The Research
Reading and writing the neural language of dreams
The Research
Studying the dreaming mind has been notoriously difficult. Because researchers have had to rely on the hazy memories people share after they wake up, an entire state of human consciousness remained unmapped.
In an experiment published in 2019, with support from the National Science Foundation and presented as part of a collective result from similar research across the world, Researchers at Paller's lab at Northwestern University succeeded in establishing bidirectional communication between lucidly dreaming subjects and the outside world, and presented empirical support that delivering learning associated sound cues during sleep can improve next-day problem solving.
In 2020, Dormio, a wearable device developed by researchers at the MIT Media Lab, was designed to track sleep stages and interface with dreams. It specifically targets hypnagogia — the semi-lucid transitional state between wakefulness and the first stage of sleep (N1). They discovered that active use of their protocol during hypnagogia can enhance memory consolidation. The system operates as follows:
The methods described above have all been incredible breakthroughs. But in order for us to make generational leaps in sleep and dream research, we need to go beyond what's currently published. We need a scalable, non-invasive way of creating a map of human consciousness in the real world and dream world. We need lots of data about neural activity (both real and synthetic), shaped in the right way, that reveals predictive patterns, and we need a way to toggle those patterns in a robust and non-invasive manner.
Under Project Morpheus, we've done just that. We've developed a novel, non-invasive, high-density EEG device for recording brain activity, and novel techniques in machine learning to analyze that activity under sleep, dreams, and virtual reality, and successfully applied our findings to create a new device, Dream Weaver, that can interface with the mind through dreams. Dreams are where the brain speaks most freely, processing emotions, consolidating memories, and composing vivid, surreal narratives out of pure neural activity. Through AI, we've learned to read and write that language, and thus, created what we believe to be the most direct and comprehensive interface to human cognition ever built.
Our research pursued that goal along two fronts:
Reading naturally occurring dreams and translating them into words and imagery.
Composing dreams and delivering them with full sensory fidelity.
Both are powered by Morpheus, our Dream Sensation Model (DSM), the first AI model that can read, interpret, and compose the neural language of dreams.
Our foundational result comes from a breakthrough recording device coupled with innovative techniques in machine learning used to process and classify brain electrical activity from thousands of naturally occurring sleep and dream sessions paired with similar analysis of brain activity under thousands of hours of immersive virtual reality experiences. Using novel, non-invasive, ultra-lightweight, high-density EEG sensors, orders of magnitude beyond current capabilities of clinical EEG, we are able to see through the skull-tissue barrier to detect nerve signals, with single neuron precision, yielding resolutions where dreams stop looking like noise. Through the right lens, we found stable, recurring structure in dream-state neural activity: consistent encodings for vision, motion, sound, emotional valence, and, most surprisingly, touch, proprioception, olfaction, and more. Collectively captured through time, we call these structures Dream Sequences or Dream Frames. They are a representation of experience, the tokens of DSM models, and the vocabulary on which Morpheus was trained.
Reading dreams was half of the puzzle. The second, and arguably even more important, was discovering a way that sleep, REM state, and dream sequences can be initiated, encoded and transmitted non-invasively, to entrain the brain's dream circuitry into the expected state without impairing its natural ability to regulate sleep-wake cycles. This required further AI-driven innovation, which we will detail in a future publication. It is this ability to rapidly transition from wake to REM state to sequence input, safely and non-invasively, that served as a key accelerator for the entire project, significantly expediting data collection and the feedback loop needed for further fine-tuning of Morpheus.
How we built Morpheus 1.0: the dream sequence representation, global traits, per-user calibration, and what we still don't understand about why it works as well as it does.
An interpretability study: following a sensory experience — the warmth of sunlight — from prompt, through Morpheus's encodings, into measured neural response and the dreamer's report.
How DSM frames/sequences compare to LLM tokens, and where they bifurcate.
When humans prompt AIs to prompt humans within dreams, and the extensive guardrails needed to ensure AI alignment.
Our layered defense and unique end-to-end biometric encryption solution.
Creating a more efficient memory store through custom dreams that enhances the brain's natural cleanup routine.
Embedding new knowledge into dream sessions, effectively locking in new memories while you sleep.
Efficacy of repetitive practice within custom dreams vs. real-world counterpart.
How prompting different sensory modalities enhances learning different topic categories.
Our work builds on decades of research into sleep, dreams, and brain-computer interfaces (BCIs):