Phatic and anti-inductive environments
Scott Alexander has an excellent article on phatic vs. anti-inductive environments. He contrasts content-free communication (phatic, “talk-for-talk’s sake”) used for social signalling  with anti-inductive communication – coming up with unusual insights, that have not yet demoted to clichés by overuse .
Epistemology and anti-inductivity
The archetypal example of an anti-inductive system is the stock market (especially under the strong version of efficient market hypothesis).
Let’s say two stock prices are historically anticorrelated – the variance in their returns moves in opposite directions. As soon as everyone believes this, hedge-fund managers will leverage up and buy both stocks. Everyone will do this, meaning that both stocks will rise. As the stocks rise, their returns get more expensive. The hedge-fund managers book profits, though, because their stocks are rising. Eventually the stock prices rise to the point they can go down. Once they do, hedge-fund managers who got in late will have to liquidate some of their assets to cover margin calls. This means that both stock prices will go down – at the same time, even though they were originally anticorrelated. Other hedge funds may lose money on the same two stocks and also sell or liquidate, driving the price down further, etcetera. The correlative structure behaves anti-inductively, because other people can observe it too.
– Eliezer Yudkovsky, Markets are Anti-Inductive
Anti-inductive environments are fascinating, because it is here that the bread-and-butter of “standard epistemology” – inductive reasoning – breaks down. I put quotes around “standard epistemology” because obviously this is exactly what hard-line Popperianism warns against (currently its most vocal proponent being Nicolas Nassim Taleb of the Black Swan fame).
Nonetheless, we do use induction all the time, trading off robustness of predictions for efficiency. It is indeed a great tool – assuming you are in an inductive environment (so definitely not in stock markets). Induction also works better, if you are close to the right solution (a local optimum in the solution space) – say ordinary, regular day-to-day problems. It does not work well with not well mapped solution spaces – say the cutting edge of science. That’s why we should rely there on falsification rather than induction.
The second key point is that anti-induction drains out any information value, the more it is acted on or decreases the utility of a given resource the more it is utilized.
Besides stock market, Scott mentions job interviewing as anti-inductive. Can we come up with other examples?
restaurant and VACATION Recommendations
These are typical anti-inductive situations: say you find a perfect vacation spot. The more people know about it, the worse it gets due to overcrowding.
If the GPS recommends the same alternative route to avoid a traffic jam for everybody, it potentially creates a new traffic jam.
The hipster Dilemma
Hipsters have it though. Saying: “I liked X, before it was cool” (for X being an indie band, foreign movie, underground writer etc.) is needed to signal your superior aesthetic taste and intellect, but at the same time devalues the resource X, by making it less obscure.
“I don’t want to belong to any club that would accept me as one of its members.”
— Groucho Marx
You want to join social circles so that they increase your status. That means that your status has to be lower than the average of the social circle. If they accept you, the average status of the group decreases. And vice versa, a group wants new members with status higher than its average status, but why would he/her want to join?
Shared resources and tragedy of Commons
Many shared resources have anti-inductive features. The experience of a great movie can be ruined by an overcrowded cinema etc. This all is closely related to the tragedy of commons.
Solving anti-induction with Fermi-Dirac statistics
Interestingly, “crowd-aversion” behavior has a parallel in physics. Fermions are elementary particles with half-integer spins that follow the Fermi-Dirac statistics. Examples are quarks, leptons (e.g. electrons) and any composite particle from odd-number of fermions (e.g. protons and neutrons) .
Fermions obey Pauli’s exclusion principle: two identical fermions cannot occupy the same quantum state simultaneously. That’s why electrons don’t just bunch up on the lowest orbit in an atom and give rise to all kinds of interesting chemistry, such as you and me.
Recommendation engines can meet the required degrees of crowd-avoidance by incorporating a fermionic replusion term in their objective function. A vacation recommendation engine therefore will not recommend both you and me the top place (giving us both a miserable time), but splits us between the few top ranked locations maximizing our overall satisfaction.
So did we solve anti-induction? Not so fast grasshopper! Maybe there is a second order anti-inductivity? As the old zen master asks: if an algorithm finds a trend, is it still cool?
I will leave as an exercise to the reader to work out the principles of a fermionic-stock market (herd behaviour? No problem!) and a fermionic utilitarianism (repugnant conclusion? Not for us!).
 Key here is that this is not a judgment statement. Content-free signalling is in fact important in many circumstances, albeit makes nerd-y types like us uncomfortable.
 Being true to Robin Hanson, anti-inductive communication is not about insights, but it is too about signalling just as phatic communicaiton, except it signals high cognitive capability and creativity (and hence a good genetic fond) instead of social alignment.
 For completeness, the counterparts to fermions, the integer spin bosons, do not mind crowds and they too can serve as a metaphor (and maybe more) in socio-economic systems, such as modelling monopolies/winner-takes-all dynamics in social networks via a Bose-Einstein condensation.