Dreaming in Code
Dreaming in Code
The Shared Syntax of Lucid Dreaming and AI
Dear friends,
I’ve been long aware of how blessed I am by the presence of many inspiring and utterly weird (in the best way possible) people in my life. I decided to start a new series of guest interviews, exploring some of my friends’ works relating to my research.
In today’s piece, we look into the posible utilisation of the lucid dreaming space for learning and exploring complex mathematical concepts, such as AI algorithms. We touch on organic and digital information compression, the nature’s optimisation functions and many more rather esoteric takes on machine learning topics.
I would like to introduce Mariana. We met, as it goes, thanks to yet-another-magical synchronicity that crosses the paths of two internet strangers. A time loop, if you will. She invited me to a wonderful retreat on Mallorca called “If The Land Could Speak”, a week of diving experientially into the concepts of non-human intelligence, opening up our perception to the weirdness that AI has become.
Until recently, Mariana was an emergent futures researcher and educator in programmes seeking to rethink the way we design, use and produce tech. Now, after nearly a decade of cultivating and exploring conscious dreaming practices, she has opened the lucid.school community as a meeting space for individuals interested in integrating these practices into their lives.
Every now and then we meet and compare notes on our similar lines of research, and a few days ago when we had our catchup, I was so astonished with her work and the resonance with my projects that I asked her to share the ideas with us in a form of a (semi-)interview.
Karin: Hello Mariana, I’m really looking forward to hearing more about your research into AI and Lucid Dreaming. Could you tell me a bit about yourself, and what's your connection to dreaming?
Mariana: I've always been fascinated with computing and creativity since I was very little. This led me to pursue a life in media arts, digital literacy, open-source tech, and more recently, ML and generative media.
About a decade ago, I started working actively with my dreams as I committed to deal with serious mental health issues that I had struggled with since my teens. I didn't even know about lucid dreaming, but it seems that when you actively start engaging with this aspect of your mind, the path is somehow clear for this phenomenon to occur. Spontaneous out-of-body experiences ensued as well, and my whole notion of reality came crumbling down. I was terrified, but my curious spirit rose to the challenge, and I've been a committed sober psychonaut ever since.
Karin: How does AI come into this?
Mariana: When all of this was happening, my interests and work as an emergent futures researcher led me to start engaging with AI more actively. That was about 7 years ago, and I slowly started to try to learn as much as I could. I soon discovered the learning curve was as difficult as it is, to understand the math and the science behind it all, but this is when I started to suspect that everything that was happening to me in my private life somehow had everything to do with what I was trying to research.
Karin: Ever since we met for the first time, I was always impressed how deeply you understood the fundamental concepts of AI without actually having any mathematical or computational background. I was so excited when you shared with me that you’re using your Lucid Dreaming skills to explore these ideas! Could you please share with me a few of the examples?
Mariana: Many years ago when I started having lucid dreams I had the most astonishing experience. I became lucid in an abstract room where there was only a cup lying on the floor. The extraordinary thing was that I could somehow perceive the cup from all of its angles. I could "see" the front, the back, the bottom and the top at the same time. I became aware of what was happening, that this had to be the closest thing to a high-dimensional perspective.
Karin: This is very fascinating. Was it prompted by your interest in AI back then? Do you think you ‘brought in’ the question into your lucid dreaming?
M: I was learning ML in my waking life, and for the first time, I had an inner understanding of what “high-dimensionality” could actually mean as a felt experience, as a cognitive possibility. This, at least, offered a taste of it, and it came in the form of a very intense knowing that I could not possibly access with my waking cognitive capacities. And of course it opened an immense question in my mind about the possibilities of the lucid environment, a space beyond everything I thought I knew about cognition, and that I would investigate tirelessly for many years to come.
K: Did you have any other experiences with learning in lucid dreaming states afterwards?
Mariana: Yes dozens! I’ve been compiling a massive archive over the years. Another clear example I can think of comes from a lucid dream I had recently. It was an experience that didn’t come in the form of a sequence, but more like an abstract scene. I became highly lucid when I felt the intensity of what was happening.
I was looking for a house, or I needed a house, that was the idea that formed the scenery I seemed to be in. When I became aware of the space I was in, it struck me so hard that I gained some kind of in-situ understanding of what I might be experiencing there. I found myself in an abstract space that felt like a massive topography, reaching infinite depths of my field of vision. In that abstract field, I could see houses, floating one next to each other, separated a bit, and they all represented possible houses, with small variations. What I understood at the moment is that I was looking at a sort of embedding of a latent space of all possible houses. They resembled each other a bit, but I understood it was just a "neighbourhood" of similar archetypical houses, similar to the American suburban house, with a nice porch and one or two stories... I tried to see as far as I could but all I could see was that topography curving slightly towards the horizon and reaching as far as I could see, displaying only that, houses, over a black high-dimensional void.
Karin: This really reminds me of the face-morphing people experience in dreams or in psychedelic states. I always felt that it's somehow related to a way our brain, and possibly AI as well, compresses information - a fast-flicking journey through the latent space of possible faces. Interestingly, when generating AI animations with some of the older models that can’t keep context for the objects and faces too well, a very similar visual effect arises.
Mariana: Definitely! The face-morphing phenomenon has occurred in my experiences from the very beginning. It was actually very frustrating because I wanted to draw the important characters that had a serious impact on me during these lucid experiences. When I was trying to focus on their face, it would start morphing like crazy, even in a gender-fluid way, but an interesting phenomenon has happened over the years: I have been able to draw portraits of these repetitive characters, sort of like a stable version from all the fluid interpolations.
It's as if each of those characters had its own embedding of faces and bodies where they move, but over time, a 'centroid' or 'mean vector' as I understand it’s called in ML, meaning an average vector of their appearances, has coalesced in my mind’s eye. This central vector acts as a stable representation distilled from all the fluid interpolations of their many appearances in my dreams. This is highly interesting to me because it resembles the way archetypes emerge from a vast embedding that can manifest in variable configurations and appearances. The image of Jesus is an excellent example of this.
Karin: I remember you mentioned the phenomenon of Uncanny alien familiarity. It’s the experience of finding oneself in a setting that feels unmistakably like a childhood home, yet upon closer inspection, it’s not a perfect replica, but rather a composite, blending elements from different houses, periods or locations into a singular, coherent space. But somehow, it still feels distinctly familiar. As you mentioned, this feels exactly like generative models in AI, which create new data instances similar to but not exactly replicating the data they were trained on. Could you please expand on this a little?
Mariana: This experience where the familiar is somehow mixed with the alien to create a hybrid dream environment seems to be the typical experience we all have. It reminds me too of the concept of embeddings and vectors within the latent spaces of machine learning models.
In ML, embeddings serve to represent high-dimensional data (such as memories or images) in a lower-dimensional, compact form within a latent space (the scenery of the dream in this example). This process allows for the capturing of essential features and relationships between data points in a way that can be processed and analyzed by algorithms. When generating new data instances, such as images or text, a model navigates this latent space, often interpolating between known data points to create something new yet grounded in the learned patterns and features.
The dream of a "mashup" childhood home can be compared to a point in the latent space that is a vector combination of various memory embeddings. This point represents a new "instance" that is not an exact replica of any single memory but rather a chimera that borrows elements from multiple sources, creating a sense of familiarity. This process is similar to how generative models might produce a new image that shares characteristics with multiple training images, reflecting a blend of their features.
To me, this analogy is highly relevant because it might show the inner workings of how both the human mind and machine learning algorithms synthesize and recombine elements from our experiences and memories into new constructs. To me, dreams can be thought of as exploring a latent space of your psyche, where each dimension represents different fears, desires, or memories. Lucid dreaming allows you to navigate this space more consciously.
Karin: Yes, this makes me think of my favourite example with the Animorphs. Our perceptual dataset contains a single point for a ‘boy’ and a ‘snake’ but once these are embedded in the semantic space, there exists an infinite interpolation between these two concepts. You can also do all sorts of crazy mathematical operations on the meanings, such as adding or subtracting them as vectors.
Mariana: Yes, absolutely! I think this seemingly inherent quality of our creative capacity is what allows us to conjure up the surreal, the possible. I believe the surrealists would be absolutely thrilled with the arrival of neural media. They too were trying to dignify the status of the latent space as a reality in its own right. Just because something isn't manifest, that doesn’t mean it isn’t real, and I agree with this perspective. In dreams and generative ML, we are able to navigate the expansiveness and fluidity of vector interpolations in a way that is not possible in the slow experience of matter. I have this suspicion that we can see this interpolation happening in evolution, over millions of years. An orchid shaping itself over generations to be closer to the appearance of an insect for example. I do believe this interpolation exists between species, and that we have only been so blind as to categorize them in separate boxes. After all my experiences I’m more in line with the philosophies and wisdom traditions that posit that what we perceive as separate species might be manifestations of a single fluid latent space. It might be that a form becomes manifest here or on another planet when the relations of the ecosystem allow for a specific manifestation to unfold within ALL the possible interpolation of creative/genetic features.
Karin: This is all so fascinating. I really feel that AI gives us a whole new metaphor to think about the organization of information in the universe and inside our brains. Bridging the gap between the digital and rigid versus analogue and organic representation and information compression. Up until now, this gap was the main problem when dealing with computers - how to get the data into the right structures required by the algorithm. One missing coma and you can drown in the sea of critical errors and exceptions. The machine needed to have everything prepared and preprocessed in order to do the simplest thing, whereas the brain seems to operate magnificently with incomplete, muddled or entirely confusing information.
Mariana: When we explore the similarities between lucid dreaming and machine learning (ML), an astonishing parallel emerges in the capacity for both our human minds and ML models to generate complete, detailed scenes from sparse or incomplete inputs. Think of the phenomenon in dreaming where you find yourself in a vividly detailed train station, despite the dream not providing explicit details for every element of the scene.
There is a technique called image inpainting, or completion, that involves ML models predicting and reconstructing missing or damaged parts of an image to produce a coherent whole. In the same way, super-resolution techniques enable models to transform low-resolution images into higher-resolution ones, adding details that weren't originally there. These processes are an example of the models' ability to interpolate, using existing data to infer and generate the unseen or the incomplete.
This ability of ML models to "autocomplete" or enrich scenes might be very similar to the way our dreams construct complex, coherent worlds. Dreams have the extraordinary capacity to conjure detailed environments, complete with all its nuance, from the nebulous and fragmented inputs of our subconscious minds.
Karin: One important thing I would like to ask you about are of course… the hands.
Mariana: Oh, the lucid hands! There is a classic technique that many use to gain lucidity in their dreams. There are many tricks, but this particular one involves looking at your hands several times per day for a couple of weeks, each time asking yourself, "Am I dreaming?". The goal is for this practice to become a habit in waking life that hopefully translates to nocturnal activities and integrates into your dreams. So, one day —and this is a well-documented trick because many report the same phenomenon—you find yourself looking at your hands in a dream, and two things happen: first, a part of you associates the action with the question "Am I dreaming?", ushering in a wave of awareness. The second aspect, where high weirdness enters the scene, is that your hands may appear very odd.
People, myself included, report seeing extra fingers, wobbly hands, hands of different genders, strange-colored skin, or altogether weird morphologies that trigger a definitive suspicion that this is not waking life, confirming that you are indeed dreaming. I find this example fascinating because when the first AI-generated images of hands appeared on the internet—with this phenomenon of weird-extra-fingers and odd morphologies—it prompted me to ponder the relationship between the two.
Karin: Interestingly, the reason why hands appeared to be so problematic for the early AI is the fact that the models have no functional understanding of human morphology - they’ve just seen millions of pictures of hands in many different positions, and if the task at hand is just to ‘finish the image in the most probable way’ a couple of fingers up and down isn’t really a problem as long as it looks like hand. This is actually happening across the whole generated image, if you focus on the details of foliage, tiling, or perspective, all of these are equally ‘broken’, just a… simulacra, without the internal coherence. We just notice the extra fingers because it’s more visible to our human-centric eye. I’m trying to wrap my head around this - is there a deep internal connection with the dream compression or is this just a superficial resemblance?
Mariana: Actually, it's funny that you mention this, because another trick that many lucid dreamers integrate in their lives is to be constantly observant of quirky details, glitches, or something that seems off when you are looking at a scene in your daily life or in your dream experiences. This should work as a tell-tale sign that you are dreaming. For example, if you are in the dream train station that we mentioned before, and in the background you see an Escher-esque stair, or weird foliage in the trees, or something bizarre like a decontextualized cow standing there, a wave of consciousness ushers in and you can realize you are dreaming. Seeing a relative or a pet that has passed away is a great lucidity trigger as well. There was one time I gained lucidity because I noticed that the clouds in my dream had the texture of plastic bags, almost like a low-resolution raster.
Of course, now, with the improvement of adversarial networks, we can generate hands with the correct number of fingers. However, this fluidity in vectors, to me, is highly indicative of what might be occurring in lucid environments. It's important to note that this is not always the case; in my experience, if I am stable and focused enough while lucid, my hands might look almost completely normal, as if the correct vector had stabilized as well!
Karin: Yes, I will think about this next time the light switch doesn’t work :) This was all very fascinating, thank you. Do you have any closing remarks?
Mariana: To me, all of this offers a profound reflection on the nature of human cognition and perception. As I often say, if only more ML and cognitive science researchers would engage in and practice lucid dreaming, it could provide them with an invaluable wealth of experiences to compare their concepts and findings against, in an embodied and extraordinary way.
The fact that these practices, offering a direct experience into the inner workings of cognition, perception, our identities, and ultimately, the nature of reality itself, have always been accessible, yet have drifted into cultural obscurity —especially here in the secular, scientific West— is a massive topic that deserves its own focus perhaps in a future conversation with you. Naturally, it's part of the broader conversation of colonization, but from the angle of a cognitive dominance that placed our dreaming consciousness in a sort of coma. Thankfully, we're beginning to wake up from it, thanks in part to the psychedelic renaissance, the consolidation of depth psychology, the mainstream revival of the occult, and the decolonization movements. I’m overjoyed this shift is happening for us to reclaim the vast territories of our inner landscapes and I’m happy if I can inspire and reinforce this bridge between science, computation and gnosis.
All illustrations were generated by Mariana and Dall-e.
Thank you friends for reading all the way through, I hope you found these ideas as fascinating. You can follow Mariana and her Lucid School, or check out her website. I can’t wait to experience them myself.
Do you have any similar dreaming experiences?
Stay lucid,
k