Another big question which likely would be better suited as a conversation… Glad to have you in the community asking such questions publicly, as there is way too much misinformation out there.
Your toy problem is problematic for a few reasons. First there is no clear relationship between your inputs and your output. Second there is a clearly identifiable solution.
Since you’re working on a toy problem, there isn’t much concern with the ‘obvious’ solution issue, but as you review the technology you need to keep the intended use in mind. If we apply the same logic to a calculator we wouldn’t have an addition button, as we can calculate 2+2 in our head faster than we can with a calculator. The complexity of the problem is what brings the value to the toolset. Similarly generative design works in the context of complex problems with no clear outcome. But as you’re looking to verify the tool works we can skip this for now.
And so let’s concentrate on the other issue: the lack of clear relationship between the inputs and the outputs. What I mean by that is that you want a net vector of 0. So the way Generative Design works is by running a random sampling of inputs, gets results from the outputs, and takes the top performing results and uses them to run the next study. Let’s say the first population of inputs were:
(16,45,52)
(30,55,36)
(42,35,32)
(55,42,12)
The ‘highest weighted’ X input would be 42 even though 16 was the ‘best’ of the X values. Y’s strongest highest weighted would be 35 which is unfortunately their best scoring though it didn’t have any really good values overall. And Z’s highest weighted is 32, despite having a better value of 12 on the board.
So the lesson learned is that if you know the direct impact of the input on the output, provide a metric for it. In this case it’d be ‘provide a minimal output’ for each of the X, Y and Z values. Admittedly not very sexy, but it would speed up the exploration quickly.
That said, it’s likely better still to provide a more reasonable toy problem. I like ‘fitting two cabins on a wooded lot’ as it provides ample input/output examples and has more value than ‘point in space’ based examples. You’ll have to fast forward though some initial ‘what’s new’ content, but I did a demo of this for the Chinamo users group a few years back which you can watch here: https://m.youtube.com/watch?v=42SylTdN3CY
Yes - there aren’t many other than ‘it must run well in Dynamo with less than 750mb of RAM for the session (this number varies but it is a good rule of thumb).
Not really much gain there. Dynamo runs in the CPU, using one core, so fast single core speed will matter the most. Generative Design is limited to 6 concurrent Dynamo instances at a time (no more, no less) so a massive amount of cores won’t help at all.
Not at this time, but it is something that has previously existed (look up the old Project Fractal), and may come back someday if/when Dynamo can run in the cloud again.
Nope.
More or less what I said before adding all the quotes… you’ll want a more diverse problem and solution. Reducing the number of inputs also helps as well (ie: utilize permutations of your lists instead of multiple get item at index nodes to reorder stuff).
As you get into ‘real’ stuff be sure to post here as the community can help quite a bit. 
