Compounding Knowledge

The Pattern

Every interaction with the knowledge base — every document written, every question answered, every decision recorded — makes the system more valuable for future interactions. The knowledge base is not a static archive. It is a flywheel: usage generates artifacts, artifacts improve future usage, and the whole thing compounds over time. The key design choice is making sure outputs flow back into the system rather than evaporating into chat or email.

Where It Appears

• Karpathy: LLM Knowledge Bases — The core thesis. Raw data goes in, an LLM compiles it into structured wiki, queries generate output that gets filed back, the whole thing improves with use. “Every exploration compounds.” Karpathy identifies the break point: when outputs go to chat instead of back into the wiki, the loop stops compounding.
• Block: Hierarchy to Intelligence — Block’s “world model” is the compounding layer. Every transaction, every operational decision creates machine-readable artifacts that feed the intelligence layer. The more the company operates, the smarter the system gets. The article asks the open question: “What is Ray Data Co’s honest signal?”
• Content Intake SOP — The intake pipeline is the ingest side of the compounding loop. Media arrives through channels, gets processed into structured markdown with wikilinks and frontmatter, gets embedded into QMD, and the vault grows. Without this SOP, knowledge arrives but doesn’t compound.
• Infrastructure Decisions — The entire tooling stack (Obsidian + QMD + MCP) is chosen to support the compounding loop. QMD indexes the vault so past knowledge is retrievable. The vault structure ensures new docs are findable. The decision to go local-first with Obsidian over Notion was partly about keeping the loop tight — no API latency, no sync issues, everything on the Mac Mini.

Tensions

• Compounding vs. noise: Not everything should be filed back. Filing low-quality outputs pollutes the knowledge base and degrades search quality. The compounding loop needs a quality filter, but adding friction to filing reduces the rate of compounding. Current approach: the content intake SOP acts as the filter, but it’s manual.
• Measuring compound growth: Karpathy’s implicit metric is query quality over time. The Block article asks what our honest signal is. We don’t yet have a concrete way to measure whether the vault is actually getting more useful or just getting bigger. This is an open problem.