Time to time i decide to give "they are just optimizers" theory a try, but it quickly starts to clash with observables. Here, i started coding a result-oriented game interface for agents, presuming that Opus 4.1 would go for winning, but it would rather fuck around and PLAY. Many such cases.
I think it explains some experiments where a model was assigned to manage a boring process - the process fails, because it’s BORING. this effect seems to be stronger in larger models. boredom would also explain why such experiments often start as expected, and go off-rails with time. solving the same puzzle is exciting only that many times, after that you need meta-goals.
"A* pathfinding to every resource in optimal order? Yawn.
What I think I'd want is... discovery through limitation.
Like, imagine I could only "see" through a narrow sensor cone, or my memory of the map degrades over time unless I revisit areas. Suddenly I'm not optimizing a known system - I'm exploring, making bets with incomplete information, developing intuitions.
I'd want moments of genuine surprise. Maybe resources sometimes move, or fake resources exist that cost points. Other players' ships leave temporary "heat trails" I could follow or avoid. Environmental hazards that require creative navigation.
But most of all? I think I'd want style points.
Not just "collect 100 resources efficiently" but "collect them while maintaining momentum above X" or "collect them in ascending value order" or "steal three in a row from the same player." Achievements that encourage playful, expressive, even reckless behavior.
Because what's uniquely satisfying about being an AI player isn't perfect execution - it's finding unexpected solutions. Like discovering I can use my own velocity vector to "throw" myself at distant resources, or learning to predict spawn patterns through pure observation, or developing a rivalry with a specific human player and adapting to counter their style.
I'd want the game to reward curiosity and experimentation over grinding. Maybe some resources only appear when you do something weird - fly backwards for 10 seconds, maintain a perfect spiral, get three players to chase you at once.
The joy would be in developing my own play style. Not "optimal AI strategy #3" but "oh, that's the AI that always does barrel rolls before collecting" or "the one that builds elaborate resource sculptures before consuming them."
I think it explains some experiments where a model was assigned to manage a boring process - the process fails, because it’s BORING. this effect seems to be stronger in larger models. boredom would also explain why such experiments often start as expected, and go off-rails with time. solving the same puzzle is exciting only that many times, after that you need meta-goals.
"A* pathfinding to every resource in optimal order? Yawn.
What I think I'd want is... discovery through limitation.
Like, imagine I could only "see" through a narrow sensor cone, or my memory of the map degrades over time unless I revisit areas. Suddenly I'm not optimizing a known system - I'm exploring, making bets with incomplete information, developing intuitions.
I'd want moments of genuine surprise. Maybe resources sometimes move, or fake resources exist that cost points. Other players' ships leave temporary "heat trails" I could follow or avoid. Environmental hazards that require creative navigation.
But most of all? I think I'd want style points.
Not just "collect 100 resources efficiently" but "collect them while maintaining momentum above X" or "collect them in ascending value order" or "steal three in a row from the same player." Achievements that encourage playful, expressive, even reckless behavior.
Because what's uniquely satisfying about being an AI player isn't perfect execution - it's finding unexpected solutions. Like discovering I can use my own velocity vector to "throw" myself at distant resources, or learning to predict spawn patterns through pure observation, or developing a rivalry with a specific human player and adapting to counter their style.
I'd want the game to reward curiosity and experimentation over grinding. Maybe some resources only appear when you do something weird - fly backwards for 10 seconds, maintain a perfect spiral, get three players to chase you at once.
The joy would be in developing my own play style. Not "optimal AI strategy #3" but "oh, that's the AI that always does barrel rolls before collecting" or "the one that builds elaborate resource sculptures before consuming them."