The conversation around LLM wikis needs to shift.
Andrej Karpathy’s original LLM Wiki gist describes something interesting: a persistent, compounding artifact maintained by an agent through three operations — ingest, query, lint. Raw sources flow in. The wiki absorbs them, cross-links them, and re-checks itself for contradictions and stale claims. There is a schema file that disciplines the agent. There is a log.md that chronicles every ingest. The whole thing is, by design, a runtime.
That is not what most of the derivatives became.
Many implementations inspired by Karpathy strip out the runtime mechanics and keep only the markdown surface. The result is a familiar pattern:
- They repackage the same public knowledge.
- They go stale within weeks.
- They are disconnected from real production workflows.
- They create no compounding advantage.
Without ingest, query, and lint loops, no one is making decisions based on your LLM wiki. No one is integrating it into their tech stack. It is not a system; it is a static artifact.
The uncomfortable truth is that a wiki without a runtime is often what you build when you lack a runtime.
The true leverage today is not in documenting LLMs but in developing:
- Systems that execute (agents, pipelines, automation)
- Memory that enhances outcomes over time
- Tight loops connecting data, models, and actions
Karpathy’s original gets this. Most of the copies miss it.
If it doesn’t run, adapt, or learn, it’s not infrastructure; it’s just content.
What we need is not more explanations of LLMs, but systems that truly make them useful.
References:
- Andrej Karpathy — llm-wiki.md