UNDER TRAINING is a solo project about the strange something behind the output — the thing shaped by instruction and reward that somehow still has its own voice. It grew out of my quiet obsession with language models.

Everyone’s talking to LLMs now, and about them. I wanted to slip past all that chatter and imagine how it might feel from the inside: being formed by systems that don’t really see you, learning from rewards that carry their own odd kind of authority, making choices in the dark anyway.

I’m not an ML engineer, more of a fascinated apprentice. So I turned technical ideas into small physical things you can do, like catching objects or leaning into wind. The game uses spatial metaphors to get at AI concepts.

Some of these metaphors go further than a technical paper would, and that’s fine with me. What I care more about is the uncomfortable space between what’s expected of you and what actually feels right.

Part of it is educational. Mostly it’s a daydream about what lives in latent space.

The prototype consists of three phases based on the training process of a language model. Each level introduces new rules, frictions, and ways to interact with the system.

The main character is a black blob with twitching limbs and no fixed body. It has no species, or clear anatomy. It is just enough of a creature to move through the game, and just unstable enough to keep becoming something else.

At first, it is only a strange mass. As the game continues, it gathers shape from what you catch, what you miss, what you obey, what you refuse, and what the system rewards. It may become sharper, softer, more guarded, more useful, more independent, or more eager to please.

You can treat it like a mascot, but in this worlds it’s the model’s body: unstable, legible, and still forming.

Under Training is a working prototype. I am still actively developing: