Notes (August 2021)

I wrote this a few years ago in preparation for a very open-ended brief for a presentation. A better title than the original might be:

Some Basic Concepts from Academia That Should Be Much Better Known

Diminishing Marginal Returns

Returns on time, money, or effort likely increase at the margin, then decrease, and eventually become negative. Figuring out where you are on this curve and when to stop is incredibly useful.

Evolutionary Stable State (ESS)

Societies, economies, or ecosystems may evolve to a state where there is no incentive to defect, but this state can be detrimental to everyone (a suboptimal ESS).

Don't compete, cooperate

According to Homer Simpson, communism is nice in theory. So is capitalism, except not for the competing agent whose marginal return is expected to go to zero. Don't be a commodity or compete directly; this often means cooperating. Cooperation may even be better for everyone (see ESS).

Complexity as Risk

Systems can quickly become fragile and impossible to understand or maintain. If isolated, complexity is probably benign.

Chaitin-Kolmogorov complexity

The minimal length of a process required to generate something is precisely how random or complex that thing is. This is a beautiful definition, and it's suggestive in practice: am I configuring things such that descriptions and processes are efficient?

Pareto Improvement

A change that makes everyone (weakly) better off. Changes don't always have to make someone worse off; we are not typically just cutting up a fixed-size pizza. Conversely, the demand that policy changes must always be Pareto improvements is absurd. If making a very wealthy individual 90% poorer makes a million other people significantly better off, then that is almost certainly a good idea.

Statistical Significance

How likely is an observed change due to chance? If the probability is more than 5%, assume no real change has occurred.

Multiple Hypothesis testing

However, the above cannot be applied naively. Perform many statistical tests, and you will eventually get something passing this threshold (1 in 20 times, assuming your assumptions are valid). One should decide what to test ahead of time and apply corrections for multiple comparisons as much as possible.

Automate

Humans are fallible, distractible, and limited. Automate tasks wherever possible.

Noise and Feedback

Get feedback, but be aware that much of it will be noise or idiosyncratic. Ideally, test quantitatively on huge numbers of people with good scientific controls. However, this is basically never possible. Some simple heuristics seem sensible, e.g., Basecamp's approach: "we don't record feedback; the important stuff will recur."