Winning Products Get Stronger Every Day

The moat is the business learning system that surrounds the software.

The question has shifted from:

How do we build something?

to:

How do we build something that gets stronger every day? Something that strengthens itself.

This is a nice X article, representing some ultra-modern thinking in the world of startups, given that AI has made code generation so quick.1

In it, Sid talks about:

1. Proprietary data that improves with use

2. Infrastructure-level trust (deep operational dependency)

3. Permission moat (regulation, licenses, compliance)

4. Structural distribution (platform default, ecosystem position)

5. Community & brand ecosystem (users building on top of you)

6. Accumulated capital / liquidity depth (financial network effects)

7. Physical infrastructure (atoms that are slow to replicate)

Things AI can copy quickly:

• features

• UI

• APIs

• models

Things that remain slow to build:

• trust

• distribution

• habits

• institutions

• ecosystems

• real-world feedback loops

These “slow to build” things all share one property in that they require time and real-world interaction to accumulate. You cannot generate them with a prompt…they’ll be a photocopy of a photocopy of a photocopy…

The best products are those that improve with user interaction. Every time someone logs in and uses the product, the business behind it learns something new: the product captures behaviour, outcomes, the corrections and edge cases: all the things that make the system smarter than it was yesterday, last week, last month, and feeds it back into the business.

We need the system itself to improve as it accumulates a history of real-world use that competitors cannot easily capture and reproduce. They might be able to copy the basic features, but they can’t do anything else because they don’t have years of interactions and learning behind them. All that data turns into knowledge.

As the system runs, it must become more accurate and reliable, becoming deeply embedded in the workflows of its end users. This is the product’s compounding history. This is what is defensible.

This would require a pipeline – the product captures signals from real usage and feeds them into a data pipeline that aggregates patterns across all usage and interactions. The system then updates itself. The interesting part is that we can use AI to take advantage of our real-world usage to improve the system, using our business and product expertise, of course. This generates new signals that are fed back into the system, hopefully creating additional users and usage.

The product stops being static and starts adapting on its own, with nudges from the business.

Will the smartest AI systems still make better decisions than human beings, or a group of human beings running the business? Personally, I don’t think so, certainly not now, though it is possible in the…remote…future. While the product system I’m talking about could update itself, it can’t do so completely on its own. The business still needs to set the goals and guardrails that will determine how the system evolves. The business will constantly have to “nudge” the system into the right shape and direction. These nudges may only be small and infrequent, but they will be absolutely required.

In some ways, these systems resemble antifragile systems. Instead of being defined by mistakes, edge cases or unexpected behaviour, they learn from them. However, unlike the definition of antifragility created by Nassim Nicholas Taleb, the underlying mechanism isn’t randomness: it’s feedback. A product with no feedback will be brittle and fragile.

https://x.com/sidrmsh/status/2029339145114374256