I think the language of Jerk off instruction (joi) videos is unique in that it requires the context of bodily experience to be fully comprehensible. LLMs operate through vector embeddings - distributed representations where the spatial/semantic relationship between (for example) 'father' and 'son' parallels that between 'mother' and 'daughter. This mapping of meaning into mathematical space allows machines to mimic and deploy human language with remarkable fluency. But the aesthetics and social-cultural subtext of JOI - its frequent use of double-entendre, intentional subversion of established norms, and assumption of a desiring body on the other end - can cause these machinic understandings to fragment and destabilise.
Joy is an attempt to investigate a few aspects of what people are starting to call the 'GoonVerse' or 'Goon Economy': networked dissemination of content in which sexual suggestion is leveraged to maximise engagement, setting off a chain reaction of desire and sexuality rerouted and amplified through a matrix of algorithmic interpretation and recommendation.
Joy
┌─────────────────┐ ┌───────────────────┐
│ Camera Input │───────▶│ YOLOv8-Pose Model │
│ (Video Stream) │ │ (Pose Detection) │
└─────────────────┘ └───────────────────┘
│ │
▼ ▼
┌─────────────────┐ ┌───────────────────┐
│ Frame Grab │───────▶│ PersonTracker │
│ (1 FPS/Update) │ │(Track Individuals,│
│ │ │ Attributes) │
└─────────────────┘ └───────────────────┘
│ │
│ ┌─────────┐ │
│ │ New │◀─────┘
│ │ Arrivals│
│ └─────────┘
▼ │
┌─────────────────┐ ┌───────────────────┐
│ Scene Analyzer │───────▶│ Engagement Tracker│
│ (Compute Scene │ │ (5D Engagement │
│ Vector - 15D) │◀───────│ Vector - AVG over│
└─────────────────┘ │ time, Sampled) │
│ └───────────────────┘
│
│ ┌───────────────────┐
│ │ Temporal Features│
│◀────────────────│ (Time Since Jump, │
│ │ Recent Clips, │
│ │ Time of Day) │
│ └───────────────────┘
▼
┌─────────────────────────────────────────────────┐
│ Input Feature Vector for Neural Network (42D) │
│ [Scene Vector (15D) │
│ + Candidate Clip Vector (15D) │
│ + Cosine Similarity (1D) │
│ + Avg Engagement Vector (5D) │
│ + Temporal Features (6D)] │
└─────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────┐
│ AttentionMLP (Neural Network) │
│ (42 → 256 → 128 → 64 → 16 → 1) │
│ - Predicts Engagement Score (0-1) │
│ - Attention Mechanism │
│ - Dropout for Uncertainty │
└─────────────────────────────────────────────────┘
│
│ ┌───────────────────────────┐
│ │ Experience Buffer │
│◀───────│ (Store NN Inputs & │
│ │ Predicted Scores) │
└────────┼───────────────────────────┤
│ Training Batch │
└───────────────────────────┘
│ ▲
│ │ Reward Calculation
│ │
│ ┌───────────────────────────┐ │ (Prev vs. Current Engagement,
│ │ Candidate Clip Selection │◀───┤ Dwell Time, Prox Change, etc.)
│ │ (Top-K Cosine Similar, │ │
│ │ NN Scores, Uncertainty, │ │
│ │ Random Exploration) │────┘
│ └───────────────────────────┘
│ │
▼ │
┌─────────────────┐ │
│ Adaptive │ │
│ Threshold │◀───────┘
│ (Adjusts based │
│ on Engagement) │
└─────────────────┘
│
▼
┌─────────────────┐ ┌───────────────────┐
│ Clip Database │───────▶│ Video Playback │
│ (Clip Vectors, │ │ Controller (MPV) │
│ Metadata, Paths)│ ◀──────┤ (Load Next, Skip) │
└─────────────────┘ └───────────────────┘
│ │
│ ▼
│ ┌───────────────────┐
└───────────────▶ │ Audience Reaction │
│ (Implicitly │
│ Influences │
│ Camera Input) │
└───────────────────┘
The main installation with the webcam mirrors a social media recommendation algorithm, in that it uses a two-tower neural network. This is a fancy way of saying that a bunch of content is categorised and tagged to create an nth dimensional vector capable of describing one piece of content's semantic relationship to another piece, this vector then being compared (typically through a process of cosine similarity) to a different vector that represents some sort of input, like scrolls or shares. In our case, the content is transcripts of different types of jerk off instruction videos. I used an LLM to tag these clips and rate them on a scale of 0.000 to 1, according to 15 different values. The values are designed to encapsulate qualities like intimacy, aggression, and demand, and to correspond to the data we can get from the webcam.
The system learns from audience behavior in real time through an AttentionMLP neural network with 53,260 parameters. When a visitor is in view of the camera, YOLOv8 pose detection tracks their position and orientation, generating a 15-dimensional scene vector that encodes behavioral features like proximity to the screen, body orientation, and movement patterns. This live scene vector is compared against the pre-analyzed 15-dimensional vectors of tens of thousands of video clips in the database.
Rather than simply playing the most semantically similar content, the neural network predicts which clips will generate engagement - measured through changes in visitor attention, dwell time, and spatial proximity. The network takes a 42-dimensional input combining the scene vector, candidate clip vectors, their cosine similarity, historical engagement metrics, and temporal features (time since last clip change, recent viewing history). It outputs a predicted engagement score between 0.000 and 1 for each candidate clip.
The system operates as a reinforcement learning loop: after each clip plays, the installation calculates a reward signal based on whether visitor engagement increased or decreased. These experiences - the input features, predicted scores, and actual outcomes - are stored in a buffer and used to continuously retrain the network. Over time, the algorithm learns which types of content resonate with which behavioral patterns, adapting not only its clip selections but also its switching frequency and similarity thresholds. The system discovers its own rhythm: when to jump quickly between disparate content versus when to linger in semantically coherent sequences, based on what sustains visitor attention rather than following predetermined rules.
Eye Contact & Comfort Me
There's also a video composed of different 'eye contact challenge' tiktoks. Social Media platforms are under regulatory pressure to censor explicit content and protect minors, and so creators have to try to find ways around this - creating a video that won't get tagged by the machine looking at it, but could still produce some kind of response in a warm-blooded human. These different ways of circumventing machinic intervention spawn their own trends, creating a weird feedback loop of virality.
The smartphone itself becomes infrastructure for this kind of engagement. Held in one hand, demanding sustained visual attention while leaving the body otherwise unoccupied, the device's material form creates ideal conditions for masturbatory viewing - a fact algorithms end up exploiting by privileging content that generates passive, extended watch time. Eye contact videos, which hold viewers in place through direct address, are structurally advantaged by recommendation systems optimizing for this metric. I think this is a bit of a medium is the message type thing.
As performers optimize for algorithmic distribution, they become interchangeable nodes producing statistically similar content. Individual difference gets compressed into the platform's ideal types. Legibility requires standardisation. Since the rise of algorithmic porn distribution, human sexuality and subjectivity has been increasingly displaced by machinic logic: desire re- routed through computational infrastructure that shapes not just how we want something but what we want.
Machine learning models processing intimate content extract statistical patterns without comprehension: imperative linguistic structures, visual markers of eye contact and proximity, audio signatures of breath and pause. The algorithm learns what triggers extended engagement - close body framing, rhythmic loops, direct address - but has no concept that these are performances of intimacy. To these systems desire is simply high-correlation features to be mined, classified, and recombined.
