One thing the tools-and-steps test undersells: it's not the step count, it's the step irreversibility. An insurance bind, a sent outbound, a redline filed with opposing counsel... every step shrinks the customer's recoverable action space, and that's the part they're actually paying a vertical company to absorb.
Labs sell intelligence per token. Rest-of-Oz companies sell accountability per outcome and accountability doesn't ship in a model release which is exactly why the moat compounds.
Thanks @Joe, but this post is mostly hopium. It's not that the labs can't go vertical. It's that they have enough capital to go BOTH vertical AND horizontal. Startups simply do not have the capital or capital-attractors to compete in both fundraising and attracting world-class AI talent. AI Application death is guaranteed as the labs continue to raise and improve their models.
Change is also inevitable, so this state won't last forever, but there is no moat for software apps against the labs. Not even the new "model-routing-cost-management" moat flavor will protect us. The labs can create low-cost models too. Come on. Let's be intellectually honest.
We need to embrace the death of software-application startups to see what's on the other side.
Joe, I am privileged to receive such a kind and fast reply. Let's debate!
Here's my proposal: All Software Application startup companies should die and become Agentic Consulting Firms. Since all the labs are attempting to build AND buy their way into the enterprise, we startups should morph into what they want to acquire: Forward Deployed Engineering Firms. We provide both trust, judgment, and deep expertise. We price consulting services, expertise services, and token-cost savings services. We temporarily sacrifice collecting margin to fund R&D on software products in order to secure our survivability as service firms in search of a new type of moat that cannot be replicated by the labs latest models and marketing.
Cute metaphor, Joe, but while you correctly diagnosed the disease (Tokenmaxxing is dead), you completely hallucinated the cure.
The 'Rest of Oz' isn't vertical software scaffolding. Software cannot act as its own physical envelope. When an enterprise agent hits a causal reasoning wall and spins into a multi-million-dollar infinite loop, your 'specialized compliance wrapper' is going to melt the cloud infrastructure right alongside the generic apps.
True governance isn't a software moat; it's a thermodynamic circuit breaker. Until you pull risk control down to the bare metal via a silicon Root of Trust, you aren't avoiding death on the Yellow Brick Road—you’re just dying in a more expensive outfit. The Physicsmaxxing era is here.
We are building “The Turing Brake” which will enable all your other investments.
What makes this useful is the everywhere-vs-great-at-one frame Joe lays out, because it pulls the AI build-vs-buy question out of the technology layer and into the strategy layer. The 11x and FurtherAI examples land harder than most a16z portfolio call-outs because they're shaped around what the customer's regulator actually requires, not what the model can technically do.
I write about AI for non-technical leaders, and the question of where to specialize comes up in every adoption conversation I cover.
The version of this at the senior level is a personnel question. Which roles on your team are you protecting against horizontal Yellow Brick Road consolidation, and which are you actively pushing into vertical depth? Most leaders treat both as the same hiring conversation. They aren't.
The four-blank answer template tightens it: one named workflow your team is going deeper on (not faster on), one named system of record you intend to own, one named regulator-or-stakeholder rule you're choosing as the moat, one named date to revisit the bet. Without the regulator-rule blank, you're building a wrapper.
Curious whether any leaders here are explicitly picking which functions to specialize and which to hand to the horizontal models.
The article’s distinction frames the application-layer question well, but understates the governance problem. In regulated work, governance is not secondary. Ask a trading desk.
Owning the workflow is not the same as governing generative behavior inside the workflow.
In complex, regulated, multi-step work, the key question is not only what the agent did after the fact. It is whether the system stayed attached to the governing objective while the work was unfolding.
The Yellow Brick Road framing makes sense for US startups, but the opportunity map looks structurally different in Chinese markets from what I track in Chinese-language sources daily.
The 'rest of Oz' thesis assumes the data flywheel comes from earning customer trust inside workflows. In China's AI application layer, the data is often assigned rather than earned -- government contracts hand SenseTime, iFlytek, or Megvii training data from surveillance networks, medical records, or factory sensors before the product ships. The moat is built through policy positioning, not customer relationships.
Same insight -- proprietary data compounds -- but the mechanism is top-down rather than bottom-up. Makes the competitive dynamics in vertical AI quite different across the Pacific.
There need to be guard rails set up for labs so they don’t run over sectors and start ups just because they can. I am pretty sure we will need to get to that point sooner than later.
FurtherAI: "What separates the two carriers is everything inside it: which risks get escalated, which loss signals matter, which appetite rule wins when two of them conflict, when a human has to sign off, which external data gets pulled in, and how the final decision gets documented." When a company wins on a workflow or data that its rivals lack, that workflow is the edge, and handing it to a shared vendor carries a heavy cost. The vendor trains on the company's own data and folds the result into its model, then sells that improved model to the same rivals the company beats today. The harder the vertical, the more of the edge lives inside that data and that process, why would you outsource your competitive advantage?
One of the more important ideas in this piece is the distinction between AI tools and AI-native operational systems.
Regulated industries don’t run on prompts. They run on workflows, approvals, controls, auditability, operational memory, and deterministic execution.
That’s why we believe the next generation of banking platforms won’t look like “AI assistants sitting on top of legacy cores.” They’ll look more like operational fabrics where intelligence, governance, and execution are built directly into the system itself.
The model is important.
The system surrounding the model is where the moat gets built.
The hedge fund / P&L test maps directly to a vertical I think you under-weighted: investment research itself.
If theses are versioned graphs (root claim, MECE branches, hypothesis, kill conditions) instead of prose summaries, agents can traverse them, propagate invalidation, and compound a track record nobody can replicate without sitting inside the workflow for years.
That’s Draw Tree (90s.pm.investing). Solo founder, Hong Kong,
MCP server publishing next week as v0.1.
Calling it the FIX protocol of AI-native research. Either it’s a category, or I’m wrong — but the kill conditions firing in production are auditable.
It’s very simple, a jack of all trades is a master of none
One thing the tools-and-steps test undersells: it's not the step count, it's the step irreversibility. An insurance bind, a sent outbound, a redline filed with opposing counsel... every step shrinks the customer's recoverable action space, and that's the part they're actually paying a vertical company to absorb.
Labs sell intelligence per token. Rest-of-Oz companies sell accountability per outcome and accountability doesn't ship in a model release which is exactly why the moat compounds.
Thanks @Joe, but this post is mostly hopium. It's not that the labs can't go vertical. It's that they have enough capital to go BOTH vertical AND horizontal. Startups simply do not have the capital or capital-attractors to compete in both fundraising and attracting world-class AI talent. AI Application death is guaranteed as the labs continue to raise and improve their models.
Change is also inevitable, so this state won't last forever, but there is no moat for software apps against the labs. Not even the new "model-routing-cost-management" moat flavor will protect us. The labs can create low-cost models too. Come on. Let's be intellectually honest.
We need to embrace the death of software-application startups to see what's on the other side.
Embrace debate! I of course document why I feel differently but appreciate you weighing in. We'll see!
Joe, I am privileged to receive such a kind and fast reply. Let's debate!
Here's my proposal: All Software Application startup companies should die and become Agentic Consulting Firms. Since all the labs are attempting to build AND buy their way into the enterprise, we startups should morph into what they want to acquire: Forward Deployed Engineering Firms. We provide both trust, judgment, and deep expertise. We price consulting services, expertise services, and token-cost savings services. We temporarily sacrifice collecting margin to fund R&D on software products in order to secure our survivability as service firms in search of a new type of moat that cannot be replicated by the labs latest models and marketing.
Cute metaphor, Joe, but while you correctly diagnosed the disease (Tokenmaxxing is dead), you completely hallucinated the cure.
The 'Rest of Oz' isn't vertical software scaffolding. Software cannot act as its own physical envelope. When an enterprise agent hits a causal reasoning wall and spins into a multi-million-dollar infinite loop, your 'specialized compliance wrapper' is going to melt the cloud infrastructure right alongside the generic apps.
True governance isn't a software moat; it's a thermodynamic circuit breaker. Until you pull risk control down to the bare metal via a silicon Root of Trust, you aren't avoiding death on the Yellow Brick Road—you’re just dying in a more expensive outfit. The Physicsmaxxing era is here.
We are building “The Turing Brake” which will enable all your other investments.
Let’s talk.
What makes this useful is the everywhere-vs-great-at-one frame Joe lays out, because it pulls the AI build-vs-buy question out of the technology layer and into the strategy layer. The 11x and FurtherAI examples land harder than most a16z portfolio call-outs because they're shaped around what the customer's regulator actually requires, not what the model can technically do.
I write about AI for non-technical leaders, and the question of where to specialize comes up in every adoption conversation I cover.
The version of this at the senior level is a personnel question. Which roles on your team are you protecting against horizontal Yellow Brick Road consolidation, and which are you actively pushing into vertical depth? Most leaders treat both as the same hiring conversation. They aren't.
The four-blank answer template tightens it: one named workflow your team is going deeper on (not faster on), one named system of record you intend to own, one named regulator-or-stakeholder rule you're choosing as the moat, one named date to revisit the bet. Without the regulator-rule blank, you're building a wrapper.
Curious whether any leaders here are explicitly picking which functions to specialize and which to hand to the horizontal models.
The article’s distinction frames the application-layer question well, but understates the governance problem. In regulated work, governance is not secondary. Ask a trading desk.
Owning the workflow is not the same as governing generative behavior inside the workflow.
In complex, regulated, multi-step work, the key question is not only what the agent did after the fact. It is whether the system stayed attached to the governing objective while the work was unfolding.
The Rest of Oz still needs runtime control.
The Yellow Brick Road framing makes sense for US startups, but the opportunity map looks structurally different in Chinese markets from what I track in Chinese-language sources daily.
The 'rest of Oz' thesis assumes the data flywheel comes from earning customer trust inside workflows. In China's AI application layer, the data is often assigned rather than earned -- government contracts hand SenseTime, iFlytek, or Megvii training data from surveillance networks, medical records, or factory sensors before the product ships. The moat is built through policy positioning, not customer relationships.
Same insight -- proprietary data compounds -- but the mechanism is top-down rather than bottom-up. Makes the competitive dynamics in vertical AI quite different across the Pacific.
There need to be guard rails set up for labs so they don’t run over sectors and start ups just because they can. I am pretty sure we will need to get to that point sooner than later.
FurtherAI: "What separates the two carriers is everything inside it: which risks get escalated, which loss signals matter, which appetite rule wins when two of them conflict, when a human has to sign off, which external data gets pulled in, and how the final decision gets documented." When a company wins on a workflow or data that its rivals lack, that workflow is the edge, and handing it to a shared vendor carries a heavy cost. The vendor trains on the company's own data and folds the result into its model, then sells that improved model to the same rivals the company beats today. The harder the vertical, the more of the edge lives inside that data and that process, why would you outsource your competitive advantage?
One of the more important ideas in this piece is the distinction between AI tools and AI-native operational systems.
Regulated industries don’t run on prompts. They run on workflows, approvals, controls, auditability, operational memory, and deterministic execution.
That’s why we believe the next generation of banking platforms won’t look like “AI assistants sitting on top of legacy cores.” They’ll look more like operational fabrics where intelligence, governance, and execution are built directly into the system itself.
The model is important.
The system surrounding the model is where the moat gets built.
The hedge fund / P&L test maps directly to a vertical I think you under-weighted: investment research itself.
If theses are versioned graphs (root claim, MECE branches, hypothesis, kill conditions) instead of prose summaries, agents can traverse them, propagate invalidation, and compound a track record nobody can replicate without sitting inside the workflow for years.
That’s Draw Tree (90s.pm.investing). Solo founder, Hong Kong,
MCP server publishing next week as v0.1.
Calling it the FIX protocol of AI-native research. Either it’s a category, or I’m wrong — but the kill conditions firing in production are auditable.