The problem with randomness is that it is overwhelmingly likely to produce worse performance, not better. So much so, that it is practically impossible to improve through random effects.
Random mutations don't have to produce better performance to end up with better results. They just need:
1. To move you to a "bigger hill". Imagine going from no eyes to skin spots that are slightly more delicate than regular skin, but are precursors of light-sensing skin cells^1. Over the course of generations, eyes can evolve.
2. To not be too deleterious. One spot of more sensitive skin won't cause the mutation to be immediately evolved away from; it's just not that bad for the individual.
When you say it's "practically impossible to improve through random effects", what is "practically impossible"? How have you weighed that against the number of individuals in a population, the number of random mutations each offspring has, and the length of time we're talking about? The question isn't "what are the chances a blind fish today has a baby with eyes tomorrow?" This question both understates the number of trials evolution has to generate something, and pins the target of evolution. Instead, we should ask "over millions of years, what are the chances that a given population evolves something useful?"
[1] I'm not claiming this is exactly how eyes evolved.
There are two different concepts at play here. The first is the notion of optimization by randomness and the second is evolution by natural selection.
For the first issue, there are is a simple way of seeing what I'm talking about; If you are trying to optimize an algorithm for given results, there are three broad techniques which I will categorize for simplicity as 'stupid', 'random', and 'smart.'
For instance, if I wanted to sort a deck of cards, I could just shuffle it until I had a sorted deck. That's 'random.' This is not the worst you can do, and only if you do worse can random be an improvement. To do worse, you have to actively prevent the deck from becoming sorted. (For example, by using an algorithm that guarantees a 1-3-2 order at the start.) Randomness is not optimization, and stupidity is optimization away from the goal. Smart is convergence on a goal.
However, evolution is not a goal directed process. Those who exist are simply those who lived long enough in the environments they happened find themselves. Randomness doesn't improve any organism or any feature of any organism: it only improves diversity, which allows for more niches to be filled. I mean, maybe this increases the availability of food or the amount of space per organism or something, but it doesn't actually improve organisms. It improves the environment by making it less hostile to life in the abstract.
So it has nothing to do with hill climbing or optimization per se.
I think we're talking past each other. I'm not really sure what claim I made you're responding to. I was responding to this claim you made, with the context of "evolution":
> ...it is practically impossible to improve through random effects.
It's true that evolution is not aimed at a specific target. I made that point in my comment.
>Randomness doesn't improve any organism or any feature of any organism: it only improves diversity...
Here is a list of examples of beneficial mutations: http://www.gate.net/~rwms/EvoMutations.html It includes things like "adaption to high and low temperatures" and "12% (3 out of 26) random mutations in a strain of bacteria improved fitness in a particular environment."
The claim is that random mutations can occur -- at some rate, low though it may be -- and are more likely to be kept through natural selection.
It's really not all that impossible for singular mutations, hence evolution. Further, adaptations composed of independently maladaptive mutations can occur. This is the point of the comment you responded to. They are unlikely, but he alluded to that.
