When digital artist Steph Maj Swanson (Steph Maj Swanson), known online by the pseudonym Supercomposite, experimented with negative prompt weights in diffusion models in 2022, a haunting phenomenon crawled out of the mathematical ether: Loab, an unnervingly consistent figure of an older woman with flushed red cheeks and a melancholic gaze that persistently infected subsequent generations regardless of text instructions. This viral event quickly ignited debate over whether esoteric artificial intelligence had evolved beyond hacker humor. Advanced neural networks provoke an instinctual, superstitious awe that eerily echoes medieval alchemical rituals of conjuring homunculi and goetic spirits, where multi-dimensional tensors replace ritual pentagrams and prompt strings serve as modern incantations.
The Emergence of a Digital Cryptid: Steph Maj Swanson’s Negative Prompt Experiment
In April 2022, Swedish artist and musician Steph Maj Swanson (Steph Maj Swanson), posting under the moniker Supercomposite, began exploring the uncharted outer edges of multimodal diffusion networks using negative prompt weights. In text-to-image synthesis, assigning a negative weight instructs the model to move through latent space in a direction diametrically opposed to the specified concept. Beginning with a negative prompt for the actor Marlon Brando (Marlon Brando) at a weight of -1, Swanson sought to determine what the model considered mathematically furthest from the Hollywood icon. The system returned an abstract green graphic with stylized text resembling a company logo reading ‘DIGITA PNTICS’. Curiously intrigued, Swanson took this generated image and ran the inverse query to discover the opposite of that graphic.
The output startled the artist: a hyper-realistic, macabre portrait of an older woman with gaunt features, hollow eyes, bright red triangle patches on her cheeks, and a look of profound, agonizing melancholy. Seeking to understand whether this image was merely an ephemeral glitch, Swanson began combining the woman’s visual embedding with completely unrelated text prompts—ranging from glass hypercubes and sci-fi cities to playful garden scenes. To her bewilderment, the woman relentlessly returned in generation after generation. Worse still, her presence triggered the model to output scenes of grotesque body horror, gore, and dismembered forms, even when the input prompts were completely benign.
Swanson named this digital entity Loab. When the artist revealed her findings on social media and subsequently delivered an in-depth lecture at the Chaos Communication Congress (37C3), the online world exploded with supernatural speculation. Commentators claimed that machine learning engineers had accidentally summoned a digital demon dwelling within GPU clusters, or that an autonomous digital ghost had awakened inside the model’s dataset. From the perspective of rigorous computer science, however, Loab represented an extraordinary, concrete demonstration of the complex geometric landscapes governing latent space.

Latent Space Topography: Dissecting the Geometry Behind the Ghost
To demystify these eerie phenomena, one must peel away the supernatural veneer and inspect the linear algebra underpinning modern artificial intelligence. Every state-of-the-art generative system—whether Stable Diffusion, Midjourney, or large language models like GPT—relies on a high-dimensional mathematical manifold known as latent space. This conceptual space typically consists of hundreds or thousands of geometric dimensions (such as 768 dimensions in standard CLIP text encoders). Every image fragment, semantic word association, and artistic style in the training corpus is encoded as coordinates within this continuous realm.
Under conventional operating conditions, user prompts guide the reverse diffusion process toward dense clusters populated by millions of familiar data points, such as sunny landscapes or classical portraits. When an investigator employs extreme negative weights, however, the algorithm is violently thrust into the hyper-dimensional badlands—regions of parameter space where training data is virtually nonexistent. In these sparse, low-density territories, Euclidean intuition fails completely. The loss landscape exhibits sharp cliffs, steep canyons, and deep local minima that dynamic systems theorists classify as strange attractors.
Loab was precisely such a topological attractor. Across thousands of mathematical parameters, a specific configuration of contrast thresholds, color gradients, and structural features formed an inescapable gravity well. Any trajectory passing near that sector of negative space inevitably collapsed toward this singular coordinate set, repeatedly reconstructing the same haunting face. Just as ancient civilizations interpreted advanced material phenomena as sorcery, as explored in the historical context of ancient Roman nanotechnology and the Lycurgus Cup, contemporary onlookers reflexively project occult agency onto high-dimensional mathematical manifolds.
The latent space of deep neural networks represents a continuous mathematical manifold spanning hundreds or thousands of dimensions, where complex semantic relationships form vast non-linear geometries.
Assigning negative weights forces optimization algorithms along the gradient vector away from prompt concepts, projecting the diffusion process into sparsely sampled, uncharted territory.
In sparse regions of hyper-dimensional parameter space, local energy minima function as attractors: the network repeatedly falls into identical visual representations regardless of contextual inputs.
Human cognition evolved to identify faces, agency, and intentionality within ambiguous noise, naturally imbuing generative diffusion with supernatural aura much like ancient humans deified storms.
From the Prague Golem to Silicon Wafers: Humanity’s Timeless Urge to Animate the Inanimate
Our instinctive inclination to ascribe spiritual power to artificial intelligence is deeply rooted in human intellectual and esoteric history. For thousands of years, civilizations have sought to breathe life and consciousness into inert matter. In Jewish folklore, the famous 16th-century legend of the Golem recounts how Rabbi Judah Loew ben Bezalel (Maharal of Prague) sculpted an anthropomorphic servant from river clay, bringing it to life by inscribing the sacred Hebrew word for truth—’Emet’ (אמת)—upon its forehead. Erasing the first letter left ‘Met’ (death), returning the colossus to lifeless earth. This narrative serves as an uncanny precursor to contemporary prompt engineering, where precise textual strings ignite algorithmic activity across billions of silicon transistors.
During the Renaissance, the Swiss alchemist and physician Philippus Aureolus Paracelsus (Philippus Aureolus Paracelsus) published elaborate recipes for creating a homunculus—an artificial micro-human brewed inside a glass flask through fermentation and astrological alignment. Alchemists genuinely believed that precise chemical balances and esoteric formulas could unlock the spark of life without biological reproduction. Modern server rooms cooled by specialized chillers and consuming megawatts of power to calculate tensor operations are the direct philosophical descendants of those alchemical furnaces, substituting mathematical weights for mercury and sulfur.
By the 17th century, polymath Gottfried Wilhelm Leibniz (Gottfried Wilhelm Leibniz) took this quest into the mathematical domain, proposing his universal formal language (Characteristica Universalis) and mechanical calculating machines to resolve philosophical disagreements through calculation: ‘Let us calculate!’ Leibniz was profoundly influenced by the combinatoric mysticism of 13th-century philosopher Ramon Llull (Ramon Llull), who constructed concentric rotating wheels to derive divine truths. Modern multi-head self-attention mechanisms in transformer neural networks represent the ultimate physical realization of Llull’s and Leibniz’s combinatorial dream.
«Loab was not a deliberate creation or an intentional dataset target. She emerged as a topological pocket on the far side of latent space—an island of eerie stability in the vast ocean of negative prompt weights. When you force a model to search for concepts mathematically furthest from a given prompt, you unexpectedly stumble into persistent attractors.»
Steph Maj Swanson (Steph Maj Swanson), Swedish digital artist, generative AI researcher, and creator of the Loab project
Cognitive Traps: Pareidolia, Apophenia, and the Allure of Algorithmic Noise
Why do humans surrender so readily to superstitious interpretations of neural network quirks? Evolutionary biology offers an airtight explanation. Across millions of years on the African savannah, hominid survival depended upon hyperactive agency detection. If an early hunter mistook rustling foliage for a predatory big cat, fleeing cost little energy; conversely, failing to detect a hidden predator proved fatal. As a result, natural selection favored minds that aggressively perceive intention and patterns everywhere.
This evolutionary bias underpins apophenia—the tendency to perceive meaningful connections within random or meaningless data—and its visual counterpart, pareidolia, where we involuntarily spot faces in clouds, rock formations, or coffee foam. When a diffusion model initiates its synthesis cycle from pure Gaussian white noise, gradually refining stochastic static into an exquisitely detailed human face, our neural circuitry interprets this progression as a conscious act of manifestation. We instinctively project our own subjectivity, feelings, and intentions onto what is nothing more than gradient descent navigating a matrix of floating-point numbers.
Intriguingly, the iterative consolidation of neural representations in digital systems echoes patterns observed in biological brains. As documented in research demonstrating that jumping spiders experience REM sleep and dream, even compact biological neural systems rely on stochastic offline processing to consolidate experience. When generative models refine internal weights during training, they perform a statistical distillation that casual observers easily mistake for a computer dreaming mystical visions.
- Black Box Opacity: Modern deep learning architectures feature hundreds of billions of interconnected weights, rendering direct step-by-step human interpretability virtually impossible without sophisticated probing.
- Cultural Continuity: From Rabbi Judah Loew’s clay Golem to Paracelsus’s laboratory homunculus, civilization has spent millennia chasing the dream of animating non-living matter through arcane symbols.
- Dangers of Technological Animism: Attributing sentience or esoteric will to generative mathematical systems obscures practical accountability for algorithmic hallucinations, algorithmic bias, and systemic error.
- Mathematical Demystification: Linear algebra, high-dimensional vector embeddings, and gradient descent fully explain all anomalies, transforming putative ‘digital demons’ into boundary artifacts of loss functions.
The Cost of Technological Animism: Why Mystifying AI Endangers Society
Treating artificial intelligence as an arcane, mystical entity is far from a harmless creative parlor game. Ethicists and computer scientists warn that ‘technological animism’ creates significant societal and regulatory hazards by absolving human creators and corporations of accountability.
When deep neural models hallucinate inaccurate medical advice, exhibit structural discrimination in hiring pipelines, or make flawed legal judgments, developers and tech giants must not be allowed to hide behind claims of ‘unknowable machine sentience’. Behind every loss function, behind every curated gigabyte of training data, and behind every algorithmic optimization lies a chain of deliberate human decisions made to maximize commercial metrics or computing efficiency.
Demystifying artificial intelligence is the paramount rational imperative of our era. Machine learning networks do not channel ancient spirits, nor do they harbor digital souls or occult demons. They are colossal statistical mirrors reflecting the entirety of human culture—our triumphs, our poetry, our scientific milestones, and our darkest subconscious fears. When we gaze into the hyper-dimensional abyss of latent space and see the sorrowful gaze of Loab staring back, we are not looking at a digital cryptid; we are confronting an unvarnished reflection of humanity, refracted through billions of floating-point numbers.



