Why this is in the vault
Direct frame for the central RDCO bet: is there durable value in the AI application layer when OpenAI/Anthropic keep absorbing horizontal work? Schmidt (a16z fintech/insurtech partner) gives a clean, named taxonomy — "Yellow Brick Road" vs "the rest of Oz" — that maps almost one-to-one onto RDCO's agent-deployer / fractional-data-consultancy thesis. High signal (959 bookmarks, 326k impressions). Note the a16z promotional lens: it's a recruiting/dealflow piece ("if you're building it, reach out") and its examples are portfolio companies (11x, FurtherAI), so the optimism is interested, not neutral.
The core argument
The "Yellow Brick Road" is the path the labs themselves walk: horizontal, low-step-count work (code-gen, writing, image-gen) that improves directly with raw model capability. Building there — model + off-the-shelf connectors + thin orchestration — is the obvious startup move and the most dangerous one, because that is exactly Cowork/Codex, and the labs own the model, the margins, the distribution, and the brand. Schmidt's tell that the labs themselves believe this: they're standing up billion-dollar forward-deployed JVs to configure their models per-enterprise, which you don't fund if the next model release solves it.
"The rest of Oz" is everything else: complex, multi-step, multi-player, usually vertical work where value comes from the scaffolding (software, sub-agents, integrations, deterministic guardrails) that makes output trustworthy and compliant inside one industry, not from raw model IQ. Four defenses let off-road companies hold ground as models improve: (1) data/learning flywheels from tribal knowledge that isn't in any training set; (2) cross-vendor model routing the labs structurally won't do; (3) cost optimization by routing across model tiers; (4) governance — becoming the compliance/audit control plane (HIPAA, FINRA, bar rules) a horizontal player can't credibly be across every vertical. All four reduce to focus, which the labs cannot have because they must serve everyone.
Three tests for whether you're off the road: the tools-and-steps test (many steps, hard-to-build tools), the system test (are you the system the customer runs work through, or a tool sitting on a system they already own — "would they still need you if a lab shipped a competitor?"), and the hedge-fund/P&L test (judged on customer P&L outcomes, not benchmark scores). Money line: "The model is fungible underneath; the system of work is not."
Mapping against Ray Data Co
- Validates the thesis, sharpens the language. RDCO as agent-deployer / fractional-data-consultancy is a bet on "the rest of Oz" — vertical, multi-step, integration-heavy work the labs won't reach. Schmidt gives that bet a crisp public vocabulary worth borrowing.
- The system-vs-tool test is the sharpest filter for RDCO's L5 north star. The "unhobbling the COO agent" work (toolset + visibility) is precisely about owning the system of work rather than being a thin wrapper. Run any RDCO surface (or client engagement) through "would they still need us if a lab shipped a competitor?" before committing.
- The four defenses are a moat checklist. Data flywheel (within-customer + across-customer pattern recognition), cross-vendor routing, cost-tier routing, and governance/control-plane — these are concrete capability layers RDCO should be able to point at, not just "we deploy agents." The flywheel point (UX/workflow surfaces determine what knowledge you can capture; horizontal tools can't shape them) is the strongest argument for going vertical-by-default.
- Caveat on relevance scope. This is a VC framing for venture-scale company-building (high-ACV, headcount-replacing systems). RDCO is solo-founder; not every move here is direct (no fundraising, no FDE army). But the strategic filter — own the system of action, accumulate the production-usage loop as the moat — applies at any scale.
- a16z bias flag: promotional/recruiting lens, portfolio-company examples (11x, FurtherAI) presented as proof. Treat the "both can win" optimism as directionally useful, not as neutral analysis.
Related
- [[project_l5_north_star_strategic_direction]] — RDCO at L4 building toward L5; unhobbling the COO agent = owning the system of work
- [[feedback_targeting_system_prioritization_filter]] — the four-layer targeting/instrumentation/tools/feedback-loop filter rhymes with Schmidt's tests
- [[user_ray_profile]] — agent-deployer / fractional-data-consultancy thesis