Enterprise AI Requires Access Control and Custom Entity Definitions

The AI landscape is more complex than a single "brain to rule them all," requiring domain-specific models and data to create effective applications.

The Reality of Enterprise AI Applications

• Foundation models are critical infrastructure but insufficient alone for enterprise applications
• The "thin wrapper around GPT" concept is fundamentally flawed
  • This mirrors the 1980s misconception of "thin wrapper around a database" that underestimated what products like Salesforce would become

Why Domain-Specific AI Applications Win

• The universe of human behavior is "fat-tailed" - rare but important edge cases matter tremendously

  • Example: Waymo's self-driving cars struggled not with rain or sleet, but with unpredictable human behaviors like driving 75mph in a 25mph zone
  • "The number of rare important crazy shit that humans do is very high"

• Successful AI applications require multiple specialized models working together

  • Example: Cursor built 14 different models to understand how developers work
  • These models incorporate countless interactions with users to build domain expertise
  • "That's not just a thin layer on a foundation model"

Enterprise-Specific Challenges

• Access control becomes critical in enterprise environments

  • AI systems need to understand who has permission to see what information
  • Training on company data creates complex permission boundaries

• Semantic differences between organizations create fundamental challenges

  • "If you look at an enterprise, find 10 enterprises, they all have a different definition of what a customer means"
  • Is a customer a department at AT&T? The entire company? A specific person?
  • These definitions matter tremendously for metrics like churn

• The technology landscape is still evolving in unexpected ways

  • LLMs have "asymptoted" as we've run out of training data
  • Reinforcement learning continues to improve linearly but doesn't generalize well
  • "If you build a great programming model it may be an idiot at math"

Implications for Founders

• The application layer offers "almost unlimited" opportunities
• Building proprietary data and domain-specific models creates defensibility
• Understanding the specific needs of your domain is critical for success
• The problems AI can solve are far broader than what could be addressed with traditional software