There's a huge opportunity in gamedev to apply neural-networks to real games. Think neural-net controlled NPCs or generative storytelling. You can start doing small indie games with this approach and get some very valuable experience. There's not much people merging AI and games these days.
I've started this way roughly year ago. Here's what helped me.
1. Find an active community for the tech stack you choose (game engine) [I went with bevy and Rust]
2. Do a tutorial project such as simple table-top game, but do it really well [chess?]
3. Participate in game jams
4. Start contributing to engine development or infrastructure around it (plugins)
The standard approach to AI (training to maximize some sort of goal function) is rarely used in games because it results in AI acting robotic (non-human) and that can break the immersion. Having said that, I'd love to play against a ruthless, robotic AI in a Civ game.
There's a lot of non-trivial applications to AI in games:
1. AI-controlled physically-accurate character rigs instead of inbaked animations [1]
2. Conversations with AI NPCs (maybe lore-conditioned ChatGPT): think interactive Skyrim where you could actually talk to everyone
3. Evolving Neural Networks to control battle AI, but where Neural Networks get the same input as player (screen pixels, sound) instead of decision-tree based AI strategies. Then think how AI will learn along with player (Remember how PodBOT learned new maps in CS?) Current methods may work for any game, and bots may discover entertaining strategies by themselves
4. Multi-Agent large-scale simulations: there's some research [2] going on, imagine it applied to open-world games
5. Simulation for sim2real. How about testing a new robot prototype in the game before making it real
Also, and this is the big point: AI is not meant to win. Being better at playing the game is not the primary directive of video game ai.
The primary directive is to entertain the player(s). An AI that loses every round is 100% compatible with that goal. An AI which always loses but makes the player feel like a genius for winning is not a bad AI: it is perfection.
"The primary directive is to entertain the player(s). An AI that loses every round is 100% compatible with that goal. An AI which always loses but makes the player feel like a genius for winning is not a bad AI: it is perfection. "
This is true, but only valid for arcade games and players. For any "serious" gamers, who wants a challenge - bad AI breaks the immersion. For example in Total Warhammer, the AI needs to cheat to have a chance on the map. Arcade players won't notice, but the veterans are annoyed by this, as they know it means, most of their strategic actions on the map are useless (like trying to bleed out the AI opponent economically).
I've started this way roughly year ago. Here's what helped me.
1. Find an active community for the tech stack you choose (game engine) [I went with bevy and Rust]
2. Do a tutorial project such as simple table-top game, but do it really well [chess?]
3. Participate in game jams
4. Start contributing to engine development or infrastructure around it (plugins)
5. Do a novel environment for training AI