My favorite thing about stories like this is that the provide a lot of data for mental modeling: Suppose you're attempting to build up a mental(/scientific) mo

My favorite thing about stories like this is that the provide a lot of data for mental modeling: Suppose you're attempting to build up a mental(/scientific) model of how the human brain works. You're trying to come up with a set of abstractions and rules for how those abstractions interact, with which you can intuit, without performing a full simulation, the answers to experimental questions. That is, your model should produce concrete predictions: if I prod the brain at A, B will happen; invisible effect C is correlated with visible effect D; nebulous phenomena E and F are well-defined instances of type G. A good model is more like a type system than a simulation; it gives you names and concepts that match (with some loss of fidelity) those used by reality. It lets you write valid programs (explanations) that match programs (physics) that are being executed by reality. The report that a person had multiple personalities and some of them were blind, leads to the logical 'program': Either (this person has a brain structure that was so bizarre it isn't related to the rest of the species' brains) [assume this is false] Or (it is possible for humans to contain multiple personalities which are not equally able to interact with sensory organs) Therefore => (it is possible for human brains to contain multiple personalities) (sensory organs are wired into personalities) (personalities can lose their wiring to sensory organs without physical damage to the organs) That is, a single report cuts out a huge swathe of model-space. With a single example we are obliged to include in EVERY model concepts of personalities (or something, er, homomorphic to them), and a concept of a wiring to the sensory organs, and the ability to express that this wiring can become broken. A perfect model must explain every example (or leave a place for as-yet-undiscovered mechanisms to explain them). New examples put new burdens on every existing explanation. They cut model-space down by, presumably, percentage chunks ("excise the subspace of model-space that doesn't model personalities and wiring personalities to sensory organs"). If you imagine that there are something like 2^N possible models at a given granularity, then, the cuts are approximately reducing us to 2^(N-1) each time. Our models with N degrees of freedom that has to explain E experiments has more like N-E degrees of freedom (disregarded 'covariance' between experimental results). What this means is that a random model is overwhelmingly likely to be wrong with each successive experiment. (In practice our models have large dark spots where we leave many degrees of freedom open - but even still, it's likely that any explicit modeling choice we make ends up being incorrect, and quickly.) (This math is pretty made-up. We could do better with something akin to Kolgomorov complexity rather than degrees of freedom, since type systems should be discounted by how complex they are so that systems more complex than reality actually is end up contributing nothing to the total.) Summary: I enjoy stories about weird neuroscience because they defy my mental model of "how brains work" almost every time, even if I keep revising it.