I’m writing this blog post after listening to a recent episode of the Decoding the Gurus podcast, wherein after an hour of very substantial discussions on the status of academic Psychology, guests Daniël Lakens and Smriti Mehta suddenly entered into an impassioned rant against the annoying antics of some “Bayesians” in their field. Listening to them describe these Bayesians parading their superior methods, and at times even pulling numbers out of their posteriors (pardon the pun) to supposedly make sloppy conclusions, made me feel like Slavoj Žižek during his infamous debate with pop psychologist and lobster fanboy Jordan B. Peterson, in which the latter’s endless barrage of hasty generalizations about “Marxists” led Žižek to ask in exasperation, “Where are the Marxists?”
“Where are the Bayesians?” I exclaim internally, while struggling to disassemble my bicycle (I might blog about that at some other time). Unfortunately for me, I think I know the answer. Unlike Peterson, who is very likely imagining the antics of his arch-nemesis, the Postmodern Neo-Marxists, as passed down by Stephen Hick’s very problematic analysis of both Postmodernism and Marxism, Bayesianism is getting something of a bad rep thanks to its association with Silicon Valley debate bros and Rationalists that now litter the Internet.
You know the ones: they brandish their Ayn Rand libertarianism on their sleeves, claiming no partisanship with either left or right wing politics, but spouting the latter’s talking points at every opportunity. They post hour-long debates on things like the Alignment Problem and whether Tech Accelerationism can bring about prosperity to all mankind (spoiler: yes, if you consider mankind = the rich people only), while completely ignoring the structure and formalism of debates, and using a thousand variations of the trolley problem to support their points.
No one comes to mind more frequently than the notorious AI catastrophizer, Eliezer Yudkowsky, who leads the blog LessWrong. I submit to the court’s consideration this blog post from 2007 titled “Bayesian Judo”, in which Yudkowsky absolutely rips a poor fellow to shreds at a dinner party for claiming that “only God can make a soul.” Yudkowsky then corners him with arguments, at one point even invoking Aumann’s Agreement Theorem, which commenters aptly recognized may not have been familiar to Yudkowsky’s opponent. Aumann’s Theorem is certainly Bayesian: it argues that two rational observers starting from the same prior are given the same dataset, then they must arrive at an agreement through identical posterior distributions. Unfortunately, the skirmish at the dinner party itself is nothing Bayesian: I doubt the poor fellow even knew what priors meant, or what a Theorem is, for that matter. But Yudkowsky presents the scene as if some formidable demonstration of rationality.
The blog post is very old, but I think it is demonstrative of the kind of brainrot that has come to associate itself by force with Bayesianism. Bayes’ Rule is a very useful way of combining evidence with prior belief, and it’s proven to be very successful within the realm of statistics and machine learning. Outside of hard numbers and mathematically constructed probability distributions, however, it just plain doesn’t work. Let’s say you wanted to use Bayes’ Rule to support a very controversial opinion regarding the origins of the COVID pandemic. All you need to do is – as Lakens and Mehta observe – pull the right numbers out of your ass and Bayes rule does the rest.
Let’s say I have a prior belief that COVID came from a lab leak: 
There you go: the data now suggests that a lab leak is much more likely than it not having been a lab leak. And to think that I even started with equally likely priors, just to make sure my analysis was fair! Hell, why even bother being fair: let’s presume instead the lab leak is actually less likely, say 
It’s now even more likely – we should definitely look into this!
Where did we go wrong? First off, Bayes’ Rule is certainly an efficient way of updating prior uncertainty about hypotheses, but it has to operate strictly within the realm of statistics. Probabilities cannot just come out of nowhere: they have to follow strict mathematical construction. And even when both of these are satisfied, garbage data and reasoning will still result in garbage analysis. This is why practicing statisticians and data analysts will perform sanity checks, diagnostic analysis, and prior and likelihood sensitivity studies to make sure all gaps are stoppered. An entire section of Andrew Gelman’s textbook, Bayesian Data Analysis, literally covers only that.
The application of Bayes’ rule in the statistical sciences involves some very complex and rigorous mathematics, at times being so intractable Bayesians had to sit and wait until powerful computers became accessible so they could actually tackle the more daunting real world problems. None of the debate bros that I’ve encountered in the wild (of the Internet, obviously, being oceans away from San Francisco, thankfully) do any of that. Instead, they tack on the label of Bayesianism to wrap otherwise rickety logic with the veneer of faux mathematical sophistication and cite unrelated theorems with scary sounding names at people over dinner party one-upmanships because, well, I guess it does sound kinda cool.
Dominic Dayta 
Leave a comment