I feel like LLMs and GANs viewed as substitutes for creativity is always going to fall flat on it's face. But the same tech viewed as Yet Another Tool for artists to use is progress. As for the ethics of using other's art (without asking) to train the AI? That seems a separate discussion. My stance below assumes that AIs in the near future are ethically sourced, regardless of how we feel about the current ones:
AI-as-substitute-for-creativity is like microwave-dinners-as-substitute-for-cooking-from-scratch. I like microwave dinners just fine, but it's not much of an artistic expression of my chef skills, and it rarely lives up to the flavor and enjoyment of a talented chef cooking a meal from scratch. I'm not much of a cook, so many times cooking from scratch still tastes worse than the microwave dinner.
AI-as-a-tool-for-artists is like... well, photoshop, or any other modern art software. It isn't destroying artistic expression just because I'm using software that allows me to undo a brushstroke, or separate layers. I've spoken to artists that eschew software and only use real paint on real canvas. I can respect the purity of form they desire, but for me, I can make better art in less time for lower costs by using GIMP. IMHO, AI won't replace artists, it will just help us do the tedious parts faster.
👉This may be a bit off-topic, but what I'd really like to explore with neural networks is less about what a fully trained AI can do for gameplay, but rather incorporating an untrained AI into gameplay, and have the players (unknowingly?) be the training data. Start the NPC goblins out with some basic decision-making:
if player detected, then approach player until within attack range.
If player within attack range, attack player.
Players will employ kiting tactics, ranged attacks, et al. Neural network will begin varying each iterative model's (goblin's) behavior randomly in small ways. The models (goblins) that achieve higher success in damaging the players will "breed" for the next generation of models. The more the players progress, the more "intelligent" the goblins seem to become, adapting to all the commonly-used cheesy player tactics. Later, mob difficulty can be adjusted by changing which training generation the players are pitted against.
I can't imagine this is a new concept, so I'm curious if y'all have seen this implemented in any games already. I'd like to see how it works.
P.S. I know there are games out there with "Adaptive AI" or "Dynamic Difficulty", but those all seem like a set of human-coded tactical behaviors pre-programmed for the mobs, and the mobs just select different behaviors based on player action. There is no evolutionary learning involved. The mobs only ever have the same set of pre-coded behaviors. I am talking about neural-network training, where no programmer has defined any specific set of tactics beyond the very basic actions (move, attack, use item, hide, use buff/debuff), and the models produce emergent tactics that may not have ever been considered by the devs. Behavior is influenced by inputs. (mob & player position, observed mob & player damage output, health, speed, [other stats], geographical & environmental features, etc.) and successful behaviors are rewarded and propagated.
