I still don't quite know how Gaussian curvature is defined. I feel like you take a tiny triangle and divide its its angle excess by its area—in accordance with how spherical triangles work—but that's probably not the usual way
(And it kinda makes Gauss–Bonnay seem a bit too easy)
@Akiva Now define the "shape operator" $S_p : T_p M \to T_p M$ to be $S_p(v) = df_p(v)$, the directional derivative of the Gauss map at $p$ in the direction of $v$. So you're moving the unit normal at $p \in M$ slightly in the direction of $v$ and thinking about how much that changes. It's a formal fact that $S_p$ does indeed land in $T_p M$ and not elsewhere in $\Bbb R^3$.
In this case, the constructive proof is simple enough, but for other stuff, I sometimes prefer the nonconstructive proofs because explicitely writing everything down gets messy. For example, prove that $a_0+a_1T+\dots a_nT^n \in A[T]^\times$ iff $a_0 \in A^\times$ and $a_i$ nilpotent for $i>0$. The nonconstructive proof is simple and elegant, but the constructive proof is non-intuitive and messy imo
To compute a directional derivative, you follow a curve on the surface. Following the unit normal along that curve gives, in turn, a curve on the sphere.
If you have a particle moving on a sphere, if its velocity had any normal component, the particle would move off the sphere (its distance from the center would increase/decrease).
First we prove that $A[x]^\times=A^\times$ if $A$ is an integral domain. But that's obvious, because $\operatorname{deg}(fg)=\operatorname{deg}(f)+\operatorname{deg}(g)$ as there can't be any cancelation, so all units have degree 0.
Now if $a_0+a_1T+\dots+a_nT^n$ is a unit in $A[T]$, then for every prime ideal $\mathfrak{p} \in \operatorname{Spec}A$, we have that $\overline{a_0+a_1T+\dots+a_nT^n}$ is a unit in $A/\mathfrak{p}[T]$, but that's an integral domain, so $\overline{a_i}= 0$ in $A/\mathfrak{p}[T]$, thus $a_i \in \mathfrak{p}$ for $i>0$. As $\mathfrak{p}$ was arbitrary, $a_i$ is in every prime ideal of $A$ for $i>0$, thus $a_i$ is nilpotent for $i>0$
@AkivaWeinberger So the shape map $S_p : T_p M \to T_p M$ is given by taking the rate of change of the normal at $S$ if you move it a little bit in the direction of $v \in T_p M$, just to summarize.
Say you change of basis $T_p M$ to have the $x$ and $y$-axes as the eigenspaces of $S_p$ (so you appropriately parameterize $M$ near $p$ on the first place)
So, just as Jacobian determinants give the fudge factor for area/volume for mappings of Euclidean space, the determinant of the Gauss map gives you the factor by which the Gauss map distorts area. That was Gauss's original interpretation. This is the Radon-Nikodym derivative to which I earlier referred facetiously.
@Balarka: As I warned in my notes, the matrix of the shape operator need not be a symmetric matrix.
@MatheinBoulomenos No need to be. I probably got unnecessarily annoyed, but actually was a little sad because you and Narcissus misinterpret my jokes as algebra-hate.
But yeah, I was about to tell Akiva that's it's basically, if you draw a small coordinate xy-square on the surface with one vertex at p, the amount of stretching the Gauss map does to the square is K