Measure is unceasing

RSS Feed, all content

Use a less coarse analysis of AMF beneficiary age and consider counterfactual deaths (2022/09/28)

tl;dr: GiveWell considers a fairly coarse division of beneficiary age, and groups children of 0 to 5 years old together. This may lead to inaccurate or inexact calculations. In addition, GiveWell doesn’t completely account for counterfactual mortality: where a beneficiary is saved from dying of malaria but dies later anyways.

Following up on Use distributions to more parsimoniously estimate impact, I was looking at the population analysis of the AMF distributions, because a previous attempt at adding uncertainty to the analysis was messier than I would have wished.

But following up on that analysis, I realized that the strategy GiveWell uses is:

$5k challenge to quantify the impact of 80,000 hours' top career paths (2022/09/23)

Motivation

80,000 hours has identified a number of promising career paths. They have a fair amount of analysis behind their recommendations, and in particular, they have a list of top ten priority paths. 

However, 80,000 hours doesn’t quite[^1] have quantitative estimates of these paths' value. Although their usefulness would not be guaranteed, quantitative estimates could make it clearer:

Utilitarianism: An Incomplete Approach (2022/09/19)

This blog post gives the sketch of a book, or maybe a long article, that’s been on my mind for a while. I wrote it last week, over the course of an hour an a half, with The Incredibles blasting on the background and with me feeling intellectually alive.

Chapter 1. Utilitarism: The building blocks

This chapter would define utilitarianism, and go over the building blocks of expected utility maximization, like I did in this post but without boring the reader to death. The building blocks are:

Use distributions to more parsimoniously estimate impact (2022/09/15)

Summary

By incorporating uncertainty into its estimates, GiveWell would produce better estimates. This is best done by working with distributions, as opposed to point estimates. For example, “$294 per doubling of consumption” is a point estimate[^1], but the following is a distribution:

An experiment eliciting relative estimates for Open Philanthropy’s 2018 AI safety grants (2022/09/12)

Summary

I present the design and results of an experiment eliciting relative values from six different researchers for the nine large AI safety grants Open Philanthropy made in 2018. 

The specific elicitation procedures I used might be usable for rapid evaluation setups, for going from zero to some evaluation, or for identifying disagreements. For weighty decisions, I would recommend more time-intensive approaches, like explicitly modelling the pathways to impact.

Distribution of salaries in Spain (2022/09/11)

The distribution of slaries in Spain looks as follows:

Salary interval (euros) %
0–13300 5.57
13300–26600 52.27
26600–39900 24.08

Forecasting Newsletter: August 2022. (2022/09/10)

Highlights

Simple estimation examples in Squiggle (2022/09/02)

This post goes through several simple estimates, written in Squiggle, a new estimation language. My hope is that it might make it easier to write more estimates of a similar sort, wider adoption of Squiggle itself, and ultimately better decisions. 

Initial setup

One can use Squiggle in several ways. This blog post will cover using it on its website and in a Google Spreadsheet. An upcoming blog post will cover using it in more complicated setups.

A comment on Cox’s theorem and probabilistic inductivism. (2022/08/31)

I’m currently reading What is this thing called science, an introduction to philosophy of science by Alan Chalmers. Around page 48, Chalmers presents the following chain of reasoning:

One attempt to avoid the problem of induction involves weakening the demand that scientific knowledge be proven true, and resting content with the claim that scientific claims can be shown to be probably true in the light of the evidence. So the vast number of observations that can be invoked to support the claim that materials denser than air fall downwards on earth, although it does not permit us to prove the truth of the claim, does warrant the assertion that the claim is probably true.

In line with this suggestion we can reformulate the principle of induction to read, ‘if a large number of As have been observed under a wide variety of conditions, and if all these observed As have the property B, then all As probably have the property B’. This reformulation does not overcome the problem of induction. The reformulated principle is still a universal statement. It implies, on the basis of a finite number of successes, that all applications of the principle will lead to general conclusions that are probably true.

Introduction to Fermi estimates (2022/08/20)

The following are my notes from an intro to Fermi estimates class I gave at ESPR, in preparation for a Fermithon, i.e., a Fermi estimates tournament.

Fermi estimation is a method for arriving an estimate of an uncertain variable of interest. Given a variable of interest, sometimes you can decompose it into steps, and multiplying those steps together gives you a more accurate estimate than estimating the thing you want to know directly. I’ll go through a proof sketch for this at the end of the post.

If you want to take over the world, why should you care about this? Well, you may care about this if you hope that having better models of the world would lead you to make better decisions, and to better achieve your goals. And Fermi estimates are one way of training or showing off the skill of building models of the world. They have fast feedback loops, because you can in many cases then check the answer on the internet afterwards. But they are probably most useful in cases where you can’t.

What do Americans think ‘cutlery’ means? (2022/08/18)

A while ago, I got into a discussion about what Americans think that the word “cutlery” means. The person I was discussing this with claimed that Americans thought that this referred to cutting instruments, and I claimed that this was preposterous, and that it clearly means all eating instruments.

At some point, I became curious enough that I asked a few American friends, and then spent $50 or so on Positly to survey 21 americans—I overpaid because at the time I wanted to get the data soon. The results are pictured above. The underlying data can be found in here, and the code for replicating the chart is here. I’m not sure if 21 subjects is enough to come to a conclusion, but it was enough to satiate my own curiosity.

A concern about the “evolutionary anchor” of Ajeya Cotra’s report (2022/08/10)

tl;dr: The report underestimates the amount of compute used by evolution because it only looks at what it would take to simulate neurons, rather than neurons in agents inside a complex environment. It’s not clear to me what the magnitude of the error is, but it could range many, many orders of magnitude. This makes it a less forceful outside view.

Background

Within Effective Altruism, Ajeya Cotra’s report on artificial general intelligence (AGI) timelines has been influential in justifying or convincing members and organizations to work on AGI safety. The report has a section on the “evolutionary anchor”, i.e., an upper bound on how much compute it would take to reach artificial general intelligence. The section can be found in pages 24-28 of this Google doc. As a summary, in the report’s own words:

Forecasting Newsletter: July 2022 (2022/08/08)

Highlights

$1,000 Squiggle Experimentation Challenge (2022/08/04)

Motivation

The team at QURI has recently released Squiggle, a very new and experimental programming language for probabilistic estimation. We’re curious about what promising use cases it could enable, and we are launching a prize to incentivize people to find this out.

How much would I have to run to lose 20 kilograms? (2022/07/27)

In short, from my estimates, I would have to run 70-ish to 280-ish 5km runs, which would take me between half a year and a bit over two years. But my gut feeling is telling me that it would take me twice as long, say, between a year and four.

I came up with that estimate because was recently doing some exercise and I didn’t like the machine’s calorie loss calculations, so I rolled some calculations of my own, in Squiggle

Some thoughts on Turing.jl (2022/07/23)

Turing is a cool probabilistic programming new language written on top of Julia. Mostly I just wanted to play around with a different probabilistic programming language, and discard the low-probability hypothesis that things that I am currently doing in Squiggle could be better implemented in it.

My thoughts after downloading it and playing with it a tiny bit are as follows:

1. Installation is annoying: The program is pretty heavy, and it requires several steps (you have to install Julia and then Turing as a package, which is annoying (e.g., I had to figure out where in the filesystem to put the Julia binaries).

More content