ok, now this is weird. an evolved algorithm decides to use the *mean* of a network's weights as its learning rate.
found a bug in this random connectivity algo where it tries to connect the same node twice, but it otherwise seems to work well. the bug is increasingly rare as node count increases and i have a couple hundred on each side in practice so it's not too big of a deal at this moment.
each file is individually compressed with Yaz0. i guess they wanted to keep all the 32-bit icons in one archive file (instead of splayed about in the filesystem) without having to load and uncompress all 0.6 MiB in memory
this is so goofy (the 4 combinators above the substation just implement an exponential moving average)
there we go
oh yeah! i made a new plastic factory design for like the 10th time and i really like how it turned out this time. iirc there's only 36 pipes to ground here
okay, that's pretty good
i modified the script to dump the computation split into xmm registers, but the result is that half the instructions are just dedicated to loading the B vectors. didn't benchmark it though; realistically you wouldn't use the full function and just have subtypes for points/lines/planes etc where most of the terms are zero
here's a slightly broken rotating cube, or at least the vertices of one
i've been busy
something like this
Cybrespace is an instance of Mastodon, a social network based on open web protocols and free, open-source software. It is decentralized like e-mail.