I can't help but think the "surprise examination paradox" rests too much in English equivocation for it to be a properly logical paradox. In particular, the fact that "surprise" changes over time, and the fact that if I've logically deduced that it is "impossible" for the test to occur on the last day then it is ipso facto a surprise if it happens then.

Sit down and make the argument really rigorous as to the definition of "surprise" and the fuzz disappears. You can get several different results from doing so, and that's really another way of saying the original problem is inadequately specified and not really a logical conundrum. As "logical conundrums" go, equivocation is endlessly fascinating to humans, it seems, but any conundrum that can be solved merely by being more careful, up to merely a normal level of mathematical rigor, isn't logically interesting.

    1.000000000

I would encourage anyone who's intrigued by this paradox to read Timothy Chow's comprehensive paper on the subject (https://arxiv.org/abs/math/9903160).

In particular, he discusses what he calls the meta-paradox:

> The meta-paradox consists of two seemingly incompatible facts. The first is that the surprise exam paradox seems easy to resolve. Those seeing it for the first time typically have the instinctive reaction that the flaw in the students’ reasoning is obvious. Furthermore, most readers who have tried to think it through have had little difficulty resolving it to their own satisfaction.

> The second (astonishing) fact is that to date nearly a hundred papers on the paradox have been published, and still no consensus on its correct resolution has been reached. The paradox has even been called a “significant problem” for philosophy [30, chapter 7, section VII]. How can this be? Can such a ridiculous argument really be a major unsolved mystery? If not, why does paper after paper begin by brusquely dismissing all previous work and claiming that it alone presents the long-awaited simple solution that lays the paradox to rest once and for all?

> Some other paradoxes suffer from a similar meta-paradox, but the problem is especially acute in the case of the surprise examination paradox. For most other trivial-sounding paradoxes there is broad consensus on the proper resolution, whereas for the surprise exam paradox there is not even agreement on its proper formulation. Since one’s view of the meta-paradox influences the way one views the paradox itself, I must try to clear up the former before discussing the latter.

> In my view, most of the confusion has been caused by authors who have plunged into the process of “resolving” the paradox without first having a clear idea of what it means to “resolve” a paradox. The goal is poorly understood, so controversy over whether the goal has been attained is inevitable. Let me now suggest a way of thinking about the process of “resolving a paradox” that I believe dispels the meta-paradox.

But it's stipulated that the test will happen on one of the days - it's not a possibility that it won't happen at all, hence the paradox.

One resolution is that what the teacher stipulates is impossible. It should really be

"You'll have a test within the next x days but won't know which day it'll be on (unless it's the last day)"

> The students convince themselves that it can't happen at all... and that's well and good, but once they admit that as an option, they have to include that in their argument—and if they do so, their entire argument falls apart immediate.

Your critical thinking is bad. The first paradox happens when the prisoner concludes that the judge lied, using a rational deduction. A second paradox happens when it transpires the judge told the truth.

> So you need to restart your argument, with your new possibility as one of the premises. And now you don't get to a contradiction at all.

It’s amusing that you stopped here without giving an actual solution. Please do tell us, which day is the test on?

I would have thought that the proof shows this problem is unavoidable. How in your view do we avoid this problem?

> F ⊢ (Prov_F(“s”) → s) → s

This is called "Löb’s hypothesis" and it's an incredible piece of logical machinery[1]. If you truly understand it, it's pretty mind-blowing that it's actually a logically sound statement. It's one of my favorite ways to prove Gödel's Second Theorem of Incompleteness.

[1] https://categorified.net/FreshmanSeminar2014/Lobs-Theorem.pd...

While I am to old to say what is taught today, this is exactly the map vs territory issue I mentioned in my other comment.

It is all there in what you would have been taught, but hidden because we tend to avoid the hay in the haystack problems and focus on the needles, because we don't have access to the hay.

As an example that can cause huge arguments, if for analysis you used Rudin, go look for an equally binary operator in that book, as every construction of the reals is a measure zero set, it is actually impossible to prove the equality of two real numbers. ZFC uses constructablity, Spivik uses Cauchy sequences etc...

If you look at the paper[1], about the equivalence of PAC/VC dimensionality it is all there, framing learning as a decision problem, set shattering etc.

Richardson's theorm, especially with more recent papers is another lens.

With the idea that all models are wrong, but some are useful, it does seem that curriculums tend to leave these concepts to postgraduate classes, often hiding them or ignoring them, in descriptive complexity theory they should have been front and center IMHO.

Assuming something is computable or learnable is important for finding pathological cases that are useful, don't confuse the map with the territory.

We have enough examples to know the neuronic model is wrong, but the proposed models we have found aren't useful yet, and with what this post provided shows that may always hold.

Physics and other fields make similar assumptions, like physics assumptions that laplacian determinism is true, despite counterexamples.

Gödel, Rice, Turing. etc... may be proven wrong some day, but right now Halt ~= open frame ~= symbol grounding ~= system identification in the general case is the safe bet.

But that doesn't help get work done, or possibly find new math that changes that.

[1] https://dl.acm.org/doi/10.1145/76359.76371

> We usually assume that (a) the entire universe is computable and (b) even stronger than that, the entire universe is _learnable_, so we can just approximate everything using almost any function as long as we use neural networks and backpropagation, and have enough data.

I don't think the assumption is that strong. The assumption is rather that human learning is computable and therefore a machine equivalent of it should be too.

Isn't this a wittgensteinian problem? ie how you interpret language is inherently subjective. But regardless of what language you use, or how you intend to bind the terms to reality semantically, apriori truth is still apriori truth. It's the only basis of objective "truth" we have access to. Throwing that out the window feels like tossing the baby out with the bath water.

Imo, this just comes down to the fact that most people would consider "standard" to be a floating signifier. I don't think the idea that mathematical concepts changing definitions when you change axioms is at all controversial in itself.