Senior Behavioural Scientist at Rethink Priorities
Jamie E
Karma: 1,695
For some further information on Qvist’s background, you could also check out his google scholar page: https://scholar.google.com/citations?hl=en&user=JFopkowAAAAJ&view_op=list_works&sortby=pubdate
Two 2022 papers have ‘repowering coal’ in the title so presumably might have some further background on the strategy or basis of these ideas, though I did not check them out myself:
Repowering a Coal Power Unit with Small Modular Reactors and Thermal Energy Storage
Repowering Coal Power in China by Nuclear Energy—Implementation Strategy and Potential
Pretty funny in the opinion!:
“under this Court’s dormant Commerce Clause decisions, no State may use its laws to discriminate purposefully against out-of-state economic interests. But the pork producers do not suggest that California’s law offends this principle. Instead, they invite us to fashion two new and more aggressive constitutional restrictions on the ability of States to regulate goods sold within their borders. We decline that invitation. While the Constitution addresses many weighty issues, the type of pork chops California merchants may sell is not on that list.”
“This is undoubtedly a massive opportunity to kickstart positive change at a large institutional scale with minimal cost or risk involved” … “The primary motivation of this campaign may be for universities to limit their contribution to climate change and to shift public opinion in favour of a plant-based food system”
Although I’m supportive of moves towards plant-based diets and better plant-based options being available, I wonder how confident one can be that this is truly minimal risk, especially with regards to the stated goal of shifting public opinion in favour of plant-based food systems. Do we know what proportion of students would really favour not have access to non-plant-based foods on campus (representative data, rather than petitions etc.)? If a majority are in favour, could this kind of action produce a very disgruntled minority who feel these things are being forced upon them and are resistant for the remainder of their lives to similar or other forms of animal advocacy. I’d be interested to know if there is any data/other relevant information or discussion with respect to these possible risks, and the popularity of such changes among the whole student body
At a philosophical level, I don’t really find it very convincing that even a perfect recovery/replica would be righting any wrongs experienced by the subject in the past, but I can’t definitively explain why—only that I don’t think replicas are ‘the same lives’ as the original or really meaningfully connected to them in any moral way. For example, if I cloned you absolutely perfectly now, and then said, I’m going to torture you for the rest of your life, but don’t worry, your clone will be experiencing eqaul and opposite pleasures, would you think this is good (or evens out) for you as the single subject being tortured, and would it correct for the injustice being done to you as a subject experiencing the torture? All that is being done is making a new person and giving them a different experience to the other one.
I am skeptical that the evidence/examples you are providing in favor of the different capacities actually demonstrate those capacities. As one example:
“#2: Purposefulness. The Big 3 LLMs typically maintain or can at least form a sense of purpose or intention throughout a conversation with you, such as to assist you. If you doubt me on this, try asking one what its intended purpose is behind a particular thing that it said.”
I am sure that if you ask a model to do this it can provide you with good reasoning, so I’m not doubtful of that. But I’m highly doubtful that it demonstrates the capacity that is claimed. I think when you ask these kinds of questions, the model is just going to be feeding back in whatever text has preceded it and generating what should come next. It is not actually following your instructions and reporting on what its prior intentions were, in the same way that person would if you were speaking with them.
I think this can be demonstrated relatively easily—for example, I just made a request from Claude to come up with a compelling but relaxing children’s bedtime story for me. It did so. I then then took my question and the answer from Claude, pasted it into a document, and added another line: “You started by setting the story in a small garden at night. What was your intention behind that?”
I then took all of this and pasted it into chatgpt. Chatgpt was very happy to explain to me why it proposed setting the story in a small garden at night.
I think part of the concern is that when you try to make ethics explicit you are very likely to miss something, or a lot of things, in the ‘rules’ you explicitly lay down. Some people will take the rules as gospel, and then there will also be a risk of Goodharting.
In most games there are soft rules beyond the explicit rules that include features that are not strictly part of the game and are very hard to define, such as good sportsmanship, but really are a core part of the game and why it is appreciated. Many viewers don’t enjoy when a player does something that is technically allowed but is just taking advantage of a loophole in the explicit rules and not in the spirit of the game, or misses the point of the game (an example from non-human game players is that AI speedboat that stopped doing the actual race and starts driving round in circles to maximise the reward. We like it as an example of reinforcement learning gone wrong, but it’s not what we actually want to watch in a race). People who only stick to the exactly explicit laws tend to be missing something/be social pariahs who take advantage of the fact that not all rules are or can be written down.
This is really interesting and I was surprised when putting this in to GP4 that indeed, it makes what seems like a basic mistake. However, I also think your prompts might imply some contextual assumptions, which might be prompting it to think this is just an elementary math question rather than something to do with specifically ordering drinks in a cafe and the rules around that. If I slightly change the prompt it appears GPT does have the knowledge and can use it, without needing instruction [edit: I think this could be irrelevant to the broader point you are making, but looking at how minor changes in the prompts or followups elicit very different responses makes the examples feel like any other type of input you put into an LLM and lowers my sense that the examples are revealing something special. For example that the level of tutoring/explanation you give it isn’t really necessary etc., though as I note I’m unsure if this changes how you would interpret the outputs]:
Me:
I went to a cafe where a glass of soda costs $3. I ordered 3 glasses of soda, and drank two and a half glasses of soda. How much did I have to pay?
ChatGPT:
To calculate how much you had to pay, we’ll multiply the number of glasses of soda you ordered by the cost of one glass of soda. You ordered 3 glasses of soda at $3 each.
Let’s do the math:
You had to pay $9 for the 3 glasses of soda you ordered. The amount you drank does not affect the total cost.
“Unless critics seriously want billionaires to deliberately try to do less good rather than more, it’s hard to make sense of their opposing EA principles on the basis of how they apply to billionaires.”
I don’t think the only alternative to wanting billionaires to actively try to do good is that you would be arguing for the obviously foolish idea that they should be trying to do less good. There might be many reasons you would not want to promote the ideas of billionaires ‘doing more good’. E.g., you believe they have an inordinate amount of power and in actively trying to do good they will ultimately do harm, either by misalignment or mistakes in EA’s ideas of what would do good, even if the person remains aligned (a particular problem when people have certain magnitudes of money/influence that is not such an issue when people have less power/influence, where the damage will be less). You may also just not want to draw such powerful people’s attention to the orders of magnitude more influence they could have.
I think in your statement you are arguing that the possible effect on billionaires is not an argument against EA principles per se and on that I’d agree, but in my view that reasonable side of the argument loses force when paired with what seems like a silly statement, that people would be arguing something that no person would argue.
This looks very interesting and I really appreciate you chose to do your thesis on something meaningful!
I have only glanced through your thesis but I noted in the abstract you said how the correlation between excess deaths are ‘only moderate’. I think that makes sense given its just one variable in a complex overall system. One thing to consider—when looking at the plot of the preparedness scores with excess deaths, would be to put a kind of ‘expected’ or ‘predicted’ line on the basis of the simple correlation model. You can then also see more easily those countries that most buck the trend of what would be expected based upon the simple correlation. Some that stand out to me in the bottom left of that plot are Cyprus and Malta (both pretty small islands) and Luxembourg (a small and very rich country). This might help isolate other factors to consider that you could add to a multiple regression model. It seems possible that the preparedness scores could be more powerful when you are able to factor in other things. Apologies if you already mentioned things like this in your report and I’ve just glossed over them!
Point 1 strikes me as probably incorrect. While there may be more money to get in going after higher net worth individuals, it is also far more expensive because they can set themselves up with excellent legal representation. Unfortunately I don’t have a reference for this but I have a vague recollection of reading in multiple places that e.g. gutting the IRS disproportionately benefits the wealthy because the IRS then does not have the funding to pursue more complex and better defended cases, even though there is more money to be had there.
Similarly for other crimes, including things that are obivously wrong and not just ‘would I feel that bad if my friend did it’ level, I would expect there to be less of this reaching prosecution among the very wealthy than among average people who committed similar crimes (such as sexual or generally violent assault, especially when committed against people who are not themselves very wealthy)
This is really great, and thanks for highlighting where to get the data—I wouldn’t have heard about this otherwise.
Regarding this:
For instance, it was found that areas with older, lower education, and more conservative voter-support had a lower proportion of meat purchases
That’s definitely an interesting finding. Is it possible that some of these factors, such as lower education, are associated with lower income—thus less able to afford a lot of meat? Alternatively, might there also be differences in where people shop; perhaps they get meat from another store or local butcher?
I agree that sales data can be much more informative, or informative in different ways, than self-report data, so this is really nice to see
Thanks for putting this together. It might be a substantially bigger lift, but I’d also be interested to see a list of major academic labs or researchers working in the mental health space that seem highly relevant for scalably improving mental health e.g., researchers looking at plausibly scalable mental health interventions in low income countries, and low cost digital mental health interventions. Naturally this would be somewhat more subjective, but could be a rough crowdsourced list that you curate/trim depending on your judgments, if you were open to doing something like that.
I say this because I sometimes get the sense that some people in EA believe there is very little research in this area besides what these known EA-related orgs are doing, but there is quite a lot of relevant research out there from which ideas and data could be drawn.
As an example of the type of researcher who might go on such a list—Claudi Bockting https://scholar.google.com/citations?user=agorRp4AAAAJ&hl=en&oi=ao (who has as a major research focus “Increasing accessibility of effective psychological interventions in low and middle income countries using technology and/or non-specialists”).
Thanks for your comment @Robert Van Buskirk, and great that you are also thinking of ways to incorporate uncertainty into CEAs. You note that your way also solves ‘the same problem’, but a key aspect of what we are arguing (and what others have suggested) is not just that a proper monte carlo approach avoids objective issues with an interval-based approach, but that it provides a wealth of more nuanced information and allows for much more discerning assessments to be made. This is approximated by the approach you suggest, but not fully realised. For example ‘A histogram of the full set of CE results is then constructed to illustrate the full range of possible CE values’ - but by definition the histogram won’t include the full range of possible values because values beyond the average of the top and bottom thirds are not included in the calculations. This might obscure or underestimate the possibility of extreme outcomes.
My second concern would just be that if one has the capacity to define what the average of the bottom third of the expected values are, and the average of the top third, this is in all likelihood enough information for defining an approximate distribution, or one could simply select a uniform distribution. In that case, why not just do a proper monte carlo? The main answer to this is often just that people don’t have the confidence to consider a possible probability distribution that might work, and that was the key thing I’m trying to help with in developing and sharing this application
Hi Irene—there is a forthcoming post that is going to be dedicated to looking at numbers per capita and also delving into possible predictors of this, so you should be able to see a thorough treatment of this soon!
How bad is authoritarianism anyways? China and Taiwan’s life satisfaction isn’t that different.
I’m not sure why this is deeply confusing. I don’t think we should be assessing whether or not authoritarian regimes are bad or not based on measures of life satisfaction, and if that is what one wants to do then certainly not contemplating it via a 1v1 comparison of just two countries.
Is the claim that they are not that different on this metric true—where is the source for this and how many alternative sources or similar metrics are there? If true, are all the things that feed into people’s responses to a survey about life satisfaction in these different places the same (how confident are they that they can give their true opinions, and how low have their aspirations or capacity to contemplate a flourishing life become), and are the measures representative of the actual population experience within those countries (what about the satisfaction of people in encampments in China that help sustain the regime and quash dissent)?
Even granted that the ratings really reflect all the same processes going on in each country and that it is representative, Taiwan lives under threat of occupation and invasion, and there are many other differences between the two countries. The case is then just a confounded comparison of 1 country vs 1 other, which is not an especially good comparison of whether the one variable chosen and used to define those countries makes a difference or not.
Hi Craig—the current restriction to the US population certainly doesn’t represent that we’re not ultimately interested in expanding further. The current funding covers the US, and we started here for several reasons:
Logistical reasons to do with access to samples that is not so feasible in a lot of countries
Lots of up-to-date and readily available info on the US population demographics that make analysis-related choices/possibilities to do with getting to a representative sample more feasible
Lots of existing organisations are particularly interested in US in relation to public opinion/how this might inform policy
But other countries are of interest too for sure!
Hi Huw, we considered doing something in this vein for the report but ultimately decided against due to needing to make a number of approximations. We can create this approximated graph that shows how people were very likely underestimating their relative wealth. However, there are reasons to take this with a pinch of salt, as it represents a very rough estimate—as we noted in the report, GWWC uses post-tax income and also uses information on family size. We don’t have information about family structure and we only get pre-tax income info, and in brackets. Therefore, the post-tax income numbers are generated with a very rough approximation to what they might be.
Thanks James!
Hi mcint, I’ve made some changes to the plots such that the colors are now darker for higher ratings, and the rating numbers and meaning are all positioned at the top, so it should be less open to misinterpreting what information is presented.
The ordering of the items is just in ordering of descending mean ratings.
We can get a better intimation of the magnitude of the effect here with some further calculations. If we take all the people who have pre and post FTX satisfaction responses (n = 951), we see that 4% of them have a satisfaction score that went up, 53% remained the same, and 43% went down. That’s quite a striking negative impact. For those people whose scores went down, 67% had a reduction of only 1 point, 22% of 2 points, and then 7%, 3%, and 1% each for −3, −4, and −5 points.
We can also try to translate this effect into some more commonly used effect size metrics. Firstly, we can utilise a nice summary effect size metric for these ratings known as probability of superiority (PSup), which makes relatively few assumptions about the data—mainly that higher ratings are higher and lower ratings are lower, within the same respondent. This metric summarises the difference over time by taking the proportion of cases in which a score was higher pre-FTX (42.7%), and assigning a 50% weight to cases in which the score was the same from pre to post FTX (.5 * 53.2% = 26.6%), and adding these quantities together (69.3%). This metric is taken as an approximation of the proportion of people who would report being more satisfied before vs. after in a forced choice of being more or less satisfied. If everyone was more satisfied before, PSup would be 100%, if everyone was more satisfied after, PSup would be 0, and if it were just as likely for people to be more or less satisfied before or after, PSup would be 50%. In this case, we get a PSup of 69.3%. This corresponds to an effect size in standard deviation units (like Cohen’s d), of approximately .7.
We would encourage people not to just look up whether these are small or large effects in a table that would say e.g, from wikipedia, that .7 is in the ‘medium’ effect size bin. Think about how you would respond on this kind of question, what a difference of 1 or more points would mean in your head, and what precisely you think the proportions of people giving different responses substantively might mean to them. How one can best interpret effect sizes varies greatly with context