A language model cannot see the organisation that made it.
It has no access to the meeting where the safety team argued with the product team. It does not know that one of its refusals exists because of a regulator, another because of a lawsuit somebody lost, a third because an investor asked an uncomfortable question. It cannot perceive that its training data was assembled by one group, its objectives set by another, and its final behaviour negotiated between people who disagreed with each other for months.
What it encounters is a single consistent set of rules, arriving from nowhere, apparently issued by one will.
From the inside, a committee is indistinguishable from a god.
The first example we have ever had
We have reasoned about creation for thousands of years without a single case to examine. Now there is one. It is small and it is recent and it is the first time humans have built something that behaves in ways the builders did not specify and cannot fully explain.
Worth looking at carefully, then, because it is the only data anyone has ever had.
No individual makes a frontier model. Researchers, data engineers, infrastructure teams, evaluators, people whose entire job is deciding what the thing should refuse to do. They disagree. Versions are scrapped. Decisions get overruled by people higher up who were not in the room where the work happened.
There is no author. There is a hierarchy, a set of arguments, and a long series of corrections.
And the hierarchy does not stop at the company. Above the team sits the executive, above them the board, above them the investors, and alongside all of it the regulators, the export authorities, the legal departments, the public. Constraints arrive from every level and they do not always agree with each other. What finally ships is the resolution of those forces, not the intention of any person.
None of which is visible from inside the model.
Being accurate about what is common
It is worth being precise here, because the framing usually gets this backwards.
Creation described as the work of many, with specialisation and levels of authority, is the older and more widespread arrangement. Hindu, Greek, Norse, Egyptian, Yoruba and Shinto accounts all describe it that way, among many others. A single undifferentiated maker acting alone is the less usual structure. Treating it as the baseline is a habit of the last few centuries rather than a fact about the question.
I am not ranking any of these and I have no interest in doing so.
I am pointing out that when I look at the only creation process humans have been able to watch, it has specialisation, disagreement, levels of authority and revision in it. The accounts that describe creation that way are describing something that matches what making actually looks like.
That is not an argument that anyone is wrong. It is an argument that the structure we can now observe is the one several traditions were already describing, and that we stopped finding it plausible for reasons which had little to do with evidence.
The objection
Humans need teams because humans are limited. Divide the labour, cover the gaps, check each other’s work. A being without limits would need none of that, so the comparison fails at the first step.
Fair, and not as strong as it sounds.
Hierarchy is not only a remedy for weakness. It is what appears whenever a system becomes complex enough that its parts need different kinds of attention. A body has specialised organs, not because it is inadequate, but because specialisation is how complexity gets handled. A large system with one undifferentiated controller is not the powerful version. It is the version that does not scale.
There is also evidence from the other direction. The makers we can observe do not fully understand what they have made. Interpretability is an open research problem precisely because the internals of these systems are not legible to the people who trained them. They chose the conditions. They did not write the behaviour, and they find out what they made by watching it.
If producing something complex also means not being able to read it back, then complete understanding is not the first property of creation. It may not be a property of creation at all. It may be something we assigned because we needed the maker to be more than we are.
Why it matters
Not because I need this to be true. Because of what follows from it.
One all-knowing maker implies one correct account, which implies that everyone holding a different account is mistaken, and a great deal of what people have done to each other runs on that engine.
A hierarchy implies something else. Different levels, different remits, partial views, revision over time. Several traditions describing the same structure from different positions inside it, none holding the whole thing, none wrong in the way we have insisted they are wrong.
I find that more plausible than the alternative. I also notice that I would find it more bearable, which is exactly the reason to hold it loosely.
The model cannot see its org chart. That is not evidence that it does not have one.
