Well, we're not quite "just an accident". We're a local maximum in a roughly 2^1.5e9-dimensional search space. Natural selection has been searching this space by means of trial and error for 3.5 billion years.

This search strategy has already stumbled across two ways of searching better. The first, DNA exchange, occurs either through sexual reproduction or plasmids. By exchanging DNA, organisms can mix-and-match partial solutions.

The second search improvement was a bit more subtle. Once organisms could learn, they could search "nearby" designs and remember what worked. And once evolution started selecting organisms for the ability to learn specific behaviors, you got a low-grade pseudo-Lamarckianism called the Baldwin effect. If I understand the math correctly, this should speed up evolution.

Natural selection finds "good" solutions (i.e., ones which have lots of offspring) far more efficiently than random chance, because it has memory from one generation to the next (DNA), recombination (sex), and fast local search (learning and the Baldwin effect).

Of course, even with written language and universities, we still can't learn all that fast. Humans can only learn at 1 or 2 bits per second times 6 billion people, assuming they're all awake, studying something useful, and so on. So there's still plenty of room for improvement, which is either impressive or scary.