What's the point of asymptotics? I presume this is a rather stupid question, but I hope some of you can find a bit of time to entertain it.
Looking at asymptotic behavior of estimators/test statistics etc means looking at their behavior as sample size approaches infinity. But if our sample size is infinite, doesn't that mean that we already have the whole population?
Case in point: consistency of OLS estimates. Under certain assumptions we are guaranteed convergence of estimates to true parameters in the limit, but at infinite sample size we already have the whole population, so why would we even want to estimate anything when every possible data point is already contained in the sample, all we have to do is to perform a lookup.
 A: It's true, in general, that consistency is not the be-all-end-all of a statistic. But neither is unbiasedness for the same reason! When you accept biased estimators, you open a whole class of Bayes estimators that beat OLS in many respects. Case-in-point: the problem of shrinkage and L2 penalization in the estimation of a mean vector for a multivariate normal sample.
For the purposes of hypothesis testing and error estimation, asymptotics go farther than just the limit of the statistic. We're also interested in whether a suitable transformation of the statistic has a known limiting distribution. In that case, we can use the limiting distribution to approximate the sampling distribution of the test statistic in finite sample sizes. The classic example is the central limit theorem where $\sqrt{n} (\bar{X} - \mu) \rightarrow_d \mathcal{N}(0, 1)$. And similar methods are used to identify critical values, p-values, power, and sample size calculations knowing full well they are only approximations. In some cases, the approximations can be improved (such as with the Student-T distribution, Agresti correction, Clopper-Pearson intervals) or exchanged entirely for exact (Fisher's Exact Test) or empirical (Bootstrap) methods. It's my opinion that these methods should be taught alongside standard testing and error estimation methods, and the latter methods used more often in practical data analysis.
The idea of an "infinite" sample size underpins all of frequentist statistics. Consider the a "frequentist" interpretation of probability: what do we mean when we say the heads probability of a coin flip is 0.5? If you actually flipped a coin an infinite number of times, it would wear down to nothing. The same holds with frequentist methods for finite sampling, that is when you sample a substantial fraction of a finite population. The sampling distributions might change somewhat, but I still conceptualize
multiple (i.e. an infinite number of) scenarios when I can replicate a particular result. The "expectation" of my estimator - and it's ultimate limit as a result of the LLN - is defined by that value. Suppose for instance I sample 30% of all surviving Siberian tigers for biometrics - say length. I can produce a CI for my length estimate. That CI is based on sampling 30% of all known Siberian tigers (as of now <400) an infinite number of times.
That said, the implementation of these methods can introduce some methodological issues (bootstrap intervals are not always valid, even with bca or double bootstrap), and some problems require analytic simplicity to provide reproducible, and communicable results. For instance, when calculating the sample size of a time to event analysis, I can base my selection of N on an exponential distribution with a known rate parameter in the control and treatment arms, duration of follow-up, and a test based on the asymptotic Wald statistic. In that case, it's easy for another statistician to verify my results.
In summary, for a didactic program, I would say learn both methods and understand their limitations. For a practical application, consider your audience and what they need to understand. And when in doubt, be conservative in your approach!
A: Most asymptotic results are closely connected to finite first-order results. We'll review this in the non-probabilistic case and then extend to the probabilistic case.
Non random case: reduced order analysis
Recall the Taylor series of the function $\sin$ around the point $x=0$:
\begin{equation*}
 \sin x - 0 = x - \frac{x^3}{3!} + \frac{x^5}{5!} - \frac{x^7}{7!} + \dots,
\end{equation*}
which most learn about in introductory calculus courses. The series helps us understand the behavior of the function $\sin$ function near $x=0$. The plot below (from Wikipedia) shows that these four terms largely reproduce the behavior of the curve near $x=0$:

Thus, to study the behavior of $\sin(x)$ near $x=0$, very little is lost in simply studying the behavior of the polynomial $x \mapsto x - \frac{x^3}{3!} + \frac{x^5}{5!} - \frac{x^7}{7!}$. Of course, the polynomial is much more susceptible to study, so this is convenient.
Non random case: first order analysis
The close relationship between the function $\sin$ and its truncated Taylor series also holds in the first order, when we approximate
\begin{equation*}
 \sin x - 0 = x - \dots,
\end{equation*}
For example, most calculus students have shown that
\begin{equation*}
 \lim_{x \to 0} \frac{\sin x}{x} = 1.
\end{equation*}
This result can be interpreted as saying that the function $\sin$ and the identity function $x \mapsto x$ are indistinguishable near $x=0$.
We stress that the limit result, although asymptotic, is merely reflecting our knowledge gained from the figure above: the two functions are very close near $x=0$.
Probabilistic case
There is an extension to Taylor series which is designed for estimators rather than functions. We'll summarise the extension and illustrate how it is behind most "asymptotic" results.
Let $P_0$ denote the true (unknown) distribution, let $\theta(P_0)$ denote the parameter of interest, and let $T(X_1, \dots, X_n)$ be the estimator using i.i.d. data $X_i \sim P_0$. Then, the following expansion is a first order expansion of the estimator around the parameter:
\begin{align*}
 T(X_1, \dots, X_n) - \theta(P) = \frac{1}{n} \sum_{i=1}^n \varphi(X_i; P_0) + \dots.
\end{align*}
The term on the left is the estimator error, i.e. the difference between the estimator and its target (the parameter). The term on the right is a first order expansion, analogous to $x$ in the $\sin$ case. The function $\varphi$ is called the "influence function" and determines the asymptotic behavior, along with the remainder.
(More details later..)
A: The first reason we look at the asymptotics of estimators is that we want to check that our estimator is sensible.  One aspect of this investigation is that we expect a sensible estimator will generally get better as we get more data, and it eventually becomes "perfect" as the amount of data gets to the full population.  You are correct that when $n \rightarrow \infty$ we have the whole (super)population, so presumably we should then be able to know any identifiable parameters of interest.  If that is the case, it suggests that non-consistent estimators should be ruled out of consideration as failing a basic sensibleness criterion.  As you say, if we have the whole population then we should be able to determine parameters of interest perfectly, so if an estimator doesn't do this, it suggests that it is fundamentally flawed.  There are a number of other asymptotic properties that are similarly of interest, but less important than consistency.
Another reason we look at the asyptotics of estimators is that if we have a large sample size, we can often use the asymptotic properties as approximations to the finite-sample behaviour of the estimator.  For example, if an estimator is known to be asymptotically normally distributed (which is true in a wide class of cases) then we will often perform statistical analysis that uses the normal distribution as an approximation to the true distribution of the estimator, so long as the sample size is large.  Many statistical hypothesis tests (e.g., chi-squared tests) are built on this basis, and so are a lot of confidence intervals.
