Karpathy reveals how to build a personal AI knowledge base with three folders

3 folders (raw, wiki, outputs) plus a single schema file (CLAUDE.md or AGENTS.md) is Karpathy's recommended minimal setup to build an AI-driven personal knowledge base. The AI ingests raw materials, rewrites them into topic-based wiki files and outputs answers; once the wiki exceeds ~10 documents it becomes queryable and self-improving, but errors can compound so he recommends a monthly AI-driven review to find contradictions and unsourced claims.

Analysis

This “three‑folder + schema” pattern is a demand accelerator for compute, inference and endpoint security rather than a death knell for incumbent knowledge‑management UX. If users wire hosted LLMs (Claude/Cursor) into local folders at scale, expect a measurable bump in API calls and cloud inference hours within 3–12 months; that flow monetizes cloud providers (Azure/GCP/AWS) and inference hardware vendors (NVIDIA/AMD) more than niche CRM/KB UX players. The mechanic: low friction → lower acquisition cost and higher activation rates, which convert lightweight experiments into persistent paid usage of underlying compute/LLM stacks.

Second‑order winners are vendors who can offer governance hooks and telemetry into the simple folder pattern. Security vendors that already own endpoint telemetry (CrowdStrike, SentinelOne) will be able to upsell DLP/ML‑aware policies to protect these emergent personal KBs; enterprises will pay a 10–30% premium for compliant ingestion pipelines within 6–18 months. Conversely, deeply integrated but heavy KM suites (highly customized Confluence deployments, boutique knowledge‑graph integrators) face revenue compression as teams adopt DIY flows; expect deal sizes to shrink even if seat counts hold.

Key tail risks: (1) data exfiltration and regulatory/legal pushback — a high‑profile breach or an LLM‑API TOS change could curtail mass adoption within weeks; (2) error accumulation inside AI‑compiled wikis — without human governance, downstream analytics are brittle and could produce systematic decision risk for firms relying on them. Watch two catalysts: influencer momentum + open‑source agent tooling adoption (fast, weeks→months) and enterprise security product integrations (slower, quarters→year).

Key Decisions for Investors

• Long MSFT (6–12 months): overweight Azure + GitHub Copilot exposure to increased hosted LLM usage. Position size: tactical 3–5% of tech sleeve. Risk/Reward: limited downside vs large upside if enterprise adoption accelerates; hedge with 3–6 month calls if volatility spikes.
• Long NVDA (9–18 months) or AMD for edge/inference exposure: buy NVDA outright or NVDA Jan-2027 LEAPS to capture secular inference demand. Size 2–4% portfolio. Risk/Reward: high concentration risk but asymmetric upside from sustained API/inference growth.
• Long CRWD (6–12 months): security demand from personal KB adoption should drive DLP/endpoint policy upsells. Consider buying CRWD or a 6–12 month call spread to limit premium. Risk/Reward: moderate upside if adoption drives enterprise policy spend; downside if macro weakens renewals.
• Short TEAM or BOX (12 months, small hedge): initiate a small short/put spread against high multiple KM players where deal sizes are most at risk from DIY folder+AI adoption. Keep exposure minimal (1–2%); risk limited by defined spread. Rationale: revenue per customer pressure even if total seats stable.