The Untrainable

In the current AI landscape, many builders and investors feel despair because frontier models seem poised to absorb almost everything. But according to Sarah Guo, this view misses a critical category: untrainable work.

Measurable, public, and legible tasks (coding benchmarks, generic customer support scripts, standardized processes) are being commoditized rapidly from both below (open models) and above (labs folding scaffolding into weights). What remains valuable and defensible is work that is private, context-rich, high-stakes, and dependent on trust, permissions, accountability, and deep integration with proprietary systems.

Real automation isn’t just about better models — it requires organizational change, domain expertise, and long-horizon judgment that models struggle to replicate without years of specific, private experience. The highest-value opportunities now sit in the “frontier/private” quadrant: translating messy real-world realities into something models can act on, while maintaining the trust and accountability that only humans (or deeply embedded systems) can provide.

This Ledger Entry expands how readers think about value creation in the AI era by showing that while frontier models rapidly absorb measurable and public work, durable competitive advantage shifts toward "untrainable" domains — private systems, high-stakes judgment, deep organizational integration, and trust-based workflows that models cannot easily replicate or train against.