I wonder whether that could be fixed by using an approach like this [1] [2], to guarantee the model output meets a schema.

[1] https://news.ycombinator.com/item?id=37125118 [2] https://blog.normalcomputing.ai/posts/2023-07-27-regex-guide...

Or two bishops of the same color

That is possible with pawn promotion

This feels like a perfect example of just whacking a neural net on something without actually applying and understanding of the domain and validating things.

There is something wrong with the rendering of the color of pieces : In https://chessdream.ai/?dreamId=ab86d6a2-470a-4238-9181-8b4d3... there are two white kings, and queens. The position is correct though when viewed on Noctie.

Edit : In fact, the kings and queen are neither black or white but a third color : grey

Edit 2: There is probably some issue with the adjust piece count, when opening the same dreamId in multiple windows. As following the same link doesn't have the same number of pieces.

Hey, you're right, adjust piece count is a local modification, so opening the same dreamId in multiple windows would let you view the position with different piece counts (and the link to Noctie uses the local state)

Thanks for the feedback! It's a simple modification to guarantee things like this – for this prototype I thought it interesting though to expose the "raw" output of the neural network, including crazy things like sometimes placing two kings of the same color. I even force the neural network to place more pieces than it intends, if you put the piece slider on max, to see what happens (and often the result is very bad)