#36 — 4 principles for building truly useful AI products

Why it matters: In a landscape where AI models improve faster than developers can build applications, startup founders need a new playbook to avoid getting tumbled by the wave of AI progress.§

The big picture

Building AI products requires entirely new strategies that challenge traditional startup physics.§

Four essential principles

1. Don't solve problems that won't be problems soon§

• Predicting AI capabilities is now part of your job
• Focus on persistent challenges rather than temporary limitations
• Example: Granola ignored context window limitations (which limited transcriptions to 30min), betting correctly that LLMs would soon handle longer meetings

2. Your marginal cost is my opportunity§

• Unlike traditional software, AI has significant per-user costs
• This creates a unique advantage: startups can provide "Ferrari-level" experiences while big tech can only offer "Hondas"
• Today's expensive models become tomorrow's affordable commodities

3. Context is king§

• Don't just write step-by-step instructions for AI models
• Treat them like smart interns who need proper context to succeed
• "Context window selection" will become a defining concept beyond AI

4. Go narrow, go deep§

• To compete with general-purpose AI assistants, focus on excelling at specific use cases
• The "wrapper" around AI often determines product success
• Narrower focus makes it easier to identify and address AI failure cases

The bottom line

While AI accelerates everything, the fundamentals remain unchanged: build something people want, sweat the details, and create products that make users feel good when using them.§