Where it is heading
Where it is heading
The half-life of predictions
In January 2023, Matt Welsh wrote "The End of Programming" in Communications of the ACM. His argument: there will no longer be any need to write computer programs, most software will be replaced by AI systems that are trained instead of programmed, and language models are virtual machines programmed in natural language [42]. Two years later Karpathy gave "vibe coding" its name [1], and a year after that the same person looked back: one year on, programming through LLM agents is becoming the default way professionals work, "except with more oversight and scrutiny" [43]. He calls this "agentic engineering" and separates the two: vibe coding raises the floor, and agentic engineering raises the ceiling.
Lay the three moments over each other and two things show. First, the names of the tools change every year but the direction has not: people move away from code and closer to judgment. Second, even the most optimistic person brought "oversight" back in. That is why the "how to look" section of How to do it well is structural and will outlast any fashion.
Force 1: the code goes away, the judgment remains
Sarkar and Drosos said expertise shifts from writing toward evaluating, steering and refining [2], and Ge et al. called the developer an intent expresser, a context curator and a quality judge [7]. Krings et al. go a step further and propose rethinking programming as an expressive and speculative practice, and code as the residue of creation [9].
On the art side this is not unfamiliar. In the shader evolution system the user is already a "creative curator" [33], and in the 737-person experiment only two things were left to the person: direction (the eye for what to change) and choice (keep this, or revert) [17]. It is the same position as when this program's organizer describes their own work as designing the rules and environments from which a narrative emerges. In a structure where the artist writes the rules, the system executes them and the audience experiences them, the artist's place was already outside the code. Vibe coding has simply opened that place to everyone.
Open question: where between the person who writes the rules and the system that executes them is "the work"? Is the prompt the work, is the result the work, or is the record of choices between them the work?
Force 2: the pull toward the average
Hito Steyerl called generated images "mean images". In her account they are statistical renderings of no object that actually exists, pointing to probability rather than to facts or truth, replacing the shock of photographic illumination with bell curves, loss functions and the drag of the long tail, a hallucinated mediocrity converging around the mean and the median [47]. The same force acts on code: web vibe coding creates a feedback loop in which a model trained on the statistical average of the internet becomes a universal homogenizer [28], and in a controlled experiment the LLM made solutions resemble each other more the more freedom a task allowed [25].
The end of that feedback loop has a name. Train a model again on data made by models and you get "irreversible defects in the resulting models, in which tails of the original content distribution disappear": the "model collapse" published in Nature [45]. A 2026 follow-up applied this to code. In the cycle where AI code enters repositories and becomes training data again, collapse is fastest without review, human-side filters such as compilation and quality rules only slow it down without stopping it, and having the model review itself was no substitute [46]. The two papers reach the same conclusion: data from genuine human interaction becomes more and more valuable.
For art this is both a threat and a definition. What the average swallows is the tail (the rare, the local, the not yet named), and that tail is where art has been. The Busan dialect, the sigimsae ornamentation of Korean traditional music, the oral testimony of survivors of the Brothers' Home, fourteen years of a collection: none of these is in any training data as an average. Item 23 of How to do it well ("put in your own material") is a declaration of position as much as a tip.
Open question: is putting in your own material enough, or, since the tool itself leans toward the average, does the labor of pushing the result away from the average become part of the work?
Force 3: the shift of authorship and credit
In January 2025 the U.S. Copyright Office set out the requirements for copyright protection of generative AI output: it is protected only where a human has determined sufficient expressive elements, purely AI-generated material is not protected, and using AI as an aid (as opposed to a stand-in) does not affect protection [49]. (This is the U.S. standard. Korea's position needs to be checked separately.) A perspective paper in Science warns of a shift in another direction: the more audiences attribute intent and agency to AI, the more credit may flow away from human artists and toward those who built the technology [44]. Practitioners have already started responding, suggesting that people treat the AI like a co-author, record their prompts and note where the AI stepped in [19]. Electronic literature goes further and embraces authorship distributed across human creativity, algorithmic processes and existing digital material as a poetics [40].
In this program, authorship comes up at the December critique. The question to ask is "what did you choose, what did you reject, and why", more than "did you use AI". If you have an answer to that, you are the artist by the Copyright Office's standard and by the audience's "guessing the intent" standard alike. Item 24 of How to do it well (keep a record of prompts) hangs on this.
Open question: is exhibiting the record of prompts, rejections and choices alongside the work transparency, a new form, or the mistake of replacing the work with an explanation?
Force 4: barriers fall, and so do wages
The economics section of the Science paper is sober. Generative tools lower the barrier to entry while making creative tasks more efficient, so the number of people working as artists may rise while average wages fall. With the same budget buying more output, fewer people make the same amount, and commissioned trades such as illustration and stock photography are pushed out first. And history repeats: when the Industrial Revolution mass-produced ceramics and textiles, handmade things became luxury goods carrying greater artistic intent [44]. In a survey of vibe coding practitioners, non-developers' first motivation was being able to make things they could not have made otherwise [6], while another study worries about dependence on API credits and powerful machines, and about exclusion and technological elitism [9]. In Jin's words, the symbolic capacities of programming languages are being privatized [10].
Applied to our situation: the Foundation covers subscription fees for four months. What will your work need to keep running after that? This is the first question after December.
Open question: in a world where the barrier is gone, what distinguishes an "artist"? Access to tools, judgment, time, or the will to move away from the average?
Force 5: control moves up, the work's lifespan moves down
Beneath the interface flattened into a single prompt box, a thicker layer has formed: remote GPUs, model providers, function calling, MCP and agent protocols. Control does not disappear; as it moves into conversation, it concentrates in model and protocol providers [10]. Price changes and service shutdowns hit individual artists directly [3]. Meanwhile, a study of 1.61 million creative code repositories reports that 5.5% have already disappeared, that generative art depends on a fragile supply chain of assets, device protocols and APIs, and that net art from the Flash era no longer runs [36]. Generated code that nobody understands is even harder to preserve.
That is why this wiki insists on "one browser page" from the first day. A single HTML page sits at the very bottom of those layers: no GPU, no subscription, no protocol, and it will still open in 2036. The higher you go (APIs, MCP, remote models), the more powerful the work becomes and the shorter its life. How high to go is a choice you make for each work, and before the December installation you will be asked: "if you unplug it and plug it back in, does it run again? What happens if the internet drops?"
Open question: what lifespan do you set for the work? Is three days of exhibition enough, or do you hope for ten years? The answer decides how many layers up you go.
Two kinds of workmanship
A design education study borrows David Pye's distinction [26]: the workmanship of risk, where the result is at risk while it is being made, and the workmanship of certainty, where the result is guaranteed. This wiki's hypothesis is that the more tools move toward guaranteeing results (certainty), the more art moves to the other side. Compton's casual creator that trades control for speed, power and surprise [38] is a pleasure on the side of certainty, and the bricoleur stepping back between brushstrokes [37] is a pleasure on the side of risk. Both are legitimate, and both are inside this program. The direction of this curriculum is to start the first-day prototype on the side of certainty (one experience of something working) and to hope you are standing on the side of risk by December.
Manovich and Arielli were already asking who counts as an artist in the age of AI in 2022, moved on to AI as a shift from tool to author in 2024, and closed with the question of aesthetic alignment: by humans, for humans? [48]. That last question is the right one. Mean images make things that can be explained. Direct experience instead of explanation, which audiences discover for themselves through play and embodiment, is something statistical rendering cannot yet make, and it is the methodology written in the first line of this program.