## Progressive Summary **Executive Summary (Layer 3):** Shopify's River story credits *osmosis learning* (watching public work) for its gains, but the durable, scalable mechanism is **capture + reuse** — the conversation becoming a searchable artifact — not ambient watching, which cannot scale to 4,450 channels. **Key Insight (Layer 2):** The headline 36%→77% merge-rate jump is *agent ratcheting* (humans writing skills back into River), not osmosis — a private feedback queue would produce the same result. **Context (Layer 1):** Skeptical companion to [[Lehrwerkstatt — Osmosis Learning Through Visible Work]], critiquing Tobi Lütke's May 2026 post. **Cross-Domain Connections:** [[Lehrwerkstatt — Osmosis Learning Through Visible Work]] **Discoverability Score:** 8/10 ## Atomic Insight Lütke braids two mechanisms and lets the headline number validate the romantic one: - **Osmosis (watching)** — judgment seeps into bystanders. His own apprenticeship worked at one workbench with a few masters; River spans 4,450 channels. You cannot stand near 4,450 workbenches — attention, not access, is the bottleneck. What replaces "hanging around" is deliberate **search**, a different and rarer skill. - **Capture + ratcheting** — public conversations become searchable artifacts and corrections get written back into the agent. This is what compounds, and it needs a captured medium, not an audience. Costs the post omits: - **No denominator** — "1 in 8 merged PRs" with 1,870 opened/week; merge *volume* and review-flooding cost are never given. - **Inversion risk** — public visibility can suppress the junior questions it means to teach. - **No backstage** — "speed of its slowest secret" treats all privacy as friction; some slowness is deliberation. Takeaway: design for **capture and retrieval first**; treat watching as a bonus, not the engine. --- *Source: [[Learning on the Shop floor]] — Tobi Lütke, x.com (May 2026)*