AI 应用层还没死，但要避开「Yellow Brick Road」！ @joeschmidtiv (a16z) 这篇文章指出：AI 应用层仍有巨大机会，但机会不在模型实验室正在全力押注的「通用智能体...

@joeschmidtiv (a16z) points out in this article that there is still significant opportunity in the AI application layer, but the opportunity lies not in the path of "generalized agents" that model labs are heavily investing in, but in the "workflow depths" of vertical, complex, and system-level applications. Founders and job seekers are anxious: Will OpenAI and Anthropic gobble up the entire application layer? https://t.co/X7ImYhRoGA" / X

The AI application layer isn't dead, but you need to avoid the 'Yellow Brick Road'!

(a16z) points out in this article that there is still significant opportunity in the AI application layer, but the opportunity lies not in the path of "generalized agents" that model labs are heavily investing in, but in the "workflow depths" of vertical, complex, and system-level applications. Founders and job seekers are anxious: Will OpenAI and Anthropic gobble up the entire application layer? Schmidt believes that this anxiety is "half right": · The right part: Labs will indeed swallow a lot of horizontal, general, and low-complexity application surfaces · The wrong part: The "application layer" is not a monolith; it cannot be generalized He uses The Wizard of Oz as a metaphor: · The Yellow Brick Road = The path labs are taking · Other places in Oz = Where startups should go What is the 'Yellow Brick Road'? Why is it dangerous? The Yellow Brick Road refers to: Taking the strongest model + ready-made connectors (Slack, Salesforce, GitHub, etc.) + simple agent orchestration → Building a universal AI colleague. The problem is that this is exactly what Cowork, Codex, and Claude Code are doing. If you are doing the same connectors, the same shallow orchestration, without sub-agents and deep configuration, and without distribution—you are competing head-on with the labs, and it's likely a dead end. The problems on the Yellow Brick Road (code generation, writing, images, etc.) share a common characteristic: Product quality improves linearly with the model's raw capability. Every additional dollar spent on pre-training/post-training improves the product. This type of problem is naturally suited for labs. Where are the opportunities in "Other Places in Oz"? Opportunities lie in complex, vertical, multi-step, multi-role problems where value is not just from model capability but also from an entire suite of scaffolding that makes the output trustworthy, compliant, and operable. Typical characteristics: · Cross-system gather context, then go through multiple human approval nodes · Involves legacy systems · Requires deterministic results, cannot tolerate ambiguity · Binds to real commercial outcomes (closing deals, underwriting, compliance review) The labs themselves admit they can't solve everything—so they are pouring heavy money into forward-deployed joint ventures (embedded joint projects) to help enterprises customize configurations. If "the next model version will solve it," they wouldn't invest that money. Why do the labs ultimately also "swallow" Other Places in Oz? 1. Data and learning flywheel · A lot of industry knowledge is not in the training set: unwritten norms, unwritten rules, practitioners' experience in their heads · Two flywheels: · Across customers: pattern recognition of similar problems · Single customer: unique exceptions and decision logic of that institution · Horizontal tools are difficult to design suitable UX to capture this knowledge; vertical players can customize interfaces around workflows 2. Model variability management · Labs can only push their own models; application companies can cross-vendor select models—the most suitable for each sub-task (open-source fine-tuning, competitor APIs, etc.) · Also do dirty work for clients: re-run evals every time the model upgrades, retune prompts for edge cases, smooth migrations · Clients get "market-leading intelligence + continuous upgrades" rather than "please migrate to our new model" 3. Cost optimization · Full Opus 4.7 = Negative margin · Vertical companies route by sub-task: cutting-edge models for hard problems, mid-range for bulk, self-developed/fine-tuned small models for narrow tasks · Labs set the "lowest intelligence you can buy for $X"; application companies sell the "lowest dollar cost required to complete this workflow" 4. Governance · Become the control plane for running AI in that vertical domain for the client: permissions, audits, what agents can do, what they actually did · Absorb regulatory complexity (HIPAA, SEC/FINRA, bar association rules, etc.) · Horizontal players cannot simultaneously be "one hundred vertical domains" of compliance partners Core trade-off: Labs must be everywhere for everyone → cannot be great at one thing. Three self-assessment frameworks: Are you in "Other Places in Oz"? Test | Yellow Brick Road (dangerous) | Other Places in Oz (opportunity) · Tool and step test | One step, one tool, results can be fault-tolerant (e.g., searching Google Drive) | Multi-step, multi-tool, output needs to pass partner/court/regulatory scrutiny · System vs tool test | Customer has "smart plugins" on their existing workflows; labs can come up with competitors and the customer can switch to you | Customers run their workflows through your system; you are the orchestration layer · Hedge fund/P&L test | Customers pay for generic capability (Claude seat can be replaced) | Customers pay for workflow-specific outcomes (closing deals, underwriting, compliance) Final judgment: Both paths will have big winners · Yellow Brick Road: Labs win—owning models + distribution of horizontal tools · Other Places in Oz: Application companies win—if they own a system of work (execution surface, data capture, governance) The model layer is fungible; the work system is irreplaceable. The next generation of enterprise software will be built outside these paths—application companies become the layer that integrates and delivers various new models, while customers depend on that system.