The idea is the following: assuming we know $x(t)$, we don't know what $x(t+h)-x(t)$ is exactly. But intuitively, as $h$ is relatively small, it's probably somewhere in the vicinity of $hx'(t)$. So that's our first estimate, and we call it $K_1$.
However, an even better approximation would be $\frac{hx'(t) + hx'(t+h)}2$. Now, $hx'(t)$ we already have. That's just $K_1$. However, $x'(t+h)$ is a different story. As we don't know $x(t+h)$, we can't just calculate $f(t+h, x(t+h))$ directly. But we know a value that is relatively close to $x(t+h)$, and that's $x(t)+ K_1$. So we use that as our best estimate to get
$$
x'(t+h)\approx f(t+h, x(t)+K_1)
$$
And that's why we use that specific value for $k$.
As for the $\alpha$ and $\beta$ (I don't think you ever want $\alpha\neq\beta$, but I could be wrong; computational methods isn't my main field), that's a result of asking for $\frac{hx'(t) + hx'(t+\alpha h)}2$ in step 2, rather than $\frac{hx'(t) + hx'(t+h)}2$. The exact same argument applies.
Higher order RK methods repeat this process, each time with cleverly chosen $k, \alpha$ and $\beta$, based on the best estimate we can make so far.