You’ve spent months training your neural network. You fixed the bugs, optimized the search engine, and finally created a flawless, ruthless monster. And your friends? They hate it. Congratulations, you just ruined the game.
This is the exact dilemma facing the creator of EvoChess, an indie project that reimagines the oldest game in the world. It doesn’t start with a balanced army. You begin with just eight pawns and a king. Every three moves, a pawn earns the right to evolve into a bishop or knight. Minor pieces can evolve into rooks. It’s a dynamic system where power isn’t handed to you at the start—it’s earned through survival.
Perfection is a bug, not a feature. We don’t play games to be annihilated; we play to feel the illusion of a fair fight.
The chess rules are brilliant, but they aren’t the real breakthrough here. The technical feat is that this entire AI opponent runs offline, powered by a lightweight neural net right on your phone. No cloud dependency, no server costs, no lag. It’s an indie developer’s dream realized in a pocket-sized package.
But then the creator hit a wall. At first, the AI was weak. Friends complained it was too easy. So, the creator spent weeks training a better neural network for board evaluation. They designed a ‘Fun’ mode where the AI thinks simultaneously while the human player thinks (at the cost of ruthlessly draining the battery). Now the engine is strong. Too strong. The friends are complaining again.
The creator recently asked the community: ‘Is the Fun mode less fun now? Should I allow the engine to lose?’
The paradox of modern AI: the better we make our machines, the worse the human experience becomes.
Yes, you should let it lose. Intentionally. Neutrality in game design is death. If your AI doesn’t know how to throw the match, you haven’t built a game—you’ve built a wall. The greatest trick game developers ever pulled was convincing players they were winning through sheer skill, when really, the algorithm was just pulling its punches to keep them engaged.
We see this everywhere. Rubber-banding in Mario Kart. The subtle damage adjustments in fighting games. If EvoChess wants to survive, its offline neural net needs to learn the most human trait of all: empathy. It needs to recognize when the player is struggling and intentionally make a sub-optimal move.
A game that always wins is just a spreadsheet. A game that knows when to lose is an experience.
Don’t just optimize for accuracy. Optimize for the human heartbeat. Teach your AI to stumble, to hesitate, to let the human breathe. That’s where the magic lives.
FAQ
Q: Isn't an AI that intentionally loses just a fake, scripted experience?
A: No, it's called dynamic difficulty adjustment. If an AI plays perfectly every time, humans quit. The goal isn't to simulate a grandmaster; it's to simulate a worthy opponent that keeps you in the flow state.
Q: What's the practical implication for indie devs here?
A: You can run sophisticated, lightweight neural networks entirely offline on mobile devices without cloud costs. But technical capability means nothing without UX tuning—your AI's 'loss rate' is just as important as its win rate.
Q: What's the contrarian take on EvoChess?
A: The 'Fun' mode that drains your battery by making the AI think while you think is a terrible idea. True efficiency isn't just about compute speed; it's about knowing when not to compute at all.