Every actuarial team I have worked in had a shared drive with a folder called something like Reports. Inside: the 2023 study report in three near-identical "final" versions, a data-definition sheet with two competing versions, a "how we treat split claims" note that half the team knows and the other half rewrote from scratch last quarter. None of it was wrong when written. Much of it has quietly drifted since. Nobody deleted anything, because nobody was sure what was still true.
That is the actual failure mode of team knowledge. It is rarely lost. It decays — the world moves, definitions drift, two documents start to disagree, and the disagreement sits unnoticed until it costs someone a week. Wikis, note-taking apps, SharePoint: every generation of tooling made it easier to store knowledge, and none of them made it cheaper to keep it true. The maintenance always fell to a person, and it was never anyone's job.
In 2026 Andrej Karpathy sketched a pattern that goes after exactly that maintenance cost, and gave it a plain name: the LLM wiki. I have been trying a version of it as a personal project — pointed at the build record of my own AI-agent experiment, the one behind the AAC talk. This note is the concept and the actuarial reading of it; the results will come later, once there is enough history to be honest about.
What the pattern actually is
Strip away the tooling and it is three layers and three loops. The layers: raw sources (memos, papers, meeting notes — immutable, the LLM reads but never edits), the wiki (interlinked markdown pages for entities and concepts — owned and written entirely by the LLM), and the schema (one configuration file that says how pages are structured, what conventions hold, and how the loops run). If you read Note 01, you will recognise the schema for what it is: a harness.
The loops: ingest — a new source arrives, the LLM reads it, updates every page it touches, flags where it contradicts something already on file, and appends to a log. Query — you ask a question, it answers from the wiki with citations, and a good answer gets filed back as a page. Lint — periodically it walks the whole thing looking for contradictions, orphan pages, missing cross-references, and claims that a newer source has superseded.
The contrast Karpathy draws is with retrieval-augmented generation. RAG re-derives an answer from raw chunks every time you ask; nothing accumulates. The wiki compiles knowledge once and then keeps it current — in his words, "a persistent, compounding artifact." And the division of labour is stated plainly: the human's job is to curate sources, direct the analysis, and ask good questions; the LLM's job is everything else — the summarising, cross-referencing, filing, and bookkeeping that no one wanted to do.
Read it as an actuary and it looks familiar
Now put on the other hat. Think about what an experience-study team actually accumulates over the years: which cedant's data arrives in what shape, which checks caught what, how a judgment call was made and why, which treatment worked and which quietly didn't. That know-how is the team's real asset — and almost all of it lives in close-out memos, old workpapers, and people's heads. In my AAC talk I called this bottleneck by name: one-off fixes, no reuse — the records pile up, but most stay case-specific. The LLM wiki is the missing second half of that sentence. It is where one-off fixes go to become reusable: each study's lessons, filed as general pages, cross-referenced, and — this is the part no shared drive ever did — kept true after the next study contradicts them.
So the actuarial reading of the LLM wiki is not "a chatbot over our documents." It is: turn every study close-out into an ingest, and every pre-study "how did we handle this last time?" into a query against pages that are actually current. Every contradiction flag is a lesson from one study colliding with a lesson from another — exactly the collision you want surfaced before the next study, not during it. And the human role stays where the profession keeps it: curate what goes in, decide what a contradiction means, sign off. The agent proposes an edit; a person approves it. If you read the AAC experiment, you have already seen this working at miniature scale — the harness playbooks grew from zero to fifty rules precisely because every retrospective was ingested, generalized, and gated. That was an LLM wiki with one topic. This note is the same loop, pointed at everything a study teaches.
Why "writes itself" is the wrong thing to be impressed by
The demo everyone shares is the writing: drop in a memo, watch six pages update. It is impressive and it is not the point. Writing pages was never the bottleneck — junior staff have always been able to write pages. The bottleneck was keeping the pages true after the world moved, and that is a decay process, which means the honest way to evaluate any knowledge system is not "how much did it write" but "how much of what it holds can I still trust, and how many contradictions is it hiding right now."
Below is a toy model of that idea — no chart to decode, just the pages themselves. The same hundred pages of team knowledge, three times over. Every month the world changes a little, so pages quietly fade; stale pages that overlap start to disagree — the red-ringed ones are contradictions nobody has noticed. The only difference between the three panels is who maintains the wiki, and how often. Watch three years pass.