally similar to the state of Fermi Paradox estimates prior to SDO ‘dissolving’ this—that is, almost everyone working on the Drake Equation put the probable number of alien civilisations in this universe very high, because they missed the extremely subtle statistical point about uncertainty analysis SDO spotted, and which I have replicated in this essay. In my opinion, Section 4.3 indicates that as long as you have any order-of-magnitude uncertainty you will likely get asymmetric distribution of risk, and so in that sense I disagree that the mechanism depends on who you ask. The mechanism is the key part of the essay, the headline number is just one particular way to view that mechanism.