This parasite could also be present in water without coliform bacteria, so standard "total coliform" tests are likely to miss them out. (link)

I think what you show here is weaker than how this section seems to be about. It's not true that individual case studies aren't relevant, it's just that it's restricting their claim to "in each category, eventually ideas are getting harder to find" where case studies are extremely relevant.

https://spectrum.ieee.org/semiconductors/design/during-the-20th-century-vacuum-tubes-improved-in-a-moores-lawlike-way

(I didn't search for others)

Although, you only think about your case in 17-18 hundreds

I think that more importantly here, TFP growth is used in the theoretical model they are working under. What he tries to do is exactly to show that the model needs alpha to decrease (or be around 0 in this case) because TFP is only growing linearly since 1950.

I think that the arguments in this section are therefore not relevant to the case against the paper. I think that they might be a case against this theory, but that's a wider claim than understood from this text

Well, looking at the data farther back in time we see that it looks more exponential, or like it wants to be exponential (which is the baseline Bloom et al consider) -

and some other countries -

(This is taken from the Long Term Productivity Database). Also, I think that https://wtfhappenedin1971.com/ might be relevant

I think that it's good to mention at this subsection that there are parts of the paper that directly address idea-looking phenomena. Say, V.A on new molecular entities

It was defined this way in Romer (1990)

-> producing each

*second

Should be \(C=(1-s)Y\)

These pages don't show content - the associated elements in the same row

One thing I don't understand here - is the goal to improve decision making in situations with long time until feedback? if so, how do you plan to iterate?

Controversial and I'm not sure if it addresses the core concern.

In the experiment, the participant were told that they were making 10 bets and wouldn't know the result of the bets until all choices were done. So expected theory seems to apply here well enough

I think that this is slightly misleading. It can be framed as choosing to work for a year for $12K instead of for a month for $2K, but these also has a very different cost in terms of work-hours.

Ruby Sass and LibSass are deprecated. Use Dart Sass instead.

https://reactjs.org/blog/2018/06/07/you-probably-dont-need-derived-state.html#when-to-use-derived-state

React Developers warn against using this function naively.

It is a $10B industry. (expensive report)

Perhaps an analysis of the political situation might be relevant.

If there is an inherent (dis)belief in (basic) science as something important to fund for some political affiliations, then that would amount to understandable bias (this study has such an analysis, but I'm sure there are more relevant resources).

If there are major lobbying entities for or against science funding, that could also be a clear sign of bias. I'd expect advocates for science to be weaker and less well organized.

Where exactly?

It's important to note that it's not clear how to actually do that evaluation. We can measure somewhat easily how capable is a scientist in her domain, but I'm not sure how big of a problem that is generally for professionals with similar amount of experience. What we'd like to measure is how a specific research group plan is likely to result in improvements to the world which is generally a hard problem when it comes to basic research.

This has several of the usual drawbacks of large institutions, which I'm not sure are overcome by the benefits of grant-handling productivity and having more public scrutiny.

Inside politics, large bureaucracy and difficulty in making agile changes all hinder efficiency.

Universities have high correlation between PR/academic status and amount of funding. This is not aligned well enough with public goods (I think).

Is this a growing trend? Is the ratio between PhDs/Post-docs to academic positions really growing?

Perhaps it has always been the case. Perhaps it is a feature and not a bug - if we assume that researcher productivity sits on a bell curve, we might not care as much about anything less than the 90th percentile (say) as they may not be worth the cost of the academic positions.

Are there countries explicitly mentioning that kind of reasoning?

What are the counterfactuals to the following examples? Would high-impact research be funded anyway, and thus funding science has high diminishing returns? (my intuition is that this is not the case because funders do not have the capacity or directive to pick the most promising research - unless the only way to estimate primisingness of research is by assessing the researcher herself, which I think that major funders are doing probably okay at)

This seems like a generally good starting point - https://cs.stanford.edu/~jsteinhardt/ResearchasaStochasticDecisionProcess.html

I think that this follows if there are reputational risks or a major group-think problem. However, I think that many successful experts would be better at detecting systemic biases than almost all non-experts and that much of the success of their work lies on this fact.

See Architecting Discovery by Ed Boyden and Adam Marblestone for an example of that kind of a framework used in neuroscience.

I take this to mean personal career risk, where the performers have a concrete research plan which is likely to work but it is not clear whether it will have impact - which is what the program manager is measured on.

and it’s hard for the performer to pull funding from the program manager.

and also, it points at empowering and training PMs as a good alternatives to creating a new ARPA-style institution

There is the recent book by Magnus Vinding- Suffering-Focused Ethics

The argument is misleading here. Calculating the momentum of a flowing river precisely is impossible due to its complexity, while it is possible to make general remarks about specific adaptive traits.

Perhaps an important difference here is that much of physics is more amenable to approximations.

5.2

Maybe the possibility of Drexler's CAIS is an argument toward increasing computer hardware.

Given much more computational resources, but little algorithmic progress, specialized algorithms might prevail and we would find ourselves in the CAIS world, and not in the AGI world.

I have some problems with this argument.

1. This section actually claims that we expect the preference of "having the most offsprings" to dominate.
2. Under sufficiently complex dynamics or cluelessness about the effects of actions on the long term, caring more about influencing the future or having the most offsprings might be less adaptive than other traits which might be much more important for a reason unknown to us.

Common sense morality values are also instrumentally useful for group selection. Expanding the "consequentialist capabilities" might also symmetrically be harmful.

This is incorrect though. Expansionism is mostly considered an instrumental value in and of itself to the people and not to evolution who designed those people.

This is supported by arguments as in paretotopian goal alignment

This is a bit misleading phrasing, I think. The three questions below are:

1. how do we pick the class of problem that the tool will solve or the space of hypotheses it will allow users to probe?
2. how well can you envision or roadmap out the possible solutions so that you can pick the best one possible?
3. once you have chosen a problem and have a roadmap of the space of possible solutions, how do you choose which path is the best?

But these are questions about the general process itself. The suggested three question for taking a topic and architecting discovery there are

1. What is the class of problems that the tool will solve? What is space of hypothesis that it will allow users to probe?
2. What is the complete set of solutions can you envision and what is the best roadmap of achieving these?
3. Among these solutions, considering the roadmap, what is the most promising path?

This is a major crux for me. I'd expect it to be the most likely scenario - a convergence to the preferences of the people alive as civilization gets more prosperous.

Well, I don't really buy this. I expect that some economic incentives will hold and perhaps be even more important than today and not allow focus on well being. Also, something might lock-in before we get a hold of our preferences.

I find this extremely important, and not something I understood before. I imagined a homogeneous future, in which every life is either good or bad, but not one in which there are vast amounts of suffering, regardless of what other accomplishments there are. I find the view that we should not allow 10^25 people to live miserable lives in order for us to have 10^35 happy people. Would I be okay with a rich civilization that expands because of use of slavery? Would I be willing to work for civilization expansion if suffering is likely to dramatically increase with it?

https://longtermrisk.org/tranquilism/

I found that my intuition (which was initially to take (2) ) was shifted more toward (1) after viewing some of the videos on suffering in the link above.

However, I think that perhaps the argument here that torture-level suffering is so horrible and unimaginable lacks the symmetric intuitions of extreme hedonic states. People go through a lot of terrible suffering in order to get another hit of heroin, and possibly wireheading really can be incredible. People having a bad trip still try LSD again because they think that it is worth the risk.

Person affecting view also supports making someone happy at the cost of creating miserable lives, which contradicts the view presented here.

This again can be explained by the intuition that in absolute value miserable is much greater than happy.

Perhaps though, this intuition is exactly the point

This may be explained hedonically via scope insensitivity and an intuition that being miserable is much more negative than the amount of positive utility one has by being simply happy.

If it would be a difference of a million to one, instead of million to thousand, I'd feel horrible not bringing about the existence of another million happy lives.

I think it is the other way around

$$H(p)=H(p,p)$$

they meant training data?

is it actually used in practice? is it preferred over neural networks?

It is converted into Lactate, which is used predominantly in the bioplastic industry.