Whether anything you build with somebody else’s model is defensible, and what would make it so. The honest answer for most products is that the moat is not the model and not the data, which makes it worth knowing what it actually is before building a strategy on a claim that does not hold.
| # | Lesson | The question it answers |
|---|---|---|
| 01 | Is Your Data Actually a Moat | What is the honest test, and how many companies pass it? |
| 02 | The Feedback Flywheel | How does usage become an eval set and then become quality? |
| 03 | The Thin Wrapper Critique | Where does value accrue when the model is somebody else’s? |
| 04 | Switching Costs for AI Products | What actually holds a customer when the model is a commodity? |
| 05 | Workflow and System of Record as Defence | Why does owning the workflow beat owning the model? |
| 06 | Data Rights, Consent and Training | What may you use, and what must you promise not to? |
| 07 | Proprietary Evals as an Asset | Why is the eval set often worth more than the training data? |
| 08 | Distribution as the Real Moat | Why do incumbents win features and lose categories? |
| 09 | Build, Buy or Wait for the Model | How do you decide when next year’s model may do it for free? |
| 10 | Capability Overhang | How do you plan against a frontier that moves every six months? |
| 11 | What Compounding Actually Looks Like | Which loops really compound, and over what timescale? |
| 12 | Worked: A Defensibility Argument | One real product, argued honestly, including what is not defensible |