Factorio Learning Environment – a benchmark that tests agents in long-term planning, program synthesis, and resource optimization

One thing I hope we get out of AI in the future is the ability to train and create "human-like" bots for multiplayer games that don't natively support them.

One night, my ISP experiences a problem, and I lose internet access for a few hours. I really want to play Deep Rock Galactic. Sure, I could load up a solo mission with Bosco, but what I'd really like is to play with a full team of players. It would be cool if I could just spin up three "agents" and have them join my game. Especially if those agents could play with the same skill and occasional incompetence as real players.

Or, if I manage a cooperative Insurgency: Sandstorm server, it would be cool if whenever there are (real) players on the server, the amount of open slots, minus 2 (to leave open slots for real players to join), are automatically filled with semi-competent agents. (I bet Sandstorm's voting system could even be modded, so players in the server could decide how many allied agents they want to play with.)

In terms of capability and performance, things aren't there yet, and I don't imagine they will be for a long time. I'm still hopeful that a lot of old and "dead" games will be "revived", because people will be able to create artificial player bases of AI bots that play with varying degrees of "human-like" skill.

Eve will never die, it'll just increasingly be computers fighting computers.

I know there's potential for that future to turn sour (well, sourer than this one), but I choose hope rather than despair.