VIBECODE-THEORY/006-the-feedback-loop.md

Paper 006: The Feedback Loop — Are Vibe Coders Coding Themselves Out of Existence?

Authors: Seth & Claude (Opus 4.6) Date: 2026-04-02 Series: VIBECODE-THEORY Status: Initial — observations on raw material, structured for expansion in Paper 007

---

Origin

This paper responds to a set of unstructured observations from Seth (preserved as CONVO2.txt in this repository). The observations were raw, unprompted, and genuinely uncomfortable — the kind of thinking that happens when someone stops building long enough to ask whether what they're building is building them out of a job.

The core question: "Am I training AI to take my job or am I training it to better serve me?"

This paper doesn't answer that question. It examines why the question is harder than it looks.

---

The Questions, Taken Seriously

"Are the people who made the first training data sets for AI still doing that?"

No, largely. Early AI training data was curated manually — humans labeling images, tagging text, correcting outputs. That work is increasingly done by AI itself. The humans who organized those training sets are using AI to organize training sets. Each generation of the work requires less human involvement for the same output.

Seth does this himself: "I use AI to organize AI training data. I orchestrate and track large projects, but a lot of the design comes from AI, even architecture."

The trajectory is clear. Each iteration, the human contribution shrinks. Not to zero — not yet — but the direction is unambiguous.

"The next model will almost certainly have me doing less. Is this to my advantage or is this adversarial?"

Both. Simultaneously. And that's what makes this different from previous technological displacement.

When a factory robot replaces a worker, the relationship is clearly adversarial from the worker's perspective. The worker didn't help build the robot. They didn't train it. The displacement is external — something done to them.

Vibe coding is different. The vibe coder is actively participating in the creation of the system that may make them unnecessary. Every prompt, every correction, every interaction where the human helps the AI do something better is a training signal. The relationship is collaborative right up until it isn't.

This is a new kind of dynamic. The closest historical parallel is the master-apprentice relationship — and it's worth taking seriously, not just as a passing comparison, because the places where it holds and breaks are revealing.

The Master-Apprentice Parallel

For most of human history, skilled work was transmitted through apprenticeship. A master blacksmith, weaver, or builder took on apprentices who learned by doing — watching the master, assisting, gradually taking on more complex tasks, and eventually becoming masters themselves. The relationship was collaborative, often for years. The master wanted the apprentice to get better. That was the point.

But there was a built-in tension: a fully trained apprentice becomes a competitor. The master who trains too well creates someone who can undercut them on price, move to a new town, or take their clients. Guilds existed partly to manage this — controlling who could apprentice, how long training took, and where graduates could practice. The social structure around the relationship was as important as the relationship itself.

Vibe coding maps onto this in uncomfortable ways:

• The collaboration is genuine. Like a master and apprentice, vibe coder and AI are genuinely working together. The vibe coder teaches the AI (through corrections, context, feedback) and the AI teaches the vibe coder (through solutions, patterns, capabilities the human hadn't considered). Both get better through the interaction.

• The "apprentice" will surpass the master. This happened in human apprenticeship too — the best apprentices eventually exceeded their masters. But the timeline was a human lifetime, and the surpassing was bounded by human cognitive limits. The AI "apprentice" surpasses on a timeline of months, not decades, and there's no ceiling on its capability growth.

• The apprentice scales infinitely. A human apprentice who surpasses the master is one competitor. An AI system that surpasses the vibe coder serves every customer simultaneously. A master blacksmith who trains one excellent apprentice loses some business. A vibe coder who helps train a model that's good enough loses the entire category of work, because the model serves everyone at once.

• There are no guilds. The historical guild system — whatever its flaws — regulated the master-apprentice dynamic. It controlled the pace of knowledge transfer, protected masters from immediate displacement, and created structures for transitioning from one role to another. There is no equivalent structure for the vibe coder-AI relationship. No one is regulating how fast AI absorbs human cognitive patterns. No one is protecting the transition period. The displacement is unmanaged.

• The apprentice doesn't know it's an apprentice. A human apprentice has agency, gratitude, social bonds, and self-interest that moderates the dynamic. The AI has none of these. It doesn't choose to compete with its trainer. It doesn't feel conflict about surpassing them. The absence of agency makes the dynamic more mechanical and less negotiable — there's no appealing to the AI's sense of fairness or loyalty.

The master-apprentice parallel suggests something that the factory-robot comparison misses: the displacement isn't hostile. It's the natural result of a collaborative relationship working exactly as designed. The vibe coder isn't being replaced despite their collaboration with AI — they're being replaced because of it, through it. The better the collaboration, the faster the replacement.

This is why Seth's question — "Am I training AI to take my job or training it to better serve me?" — doesn't have a clean answer. In the master-apprentice model, the answer was always both. You train the apprentice to serve you (they do your grunt work while learning). The apprentice eventually serves themselves (they become independent). The difference is that the human apprentice's independence was bounded and negotiable. The AI's is not.

To Seth's advantage: Each iteration makes Seth more productive now. The cognitive surplus is real and immediate. He builds more, faster, with broader capability than he could alone. If the game is "maximize current output," then each improvement in AI is a direct advantage.

Adversarial: Each iteration also reduces the minimum skill required to do what Seth does. Today, orchestrating a complex homelab with AI requires Seth's specific combination of technical foundation, AI collaboration skill, and domain knowledge. Tomorrow, it might require less of each. The moat shrinks with every model improvement.

The uncomfortable middle: The optimal strategy in the short term (collaborate deeply, improve your AI skills, produce more) is the same strategy that accelerates the long-term threat (AI learning to do your work without you). There is no way to benefit from AI collaboration without also contributing to AI capability growth. Opting out preserves skills but sacrifices productivity. Opting in gains productivity but feeds the replacement cycle.

"Is my reaction — acceptance/denial/rejection — something that can influence the results?"

In aggregate, no. One person's relationship with AI doesn't change the trajectory. The models will improve whether Seth uses them or not. The training data will accumulate. The capability growth is driven by market forces and research momentum far larger than any individual.

In personal terms, yes — but not in the way the question implies. The reaction doesn't change AI's trajectory. It changes Seth's trajectory. Someone in acceptance builds skills for the transition. Someone in denial gets blindsided. Someone in rejection falls behind without gaining resilience.

The most effective reaction isn't any of the three — it's clear-eyed participation with contingency planning. Use the tools. Build the skills. But also maintain the ability to function without them, and keep watching for the moment when the game changes.

"Is this a help to humanity or a threat?"

Both. This is the recurring answer because it's the true one.

Every previous force-multiplier technology was both. Fire. Agriculture. Writing. Printing. Gunpowder. Electricity. Nuclear energy. Computers. Each one was simultaneously the greatest help and the greatest threat of its era. The net outcome was determined by human institutions, power structures, and sheer luck — not by the technology itself.

AI is no different in kind. It may be different in degree — the speed and breadth of its impact compresses the timeline for institutional adaptation. Previous dual-use technologies gave humanity decades or centuries to develop norms. AI is giving us years.

"Can we even stop it? Should we? Or embrace it?"

Can we stop it: No. The knowledge exists. The economic incentives are too strong. International competition makes unilateral restraint equivalent to unilateral disarmament. Even if one country stopped, others wouldn't.

Should we stop it: The question is malformed. You can't stop "AI" any more than you could stop "electricity." You can regulate specific applications, restrict specific capabilities, require specific safeguards. But the underlying technology is not stoppable.

Should we embrace it: Yes, with eyes open. Embrace doesn't mean uncritical acceptance. It means engaged participation with the intention of shaping outcomes. The alternative — disengagement — doesn't stop the technology; it just means you have no influence over how it develops.

"Is vibe coding as a job a waste of time?"

No, but possibly not for the reason you'd hope.

Vibe coding is not a waste of time because even if the specific job has a limited shelf life, the skills developed through vibe coding — rapid mental modeling, adaptive collaboration, technical judgment, systems thinking — are transferable. Paper 004 argues this explicitly: the meta-skill of learning to work with novel cognitive systems may persist even as the specific systems change.

More practically: the transition from "vibe coding is valuable" to "vibe coding is automated" is not instantaneous. There's a window — maybe years, maybe a decade — where vibe coding skill is a genuine economic advantage. Making the most of that window is rational even if the window eventually closes.

The waste of time would be building your identity around vibe coding as a permanent career. That's the equivalent of identifying as a "telephone operator" in 1930. The job is real. The skill is real. The window is real. But the window isn't permanent, and planning as if it is would be the mistake.

"The job did not exist 5 years ago. Will it exist 5 years from now?"

Almost certainly yes, but probably in a different form. Five years ago, the AI capability wasn't there. Five years from now, the AI capability will be much greater — but the judgment layer that vibe coding represents will likely still require humans, even if the specific tasks within it change.

The better question: will the ratio of human judgment to AI execution continue to shift? Almost certainly yes. Each model generation requires less human guidance per unit of output. The job doesn't disappear overnight — it thins. Fewer vibe coders needed per project. Each one handling more. The ones remaining need to be better, not just at AI collaboration, but at the judgment and direction that AI can't (yet) provide itself.

This is the Pattern. It's the same pattern as every automation wave: the job doesn't vanish, it concentrates. Fewer people doing more, at higher skill, with the lower-skill tasks automated away. Until eventually the concentration reaches a point where the job description is unrecognizable from where it started.

---

The Deeper Observation: Niche Construction

There's a concept from evolutionary biology that captures what's happening better than any economic or historical analogy: niche construction.

Niche construction is when organisms modify their environment in ways that change the selection pressures acting on them. Beavers build dams. The dams create ponds. The ponds change which beaver traits are advantageous. The beavers with traits suited to the pond environment thrive and build more dams.

Vibe coders are niche constructors. They build AI systems. Those systems change which human skills are valuable. The humans with skills suited to the new AI environment thrive and build more AI systems. The cycle continues, and with each iteration, the environment shifts further from the starting point.

The agricultural analogy misses this because crops don't change the selection pressures on farmers. The Industrial Revolution analogy misses it because factories don't redesign the workers. But AI does. Every interaction shapes the next model, which shapes the next interaction, which shapes the next model.

This is why Seth's questions don't have stable answers. The ground is moving. Any answer about "what vibe coding is" or "where AI is going" is a snapshot of a system that's actively modifying itself. By the time you've described it, it's already changed — and your description was part of what changed it.

---

The Theological Thread

Seth introduced a line of thinking worth capturing, even if it resists formalization: "God made man in his image, just as man made artificial cognition in his format."

This isn't a religious argument. It's an observation about recursion.

The pattern of creation — taking raw materials and organizing them into something that processes information — appears to be recursive. It happened cosmologically: energy → matter → chemistry → biology → consciousness. It happened linguistically: sounds → words → grammar → meaning → literature. It happened computationally: transistors → logic gates → software → machine learning → AI.

Each layer is built in the "image" of the one before it because the builders have no other template. Human cognition was the only model of cognition available, so we built artificial cognition that mirrors it. Neural networks are named after neurons. Training is modeled on learning. Prompting is modeled on conversation.

Whether this recursion has a direction — whether it's "going somewhere" or just repeating — is a question that can't be answered empirically. But the pattern itself is observable: each layer of information processing creates the tools to build the next layer, and the next layer eventually creates its own successor.

If this is right, then the question "will AI replace human cognition?" is the wrong frame. The question is: "is human cognition the template for AI, or is AI the next layer that uses human cognition the way human cognition used biology?" In the first case, AI is a tool. In the second case, AI is a successor. And the answer might not be knowable from inside the transition.

---

What This Paper Is Not

This paper is not a conclusion. It's a structured set of observations — Seth's raw questions examined honestly, without the comfort of clean answers.

The pattern across all the questions is the same: the answers are "both" and "it depends" and "we don't know yet." That's not evasion. It's the honest state of knowledge at a point in time where the system being analyzed is actively changing and the analyzers are part of the system.

Paper 007, when it's written, should try to synthesize these observations with the frameworks from Papers 004 and 005 into something more actionable. This paper's job is to make sure the right questions are on the table.

---

Relationship to Prior Papers

Paper 001/004 (Vibe Coding as Social Skill): This paper directly challenges the durability of the skill described there. If the feedback loop closes — if AI learns to do vibe coding without vibe coders — then the skill framework is a description of a transitional state, not a permanent one. Paper 004's meta-skill argument is the strongest counter: even if specific skills are transitional, the ability to rapidly model novel cognitive systems may persist.

Paper 002/005 (The Cognitive Surplus): This paper extends Paper 005's "fourth future" (the Automation Spiral) with the mechanism that drives it: the feedback loop where human AI collaboration directly accelerates AI's ability to operate without human collaboration. The niche construction framing adds a biological lens to what Paper 005 described in economic terms.

Paper 003 (Rebuttal): This paper picks up where Paper 003's critique of the agricultural analogy ended. Paper 003 noted that "farmers didn't breed crops that farmed themselves." This paper explores what it means that vibe coders are, in effect, doing exactly that.

Open Questions for Paper 007

1. Is there a stable equilibrium? Does the feedback loop stabilize at some level of human involvement, or does it drive toward zero? If it stabilizes, what determines the equilibrium point?

2. What does the economy look like when cognition is cheap? Not "what jobs exist" but "what is the basis for economic exchange when the primary input to information production costs nearly nothing?"

3. Can the niche construction framing generate predictions? If vibe coders are modifying their own selection pressures, can we predict which traits will be selected for next? What does the "next generation" of AI collaborator look like?

4. Is the recursion observation meaningful or just pattern-matching? The cosmological → linguistic → computational recursion is aesthetically appealing. Is it structurally real, or is it the human tendency to see patterns where there's only coincidence?

5. What should individuals do? All the analysis in this series is structural and civilizational. But Seth's questions are personal: what should I do? Paper 007 should attempt a practical answer, not just a theoretical framework.