Why should you create personal knowledge bases for quicker responses to queries? Andrej Karpathy explains

Karpathy's proposal for LLM-maintained personal wikis drew ~16 million views and over 100,000 bookmarks in two days, sparking rapid adoption and praise from figures like Jack Dorsey. He recommends replacing retrieval-augmented generation (RAG) with a three-layer system (raw sources, an LLM-updated plaintext wiki, and a guiding schema) that compiles knowledge once and keeps it current, shifting users to curators while the LLM handles summarisation, linking, and maintenance for use cases from personal tracking to research and internal team knowledgebases.

Analysis

A shift from per-query retrieval to a compiled, persistent LLM-maintained knowledge layer materially changes cost and product economics: expect front-loaded compute and human-in-the-loop curation costs up front, then ~5-20% lower marginal token/API spend per active user once a useful wiki exists. That front-loaded work creates a natural product wedge for firms who can offer reliable ingestion, schema governance, and audit trails — those capabilities become the new gatekeepers to enterprise adoption rather than raw retrieval latency or vector search accuracy.

Second-order winners are platforms that bundle editor/UIs + enterprise control (hybrid/local-first sync) because customers will pay to host canonical, auditable knowledge on-prem or in managed clouds; conversely, pure per-query API monetization models face margin pressure as usage profiles shift from many small queries to fewer, heavier compiles and periodic refreshes. Infrastructure demand will reallocate: more emphasis on orchestration, bookkeeping, and fine-tuning pipelines (steady recurring revenue), less on bursty GPU-hours for repeated reconstruction of the same facts.

Key risks and timing: real-world adoption needs standardized schemas, validation workflows and liability/forensics for persistent errors — if a compiled wiki embeds hallucinations they persist and amplify, creating regulatory and customer trust blowback within 3–12 months. Market catalysts to watch over the next 6–18 months are product integrations from major cloud/office vendors that expose wiki-as-core-primitive, pricing shifts from call-based to repo-based billing, and enterprise case studies showing >30% time savings in knowledge work — those will decide whether this is a niche productivity hack or a platform rewrite.

Key Decisions for Investors

• Buy Microsoft (MSFT) 12-month calls (10–15% OTM). Rationale: MSFT is best positioned to package compiled-wiki primitives into Office/Copilot and capture subscription economics. Target 2.5x return if enterprise integrations accelerate; stop-loss at 50% of premium if no material product announcements within 6 months.
• Buy NVIDIA (NVDA) 12–24 month LEAPS (ITM or slightly ITM calls). Rationale: even with lower marginal per-query spend, aggregate compute for initial compilation, fine-tuning, and private on-prem inference keeps demand for accelerators intact. Risk/reward ~3:1 assuming continued AI enterprise spend; hedge 25% with short-dated puts if guidance shows slowing data-center orders next quarter.
• Buy Salesforce (CRM) 9–12 month calls or add to core long position. Rationale: CRM and customer-support workflows are early beneficiaries of compiled knowledge (reduced support handle time, higher automation), and Salesforce can monetize via bundled knowledge-products. Target 1.8–2.0x in 9–12 months if product pilots convert to paid deployments; trim position if churn or pilot-to-pay conversion <10% after three quarters.