Foundation models upgrade every 6 to 12 months. Nobody at the application layer controls that curve. The only thing we choose is our posture relative to it.
Draw two lines, and there are exactly three postures.
Zero angle: overlap, get swallowed
Doing what the model's native evolution already does: general chat, general orchestration, general summarization. Every release replaces a slice of your product with native capability. This is the geometry of wrapper death.
Negative angle: betting on the model's flaws
Building value on what the model cannot do today: elaborate prompt chains, patches for weak reasoning, heavy scaffolding that corrects the model. Every time the model improves, these assets depreciate. You must build some of this so the product works today, but you cannot bet on it.
Positive angle: ride the curve
Value and model capability are multiplicative: value ≈ model capability × private context. The stronger the model, the more judgment it squeezes out of the same data, semantics and profiles. Every release is a free upgrade.
Others fear releases. We look forward to them.
The angle is per component
A company is not a point but a portfolio, and each component has its own angle. Ask every component the same question: when model capability goes up a notch, does this appreciate, hold, or depreciate?
- Negative derivative · consumables: multi-agent orchestration frameworks, prompt chains, patches and fallbacks for model flaws. Build them because you must, but build them thin and removable, and depreciate them in your head from day one. Never write them into the moat story.
- Zero derivative · positions: distribution, customer trust, compliance, workflow embedding. Model upgrades don't touch them, but watch for general agents growing browsers and tool hands and eroding them from the side.
- Positive derivative · first-class assets: the private industry database, transaction flow, tacit policy library, knowledge graph and semantic metric layer, user profiles, private eval sets. On the day a new model ships, these appreciate automatically.
There is only one strategic move: keep shifting investment from negative-derivative components to positive-derivative ones.
Two things easy to miss
Cost is a curve too. Beyond capability, inference cost keeps falling. High-frequency scanning tasks should be designed at the scale that makes sense after a 10× cost drop, with budgets and degradation boundaries in place.
Migration speed is itself an angle. Build a private eval suite from real questions, real queries, report adoption and transaction outcomes; replay it when a new model ships, then decide whether to switch. A team that can safely move to a new model in a week and one that needs a quarter stand at different points on the same curve.
One line
Be a complement to the foundation models, not a substitute. Complements get multiplied at every release. Substitutes get subtracted.