Models Trained on Models Trained on Mimicry

In a time when LLMs and LRMs are marketed as tools that “reason,” many of us are now using these same tools to generate business models, especially for AI-driven ventures. But what are these models really doing?

Remixed from ivodich

They’re not reasoning. They’re remixing.

SaaS emerged as a structured mimicry of stability: recurring revenue, low churn, scalability. It made sense because it looked like what venture capital wanted at the time: predictable, expandable cash flows with high gross margins. I remember, because I learned SaaS from “my VC” who was training its analysts to be experts in the emergence of this promising model and to reshape the logic of founders who were used to the older practices of licensing and services.

The logic was seductive. And it worked.

But now, when we ask LLMs to generate business models for new ideas, especially in AI, we’re pulling from a vast training corpus of other mimicries. Pitch decks. VC blogs. YC applications. All of them optimized not for the thing being built, but for the story that secures optionality.

“Artificial intelligence learns to reason”1

This is meta-mimicry, as Shannon Vallor puts it. A surface-level coherence that sounds like thought. A convincing sequence of moves that looks like logic. But there is no reasoning behind the curtain—just the smooth probability surface of what sounds like reasoning.

Which sounded like me, in my early years of business model swagger, just making it up as I went along, crossing my fingers behind my back, hoping people wouldn’t question my assumptions. Waving hands. Pointing forward.

And so, we now see the proliferation of AI-generated business models for AI businesses, justified by AI-generated strategy rationales, tuned to investor expectations that were shaped by previous generations of performative models. It’s mimetic turtles all the way down.2

This is not a model of contribution.

It is a model of persuasion.

What’s missing is the thing that most business models forget to ask: What is this company here to do? What system will it transform? What relationships will it rewire? What contribution: material, social, ecological, will it make? Geez, even to margins, will the unit of use ever cost less to generate than the company is paid?

And what will be lost if it scales before figuring that out?

I think lots about Reason, having named my company after the formerly-named Reason Street in NYC.

At Reason Street, I’ve always understood business models as stories. But not just stories to win capital, they are speculative blueprints for how value flows through an ecosystem. When done well, they clarify trade-offs, surface values, and help founders choose what not to do.

When done poorly, or trained on mimicry, they obscure reality. They manufacture confidence in copy-paste answers.

And when we automate the creation of business models without confronting this meta-mimicry, we risk building companies optimized for investor performance, not world performance.

We start to forget that the true reasoning of business happens not in step-by-step chains of thought, but in situated action. Through messy experiments. Through constraints. Through dialogue with regulators, customers, partners, and skeptics. Through the friction of the real.

So here’s a provocation:

Don’t ask AI to generate your business model.
Ask it what kinds of business models it has seen, and what business model their owners have chosen for them.

Then start again.