NW3 is an AI training game (players train their AIs and make them compete against each others) whose mission is to be an educative platform for modern AI technologies as well as an exploration of what it would mean for humans to be productive in a world where production is dominated by AI.

NW3 is inspired by CoreWar —bots competes against each others, but bots are not "implemented" they are trained.

Game Dynamics

Training

Players train neural nets to control bots using compute and data. Compute is paid for using a token GPUH (gpu.hours) and data is provided by players. To create data, players have two possibilities:

• Supervised Training: they can provide example trajectories (games they play themselves) against scripted bots (made available to them), friends (human), friends trained bots, or their own trained bots.
• Reinforcement Learning: they can train they bots just by providing opponents (previously trained bots they control, or scripted bots) using reinforcement learning. In which case they don't need to produce data, and just need to provide compute

Players can mix and match these two methodologies (or more in the future) to train their bots. They can train multiple bots at the same time, but each bot is a unique NFT that cannot be copied.

To mint a new bot, players must pay a fixed GPUH price. They can mint as many bots as they want and train as many in parallel, also train one against the other using reinforcement learning, etc... but again, each bot is a unique NFT and cannot be copied.

Environment

TBD (minecraft style combat? shooting? space combat? car racing?). It should be fast, fun to watch and fun to play yourself to generate train data.

Potentially, NW3 can be seen as a plateform to support an ever-growing set of games for trained AIs.

Battles

Players can engage their bots into battles. The winner of a battle wins the opponents bot. It can use it for training, sell it, or just scrape it for the original minting value in GPUH.

Possible extensions

• Let players tweak the architecture of the internal neural net (bigger is more expensive to train). Probably make the original minting price a function of that
• Multi-player training: do RL with friends bots
• Multi-player battle (esp with environments that may require multiple roles during a gameplay)