In a companion essay I argued that there is a class of AI infrastructure that participates in execution without generating, and that current observability vocabularies do not see it. This essay is about why. The argument here is not technical. It is about how standards are shaped by the institutional position of the people doing the shaping — and why that pattern, once seen clearly, is more important than any single attribute or namespace decision.

The frame I want to borrow for this is second-order cybernetics. It will do most of the structural work in what follows, so it is worth spending a moment on what it means and why it fits.

A short detour through second-order cybernetics

First-order cybernetics, as Norbert Wiener and the early cyberneticians developed it, studied systems that regulate themselves through feedback: a thermostat, a homeostatic process, a control loop. The observer of such a system was assumed to stand outside it. The system was the thing being studied; the observer was a neutral instrument.

Second-order cybernetics, articulated most clearly by Heinz von Foerster, took that assumption apart. The observer, von Foerster argued, is not outside the system. The observer is part of the system. What gets observed depends on the position, the instruments, and the categories the observer brings with them. The act of describing a system changes what counts as a feature of the system, which changes what gets noticed in future observations, which changes the descriptions, and so on. There is no neutral observer. There are only observers shaped by their position relative to what they are looking at.

Standards work, I want to argue, is second-order cybernetic in exactly this sense. A technical standard is not a neutral description of a class of system. It is a description produced by observers who are themselves embedded in — and shaped by — the institutions, incentives, and deadlines around them. The vocabulary that gets standardized is not the vocabulary that best describes the systems in question. It is the vocabulary the observers were positioned to produce.

This is a structural observation, and recognizing it is the first step to working with it rather than being surprised by it.

Who is in the room shapes what the standard sees

Most contributors to large open technical standards are paid by someone to be there. This is not a secret and it is not a scandal. Open-source standards work is real engineering effort, and at the scale the modern internet operates, that effort needs to be funded. Standards bodies are populated overwhelmingly by employees of companies whose products will eventually emit or consume telemetry shaped by the standard.

This produces a predictable pattern. The categories of system that receive careful, well-developed vocabulary are the categories that someone in the room is being paid to ship. A model invocation gets first-class observability because the companies building model-serving products want their products to be observable. A tool call gets first-class observability because agent framework companies want tool use to be observable. A retrieval step gets first-class observability because vector database vendors want retrieval to be observable.

None of these are bad outcomes. The vocabulary is genuinely useful, and the people producing it are mostly doing good engineering work under real constraints. But the cumulative effect of many such observers, each shaped by their own commercial position, is that the standard ends up describing the union of what the participating companies build — and not much more.

What sits outside that union? Roughly:

These categories are not invisible. People know they exist. They just do not have anyone in the room whose paid task is to give them vocabulary.

Content-blind AI infrastructure, the category I discussed in the companion essay, sits squarely in that set. It exists. It is used in production. It is implemented in several well-known open-source projects. And the standardization of vocabulary for it lags behind the standardization of vocabulary for model invocation, not because it is less important, but because nobody is being paid to push it through.

Deadlines as a hidden axis of standards work

There is a second layer to this pattern that is less often discussed: the observers are not only positioned by their employers, they are positioned by their deadlines.

A contributor who is being paid to land a particular proposal in a standard is doing so on a quarter-by-quarter basis, with internal commitments about when the work will be done. They have a roadmap. They have stakeholders. They are accountable for landing the specific set of attributes they came in with.

This produces a third-order effect. Even when a paid contributor privately agrees that an adjacent category of system deserves vocabulary too, they cannot easily expand their scope to cover it without missing their deadline. The structurally rational move is to treat adjacent work as orthogonal, scope it out of the current proposal, and continue toward the target the contributor was originally tasked with. This is not territorial behavior, exactly. It is what landing a proposal on a deadline looks like.

The cybernetic shape of this is worth pausing on. The contributor's behavior is being regulated by an external feedback loop — their employer's roadmap — that has nothing to do with the technical merits of the standard. From inside that loop, the right move is to keep scope tight and ship. From outside the loop, the same move looks like gatekeeping. Both readings can be true at once, because the contributor is responding correctly to the system they are embedded in, even though that system is not the standards body itself.

Second-order cybernetics gives us a precise way to say this: the observer's feedback loop and the system being observed are not the same loop. Confusing them — on either side — produces avoidable misunderstanding.

Where the unpaid contributor sits

I want to be honest about my own position in this, because the whole point of a second-order frame is that you cannot pretend to be outside what you are describing.

I am not an employee of a company shipping AI infrastructure. I am currently unemployed and doing this work in my own time, without an employer's roadmap shaping what I look at. That is relevant to this essay in two ways.

First, it is part of why I noticed the gap. With no commercial product to position, I had no reason to scope my attention to the categories that would help that product. I was free to instrument a small unusual system and follow whatever frictions appeared, without needing to defend the choice to anyone.

Second, it is part of why the gap is hard to close. The contributors best positioned to notice underrepresented categories are often the contributors least positioned to spend months pushing them through a standards body. The structural asymmetry between “I have time to notice” and “I have institutional standing to push” is part of why the landscape of standardized vocabulary looks the way it does.

The work I am doing is work I want to be doing, and the freedom it gives me to look at unusual systems without a roadmap is real value, of a kind that is hard to find inside a paid role. But it is also true that the people who can most easily see the gaps are the people with the least power to fill them, and the people with the most power to fill them are the people most constrained about which gaps they are allowed to look at. That asymmetry is a structural feature of how standards get made, and it deserves to be named.

What this is not

It would be easy to read this essay as cynical, and I want to push back against that reading explicitly.

This is not an argument that paid contributors are doing bad work. Most of them are doing excellent work, under real constraints, and producing vocabulary that genuinely helps the ecosystem. I have used their attributes. I will keep using them. The friction log that started this whole exploration was only possible because the vocabulary they built was already in place to push against.

This is not an argument that standards bodies are corrupt or captured. Open standards processes are, on balance, one of the better mechanisms our industry has for producing shared infrastructure. The fact that they reflect the commercial positions of their contributors is a property of any large coordination effort, not a unique failure of standards bodies.

This is not an argument that the right move is to walk away from standards work. The right move, I think, is to engage with a clear understanding of what is actually going on — to contribute where contribution is possible, to publish independent work where direct contribution is blocked, and to keep naming the structural patterns that shape which contributions land.

What this is, I hope, is an argument for a more accurate picture of how technical vocabulary gets made. The picture in which a standard is a neutral consensus among disinterested experts is wrong. The picture in which a standard is a corrupt vehicle for vendor capture is also wrong. The accurate picture is the cybernetic one: a vocabulary shaped by observers, whose positions shape what they can see, whose outputs in turn shape what the next round of observers will see, and so on. Recognizing that loop is the first step to contributing to it well, from whichever position you happen to occupy.

What this means for unpaid contributors

If you are doing AI infrastructure work without an employer backing your contributions, the situation is genuinely asymmetric, and there is no point pretending otherwise. But there are a few moves that I think work better than others.

Publish independently. A piece of writing hosted on your own site, with your own framing, lives differently than a comment in a long issue thread. It is citable. It is durable. It can be found by people who are not currently looking at the standards body at all. The work I described in the friction log will reach more thoughtful readers as an essay on my own site than it would as a series of replies inside someone else's proposal thread.

Treat the standards body as one venue among several. Standards bodies are where vocabulary gets ratified, but the thinking that produces vocabulary happens in many places: blog posts, talks, working group calls, side-channel conversations, academic-adjacent venues. An idea that cannot enter the standards body directly may still propagate through these other venues and arrive there later, carried by someone whose institutional position lets them push it.

Take the cybernetic frame seriously. When a paid contributor reroutes your contribution as “orthogonal,” do not read it as a verdict on the quality of your work. Read it as a signal about the feedback loop they are embedded in. The signal is real information. It tells you that the work you are doing is outside the scope of what they are paid to land in this cycle. That does not mean it is outside the scope of what is worth doing.

Resist the comfortable narrative. It is tempting, especially after a disappointing interaction, to settle into a story in which paid contributors are territorial pitchers and unpaid contributors are pure deep thinkers. That story is too clean. Most contributors, paid or not, are doing some mix of genuine thinking and structural positioning, and the mix shifts depending on the day, the topic, and how close they are to a deadline. Keeping a clear eye on this protects your own thinking from drifting into resentment, which is a worse engine for deep work than curiosity is.

Closing

The technical companion to this essay tries to make a category of AI infrastructure visible. This one tries to make a sociological pattern visible. Both are about legibility — about getting things into the field of view that have been sitting just outside it.

Second-order cybernetics says that observation is never neutral, and that the most useful thing an observer can do is to include themselves in the description. I have tried to do that here. The pattern I am describing — standards shaped by their contributors' commercial positions — applies to me too. My freedom to write this essay is a function of my being outside the institutional structure that would otherwise constrain it. That is not a virtue. It is a position. Different positions produce different things, and that is fine.

What I hope is that more of us — paid and unpaid, in the room and outside it — can name the loops we are embedded in clearly enough to do good work despite them. Not in a spirit of cynicism. In a spirit of careful, honest looking, which is what the cyberneticians wanted in the first place. 🌙

Status: Essay published · companion piece on the technical category here · further exploration ongoing.