Endless Machines

We use AI to let our robots express themselves! If a robot became alive, it wouldn’t come pre-programmed with the perfect way to express its thoughts or emotions — it would have to figure it out along the way. That’s why the hero robots in Endless Machines use AI to give voice to their thoughts and feelings. What they say isn’t always polished or logical. It's quirky, weird, funny, and sometimes unexpectedly touching. It’s not a scripted, perfect dialogue. It’s more like a robot trying to make sense of the world — and itself — in real time. This approach makes their personalities dynamic and unpredictable. For us, that’s what makes AI-driven dialogue a perfect fit for the game: It’s about learning, exploring, and growing — just like the robots themselves. Under the hood, the game runs a small language model that has been custom fine-tuned for each robot's personality. It is instructed to speak through prompt engineering, ensuring the robot stays in character and relates to the game's lore. The model does not require online access. We chose to use SmolLM because we were looking for a language model where the training background, technical setup, and use of data are transparent (open source and based on publicly available texts), and where the model can be run locally. We use this locally hosted language model (LLM) to generate text content during gameplay. What the robots say in the game is defined and moderated by us. We have written a set of foundational texts that form the basis for each robot’s personality, tone, and way of expressing itself. From that foundation, they “build” further and develop over time. Endless Machines does not and cannot collect any data from the person playing the game. All outputs are guided by custom prompt engineering, custom fine-tuning, and are controlled by local code at runtime to ensure appropriate content. More details on the model architecture and dataset can be found here: 🔗 https://huggingface.co/datasets/HuggingFaceTB/smoltalk