exxos's Atari, Amiga & retro forum

Been trying to build a new autorouter, I get the impression its not going very well :P

:lol:

Reminds me of a video I saw once about the shapes of webs spiders make after being fed with LSD.

(Yep; you read that last sentence correctly!)

mrbombermillzy wrote: 27 Aug 2026 16:39 Reminds me of a video I saw once about the shapes of webs spiders make after being fed with LSD.

Yeah I remember that video, was great !

It didn't build the web it built a hammock :lol:

Still trying to get AI to do something "useful" with a PCB. Only thing it managed so far was to move IC5 to the right a bit with moving vias and tracks in KiCad. It actually did a good job. That saved a lot of clicking. But even after giving it answers to routing problems, it still manages to somehow mess it up :roll: You would think in 2026 there be a router which doesn't suck by now !

Well this is going well then... Not.

Still trying to train A.I. to route boards as good as I can do them.. This one actually I routed myself and uses way less via's than any auto router has achieved yet.. It seems a bit difficult to try and teach A.I. basically lifelong experience on how to do things...


Apparently there are a set number of algorithms used throughout the industry for routings I'm getting those ported over, mixed in with a "bias" on how I want things done.

I have had some limited success overall so far. Basically took two days to do something I could have probably done in a few seconds. But the point being, because of my RSI pain in my hands, my actual time being able to work on PCBs and such is basically non-existent these days. Plus with not sleeping well I cannot concentrate on such things for long anyway. So even though technically I could do it quicker in some respects, in other respects it would actually take me a long time maybe even weeks.

So even if A.I. takes all day to do something relatively simple, all I have to do is check the end results. Basically the same as what any other router does but I want a router I can control how it routes myself.. Which is the whole point of the exercise..

Algorithms suck :lol: but explains why all water routers seem to also suck if they are all doing basically the same things.

So far its best effort has used 70 vias vs my 26, and there were several tracks it could not even complete!

Some trying to come up with a new algorithm for routing PCBs, this may turn into a bit of a problem :lol: :roll:

This is how long Claude has been going for one single session so far today :lol: :roll:


Still trying to get into Ruta PCB as good as me. But it seems to keep coming to the same conclusion that less vias = more unrooted tracks. And the same in reverse, that more vias = less routed tracks. But no matter what combination it goes through, it simply has not got anywhere near how I routed it manually yet.

Even more bizarre thing which it is currently churning through, is that I literally gave it the final board layout, asked it to remove one track, and uses same router to see if he could replace that singletrack, and so far it has basically failed twice and each one of those runs was about four hours long.

I have given it multiple ways and skills on how I do the boards and why I do the things I do, but every single change seems to somehow make things worse not better.

I gave up with the regular algorithms as they was just even worse to what I was building myself anyway . But as to why it still keeps failing miserably, Is really baffling.

I noticed the same sort of behaviour in Eagle as well. Where you want to try routing literally a few millimetres apart and it will somehow manage to find the most convoluted long winded lengthy route imaginable on the board every single time.

I saw a EEVblog video a few days ago. Something about some AI and wanting PCB designers to design with its router to basically train their AI model on how people route PCBs. Not sure that's wise if the general public are going to be using the system :lol:

Maybe I should just give the AI every single board I have ever done to give it examples maybe that would help :shrug: pretty much running out of ideas at this point :(

Wait.. what...


I think I have just been out routed ! :shock:

I see one mistake it made where it could save length on a track which will get track length down a fair bit...

This is starting to read like the Big Blue vs Kasparov chess match from the 90's. :lol: