The author is pushing back against the common view that "uncertainty quantification = measuring how wrong your model might be." That's only half the story. Even a perfect model — one that has learned the true data-generating process exactly — still has uncertainty to report, because the world itself is noisy.
Think of a concrete example. You're predicting house prices from square footage. Even if you knew the exact true relationship between square footage and price, two 1,500 sq ft houses in the same neighborhood will sell for different prices — different sellers, different conditions, different buyers. That variation isn't your model's fault. It's baked into reality.
Mathematically, the true conditional distribution p(y|x) has some width, some spread. For a given input x (say, 1,500 sq ft), there's not one true y but a distribution of possible y values. UQ's first job is just to faithfully represent that spread rather than collapsing it to a single number.
What "marginalized-z spread" refers to: the author mentioned earlier (in a prior part of the conversation) that there are hidden variables z — things like the house's condition, the seller's urgency, market timing — that you don't observe. The true conditional is really:
$$ p(y|x) = \int p(y|x, z) \, p(z|x) \, dz $$
Even if p(y|x,z) were very sharp (if you knew everything, the price is nearly determined), once you marginalize over the unobserved z, the resulting p(y|x) has spread. That spread is aleatoric uncertainty.
How models capture this in practice:
A standard regression model outputs a single prediction ŷ = f(x). That throws away all information about spread. A heteroscedastic regression head instead outputs two things: μ(x) and σ(x) — a predicted mean and a predicted standard deviation that can vary with x. This lets the model say "for this particular input, I expect high variance" (maybe a house in an erratic market) versus "for this input, I'm confident" (maybe a house in a very uniform subdivision). For classification, the softmax vector itself represents this spread: a [0.7, 0.2, 0.1] output is the model's attempt to represent the true class probabilities for that input.
The key property: aleatoric uncertainty does not shrink with more training data. If two identical houses genuinely sell for different prices 30% of the time, no amount of data changes that fact. It's a property of the world, not of your ignorance.
This is the part most people mean by "uncertainty quantification" — how far is your learned model p̂(y|x) from the truth p(y|x)?
The author makes a crucial point: you can never measure this gap directly. If you could, you'd just correct your model by that amount and have a better model. This sounds almost tautological, but it's important. The true p(y|x) is the very thing you're trying to learn. You can't compare against something you don't have.
So how do you operationalize it? Through disagreement among plausible models.
Instead of committing to one set of parameters θ (one model), you maintain a distribution over parameters — the posterior p(θ|D), which tells you "given my data D, which parameter settings are plausible?"