If one AI plays every character in a story, it looks smart — give characters private information, and it quietly breaks.

This explores how arranging knowledge under one master scheme (one taxonomy, one 'view from everywhere,' one slot for each fact) narrows what counts as knowing. To be direct: the retrieved notes don't discuss encyclopedias, library classification, or the history of ordering knowledge. What they do contain are several AI findings that hit the same limit from another angle. Each shows what gets lost when knowledge is treated as one flat store of propositions seen from a single vantage point.

Start with the 'view from everywhere.' An encyclopedia speaks as if from no particular position, with every fact visible at once. LLM social simulations work the same way when one model plays every character, and that is exactly where they look most capable. Give each agent private information, though, and they fail systematically Why do LLMs fail when simulating agents with private information?. The omniscient setup hid the real work of knowing: figuring out what someone else knows, asking, and building shared ground. A single ordering that sees everything quietly drops perspective-bound knowing, the kind that only exists because people know different things.

Next, the assumption that knowing means holding a stored proposition. LLMs sometimes judge whether one statement follows from another by checking whether the conclusion appeared in their training data, not whether the premise actually supports it Do LLMs predict entailment based on what they memorized?. That is what an encyclopedia-shaped mind does: it recognizes familiar entries and misses the reasoning that connects them. A related failure is 'Potemkin understanding,' where a model explains a concept correctly, can't apply it, and can see that it failed Can LLMs understand concepts they cannot apply?. Knowing about something and knowing how to do it come apart, and a system organized around entries only ever holds the first kind.

Two more threads suggest other models of knowledge. One argues that LLMs show Saussure's idea that meaning comes from relations between words, not from words pointing at things Can language models learn meaning without engaging the world?. On that view, knowledge is a web where every term is defined by its neighbors, not a ladder of categories. The other argues LLMs may hold tacit knowledge, the kind you can act on but can't write down as an entry, although the causal evidence so far rests on a single fact-editing case Do language models possess tacit knowledge in Davies' sense?. Neither relational meaning nor tacit know-how fits a scheme that gives each fact one place in one hierarchy.

The last finding is the one you might not expect. A formal result shows that facts stored inside a model are capped by its size, and that cramming in more facts overwrites what was already there. Looking things up with external tools removes that cap Can models store unlimited facts without growing larger?. Read alongside the questions above, this suggests the opposite of one big ordering may be a good alternative: keep facts in many outside stores organized in different ways, and treat knowing as the skill of moving between them, not as owning the one correct map. If you want material that tackles encyclopedic ordering itself, such as Diderot or library classification systems, the collection doesn't seem to have it yet.