Unfortunately as soon as there is a tactic in the position (look at the simple mate in 2 puzzles) maia is unable to correctly predict what 99% of players above a quite low level would play and instead focuses on positional moves. While the positional evaluation is very interesting and no doubt useful, it makes maia as an opponent useless above a certain level since any tactic will bring it down unlike the equivalent human. The way to win would be to create simple threats which wouldn't fool any human, but the engine remains oblivious. In analysis mode one could try to reconcile stockfish with maia to determine a human-like move, but again it's not clear whether at that level a human would play stockfish's move without the human solving the tactic oneself (i.e. it could be so deep or with so many options most humans would not take the risk and requires judgement based on knowing what other players are like).
This perhaps could be solved in a different way. While there might be a neural net / pattern recognition method (there must be as that's what our brain does for the intuitive bit) which would be worthy of research, and more heuristic approach (what we do after) with some tuning based on level might better propose tactical solutions. Even at the 1200 level players are aware of checks, captures, threats. The latter are harder to recognise, and would be a challenge in many cases. Finding all checks and captures in a position is trivial, and from that a tree could be built which then the positions during the tree could be analysed by maia (brute force but not considering checks, captures, threats not everything a-la stockfish). The simple 'I go here, he goes there...'. Threats could be determined potentially by giving weight to features in the position such as creating alignment or forkable positions, removing defenders or adding attacks, creating a mating net (restricting squares etc). There then needs to be some kind of temporal penalty to make it more human like (effectively the deeper or more complex the longer it takes for the calculation to happen - tunable parameter or switchable off e.g. during non-clock play or when not wanting the same 'tempo' of play as against a real human), perhaps also time management can be tunable as well so time-trouble or having to bail-out can be simulated based on whether the calculation seems to be yielding results and whether the outcome will be devastating enough to be able to finish in the remaining time vs that understanding and personality). The deeper and the more thick the branches, then the temporal penalty worsens but also the fog and chance of making a mistake (perhaps some study of such positions in games could reveal to how to best pick wrong moves or when to fog). Often tactical ability or being able to solve a problem is based on intuition as well which might weight how long it takes, the fog, or whether the tactic is solved (for example not instantly seeing pieces or threats far from the action is an intuition failure which is the cause of many players' tactical difficulties below master level as is not getting much of the answer via intuition and having to rely on slow manual search).
There also has to be the same process in defence/handling counter-attack, threats or tactics recognised and looking for ways to break the tactic, counter attack or take the best of several bad options. This could be done as my threats, then his threats and a reconcile and trying to interleave the two, or at strong levels trying to do both together (not sure here how professionals calculate this sort of position).
Finally there has to be some reconciliation between the positional (existing) and tactical evaluation, perhaps factoring in position type likelihood (is it a tactical position - lots of things interacting or a positional one with no clear arena of action), playing style and rating strength to whether the tactical route is taken or the positional one is taken. In some positions the positional evaluation won't be considered (king hunt etc) or in others (pieces swap, pawn structure gets damaged) it is necessary again applying some kind of fog factor.
Unfortunately as soon as there is a tactic in the position (look at the simple mate in 2 puzzles) maia is unable to correctly predict what 99% of players above a quite low level would play and instead focuses on positional moves. While the positional evaluation is very interesting and no doubt useful, it makes maia as an opponent useless above a certain level since any tactic will bring it down unlike the equivalent human. The way to win would be to create simple threats which wouldn't fool any human, but the engine remains oblivious. In analysis mode one could try to reconcile stockfish with maia to determine a human-like move, but again it's not clear whether at that level a human would play stockfish's move without the human solving the tactic oneself (i.e. it could be so deep or with so many options most humans would not take the risk and requires judgement based on knowing what other players are like).
This perhaps could be solved in a different way. While there might be a neural net / pattern recognition method (there must be as that's what our brain does for the intuitive bit) which would be worthy of research, and more heuristic approach (what we do after) with some tuning based on level might better propose tactical solutions. Even at the 1200 level players are aware of checks, captures, threats. The latter are harder to recognise, and would be a challenge in many cases. Finding all checks and captures in a position is trivial, and from that a tree could be built which then the positions during the tree could be analysed by maia (brute force but not considering checks, captures, threats not everything a-la stockfish). The simple 'I go here, he goes there...'. Threats could be determined potentially by giving weight to features in the position such as creating alignment or forkable positions, removing defenders or adding attacks, creating a mating net (restricting squares etc). There then needs to be some kind of temporal penalty to make it more human like (effectively the deeper or more complex the longer it takes for the calculation to happen - tunable parameter or switchable off e.g. during non-clock play or when not wanting the same 'tempo' of play as against a real human), perhaps also time management can be tunable as well so time-trouble or having to bail-out can be simulated based on whether the calculation seems to be yielding results and whether the outcome will be devastating enough to be able to finish in the remaining time vs that understanding and personality). The deeper and the more thick the branches, then the temporal penalty worsens but also the fog and chance of making a mistake (perhaps some study of such positions in games could reveal to how to best pick wrong moves or when to fog). Often tactical ability or being able to solve a problem is based on intuition as well which might weight how long it takes, the fog, or whether the tactic is solved (for example not instantly seeing pieces or threats far from the action is an intuition failure which is the cause of many players' tactical difficulties below master level as is not getting much of the answer via intuition and having to rely on slow manual search).
There also has to be the same process in defence/handling counter-attack, threats or tactics recognised and looking for ways to break the tactic, counter attack or take the best of several bad options. This could be done as my threats, then his threats and a reconcile and trying to interleave the two, or at strong levels trying to do both together (not sure here how professionals calculate this sort of position).
Finally there has to be some reconciliation between the positional (existing) and tactical evaluation, perhaps factoring in position type likelihood (is it a tactical position - lots of things interacting or a positional one with no clear arena of action), playing style and rating strength to whether the tactical route is taken or the positional one is taken. In some positions the positional evaluation won't be considered (king hunt etc) or in others (pieces swap, pawn structure gets damaged) it is necessary again applying some kind of fog factor.