AI is in a sedimentation phase

The report argues that AI isn’t in a phase of explosive acceleration but in a “sedimentation” phase where value settles around verifiable, persistent workflows rather than ever-smarter models.
Core thesis

- Demand for AI is polarizing into two extremes—programming and creative roleplay—while middle tasks like generic Q&A and translation remain thin.
- This pattern is driven by **verification**: where outputs are easy to test and automate, usage compounds; where correctness depends on human judgment, progress stalls.

## Verification as the key divider

- In programming, verification is cheap, repeatable, and automatable: code can be executed, tested, and re-run in tight feedback loops, so improvements compound and elasticity (responsiveness of demand to quality/price) stays high.
- In judgment-heavy domains (legal memos, financial analysis), correctness reduces to expensive, non-repeatable human review, so small quality gains (<5%) do not justify large adoption, flattening the cost-usage curve.

## Retention and the “glass slipper” effect

- Models like Gemini 2.5 Pro and Claude 4 Sonnet showed around 40% retention by month 5 in mid‑2025, indicating that when a model fits a specific, high-value workflow very well, users get locked in.
- However, this fit is only revealed through high-cost real-world verification, so value accrues not to generic “smartest model” claims but to specific embedded workflows where state survives repeated use.

## Constraints, infrastructure, and where value settles

- Value pools where something important cannot scale quickly: examples include silver and gold miners and semiconductor manufacturing outperforming broader IT, presented as “constraint trades.”
- In AI, models iterate quickly, but infrastructure layers that preserve state—memory systems, orchestration, recovery, observability—scale slowly and harden through failure, so value accumulates there rather than in the models alone.

## Strategic implications with verification emphasis

- For builders, the recommendation is to focus on **stateful systems** (memory, orchestration, cache management) that preserve continuity so verification and improvement happen over persistent workflows, not one-off prompts.
- For operators, investing in verification infrastructure—even crude multi-judge setups—creates structural gains because it turns messy judgment work into something partially testable and optimizable over time.
- For investors, the suggested edge is to look for “sedimentation” plays in infrastructure and constrained resources, where systems slow down and value quietly accumulates instead of chasing headline capability jumps.