Author here. Lots of great points being made. I want to throw in a crazy prediction.
I think routing is basically an image transformer problem without a good dataset right now. If the eye of the giant AI companies turned to autorouting and large synthetic but physically simulated circuit datasets were created, we would basically be done and autorouting would be a solved problem.
This means that all the work I’m doing now on heuristic algorithms, as well as all the hard work done by humans, will probably not be needed in the future. I just don’t see autorouting as being more difficult (in terms of solution space size) than the art being produced by transformer models right now.
I’m saying this because you’re right, these heuristic algorithms can only get us so far- the problem is really difficult. But our human intuition, the magic black box operation we do, doesn’t seem to be too far beyond the emerging transformer image models.
The major difference is in PCB, every single track has to abide by the rules, no exceptions are allowed if you want your board to work.
While AI-generated art is full of small defects which people are just ignoring, who cares about non-natural shadows or unrealistically large fingers.
It is possible to iteratively combine AI with DRC checker and loop until all is good, but it's not obvious to me that this would be performant enough, or if this system will stay in some sort of local minimum forever once the circuit is complex enough.
The same way claude never outputs code that has a syntax error, the image transformers will output DRC compliant “images”!
I think spatial partitioning can help solve issues with minor DRC violations as well- it should be easier to correct an image than to generate one from scratch. But I’m not even sure it’ll be necessary because of how coherent image models already are.
Claude doesn't usually produce code that actually works though. Passing DRC is one thing (no syntax errors). Actually works is another (compiles and executes with the desired effect as a complete application).
And you don't even get to use unit tests to check correctness.
You're suggesting the robots can learn the routing algorithms and rules just by looking at a bunch of pictures?
Sure, maybe, given a super-massive amount of data.
I see it as the difference between "I want a photo-realistic portrait of a beagle" and "I want a photo-realistic portrait of my neighbor's beagle, Bob". The first one there's some general rules as to what makes something a 'beagle' so is not too hard while the second has specific constraints which can't be solved without a bunch of pictures of Bob.
To address the specific issue, an AI would have to learn the laws of physics (aka, "Bobness") from a bunch of pictures of, essentially, beagles in order to undertake the task at hand.
I think maybe the best way to get this data set is to subsidize a few dozen electronics recycling centers for every unique microCT scan they send you. Lease them the tomography equipment. They increase their bottom line, you get a huge dataset of good-to-excellent quality commercial PCB designs.
Very fun idea, I had not considered training on existing work (IP is so sensitive I just couldn't think of a way to get enough)
My approach is slightly different for building the dataset. I think we should bootstrap an absolutely massive synthetic dataset full of heuristically autorouted PCBs to allow the AI to learn the visual token system and basic DRC compliance. We then use RL to reward improvements on the existing designs. Over time the datasets will get better similar to how synthetic datasets are produced whenever a new LLM model is released that make training subsequent LLMs easier.
I think people are underestimating the number of PCBs that are needed to train a system like this. My guess is it is well over 10m PCBs with perfect fidelity. It will make sense to have a large synthetic data strategy.
Before you splurge on hardware to extract data it would be much cheaper and faster to just buy it in Shenzhen. All the Apple stuff has been reverse engineered, this is how apps like ZXW have scans of all pcb layers. Random google search https://www.diyfixtool.com/blogs/news/iphone-circuit-diagram...
I think routing is basically an image transformer problem without a good dataset right now. If the eye of the giant AI companies turned to autorouting and large synthetic but physically simulated circuit datasets were created, we would basically be done and autorouting would be a solved problem.
This means that all the work I’m doing now on heuristic algorithms, as well as all the hard work done by humans, will probably not be needed in the future. I just don’t see autorouting as being more difficult (in terms of solution space size) than the art being produced by transformer models right now.
I’m saying this because you’re right, these heuristic algorithms can only get us so far- the problem is really difficult. But our human intuition, the magic black box operation we do, doesn’t seem to be too far beyond the emerging transformer image models.