Paul Buchheit: Why Evals, Not Code, Are the Real AI Moat

YC now sees entire batches averaging 10% weekly growth and companies going from zero to $12M ARR in a year, often with very small teams; hitting $1M ARR in 6–12 months is becoming a baseline expectation for strong AI startups.

Instead of convincing customers they need AI, founders find enterprises under internal pressure to adopt it; the bottleneck is no longer demand but building agents that truly perform at or above human level for tasks like support and sales.

Founders report that meticulously labeled evaluation datasets and refined prompting strategies are often more valuable than their codebases, enabling reliable, predictable AI behavior that competitors can’t easily copy even if they use the same base models.

Successful teams frequently throw away old architectures, switch tools (e.g., from vector DBs to PGVector), and rewrite stacks every few months, gaining an edge over larger organizations that can’t update as quickly.

Tools like Cursor and Claude are becoming must-haves; some founders won’t hire engineers who don’t use AI codegen, and designers are skipping Figma in favor of text-to-code workflows, raising the baseline of what one person can produce.