Nice idea, totally worth some attention!
There are many possibilities... and Lichess has already been improving a lot recently, so i believe that welcoming ideas and discussing them is a nice way to bring meaningful contributions to the community :)
* 5 different positions might be presented, each player would take (timed) turns discarding one of them, until one position remains; on a few seconds timeout, the selection would be random.
* After using Stockfish in low depth possibly to discard too unbalanced positions, some variation of Maia9 (human-like playstyle neural network built on top of Lc0; search 'maiachess' for details) might be a good alternative for evaluating how fair a position would be for humans to play (specially for complex positions; imagine a tangled mess of active pieces), and small modifications of the position might bring balance without having to generate everything from scratch again (given that randomly generated positions are often highly unbalanced).
* Random positions (instead of database selection) might generate fun games that wouldn't be possible with classical or fisher random chess, and the major difficulty would be on how to generate random positions that are human-fair and fun to play. Possibly a suitable task for neural networks, so this 'experimental variant' might gradually improve with user feedback; but surely it would take hours of coding and valuable computing resources, so it might just not be worth the effort.
* Game phase/complexity selection. Sometimes you're simply willing to train endgames (or anything), but the ones you face in your regular chess experience are obviously biased by your opening choices and playstyle, and also a large portion comes to an end way before reaching an endgame. So such filter might come as a valuable training tool, if not a source of fun and novelty by itself.
All of it would surely demand a considerable effort by the Lichess devs, so it might take a while to plan and think about implementing something like this. But they've already been bringing awesome news with the puzzles (database, dashboard, streak, storm, racer, and counting!), so my solid bet is that Lichess will continue to keep growing in many directions, and this one might pretty well be one of them. :D
There are many possibilities... and Lichess has already been improving a lot recently, so i believe that welcoming ideas and discussing them is a nice way to bring meaningful contributions to the community :)
* 5 different positions might be presented, each player would take (timed) turns discarding one of them, until one position remains; on a few seconds timeout, the selection would be random.
* After using Stockfish in low depth possibly to discard too unbalanced positions, some variation of Maia9 (human-like playstyle neural network built on top of Lc0; search 'maiachess' for details) might be a good alternative for evaluating how fair a position would be for humans to play (specially for complex positions; imagine a tangled mess of active pieces), and small modifications of the position might bring balance without having to generate everything from scratch again (given that randomly generated positions are often highly unbalanced).
* Random positions (instead of database selection) might generate fun games that wouldn't be possible with classical or fisher random chess, and the major difficulty would be on how to generate random positions that are human-fair and fun to play. Possibly a suitable task for neural networks, so this 'experimental variant' might gradually improve with user feedback; but surely it would take hours of coding and valuable computing resources, so it might just not be worth the effort.
* Game phase/complexity selection. Sometimes you're simply willing to train endgames (or anything), but the ones you face in your regular chess experience are obviously biased by your opening choices and playstyle, and also a large portion comes to an end way before reaching an endgame. So such filter might come as a valuable training tool, if not a source of fun and novelty by itself.
All of it would surely demand a considerable effort by the Lichess devs, so it might take a while to plan and think about implementing something like this. But they've already been bringing awesome news with the puzzles (database, dashboard, streak, storm, racer, and counting!), so my solid bet is that Lichess will continue to keep growing in many directions, and this one might pretty well be one of them. :D