The Dream Machine

The Dream Machine was a 2024 Gervi Labs research project exploring whether an AI system could simulate something akin to human dreaming. The goal was simple: create a closed loop between a language model and an image model, let them run for hours without interruption, and watch what emerges.

The experiment combined ideas from Jungian psychology and modern neuroscience. Jung’s work on symbols, archetypes, and the unconscious provided one lens. Research on the brain’s default mode network, memory consolidation, and emotional processing provided another. We treated this as a frame, not proof. The aim was to observe what happens when two different model types interpret each other again and again.

The system ran for eight hours. Each cycle began with a high-temperature prompt to the language model, which produced a detailed description. A diffusion model turned that description into an image. The image was then interpreted by the language model, which produced the next prompt. This loop repeated 1,532 times.

The result was a long sequence of images and texts that shifted, drifted, and returned to earlier themes. It produced patterns that felt dreamlike without being human. The output was compiled into a video and a dataset for later analysis.

What the project showed was not that AI dreams, but that iterative model-to-model feedback produces surprising behavior. It opened up new ways to think about artificial creativity, symbolic transformation, and long-form generative processes. It also gave us a foundation for future work where AI systems generate, interpret, and evolve content over extended periods.

The Dream Machine sits at the intersection of research and artistic exploration. It was a study in emergence, a technical experiment, and a piece of generative art in its own right.