Andrej Karpathy published LLM Wiki. Here is why I think RAG still makes sense for some things.
A pattern that went viral in April 2026. My instinct is that RAG and LLM Wiki solve different problems. I shared that thought on X and the conversation was interesting.
On April 3, 2026, Andrej Karpathy published LLM Wiki. The core idea: instead of re-deriving answers from raw documents on every query, an LLM builds and maintains a persistent, structured knowledge base that compounds over time. The follow-up GitHub Gist went viral the next day.
The framing going around was that RAG is broken and this is the replacement.
I had just shipped RAG search on this site. So I had thoughts.
RAG and LLM Wiki are solving different problems.
My read: RAG is for "find the relevant piece and answer." LLM Wiki is for "synthesize everything I know over time." One is a retrieval pattern. The other is a knowledge accumulation pattern.
LLM Wiki makes a lot of sense if you are reading papers for months and want a compounding knowledge base. The AI reads each source, updates entity pages, flags contradictions, and builds cross-references. Nothing disappears into chat history. The wiki gets smarter with every source you add.
RAG makes sense when your corpus is already structured, and the job is finding the right chunk quickly. Which is exactly what ASK ✦ does on this site. You ask a question, it finds the relevant field notes, streams back a cited answer. No synthesis needed. No compounding knowledge base needed. Just retrieval.
The thing that helped most was starting from structured content.
One thing I noticed building the RAG search here: because the field notes live in a headless CMS, the chunking logic was clean from the start. The document already knows where sections begin. No scraping rendered HTML. No guessing at heading boundaries. The structure is already there.
That is not a RAG vs LLM Wiki argument. That is just a good data foundation, making everything downstream easier. Both patterns benefit from it.
The conversation on X was worth having.
I replied to the thread with my take, as an instinct rather than a conclusion. The replies suggested others were thinking the same way. Not RAG vs LLM Wiki. Both valid. Different jobs.
Karpathy himself frames it clearly in the Gist: most RAG systems re-derive knowledge from scratch on every query. LLM Wiki compiles it once and keeps it current. That is a real difference. The question is just which job you are actually trying to do.