Model-Market Fit

• Legal AI was stuck for years until GPT-4 crossed the capability threshold in March 2023—within 18 months, the market minted more unicorns than the previous decade combined, including Casetext's $650M acquisition
• The test for MMF: Can the model produce output customers would pay for without significant human correction? Legal reasoning hits 87% accuracy (production-grade), financial analysis sits at 56% (not there yet)
• "Human-in-the-loop" positioning reveals missing MMF—when it's a feature, the AI does the work; when it's a crutch, humans are compensating for AI that can't perform the core task
• The dangerous zone is betting on MMF that's 24-36 months away—close enough to seem imminent, far enough to burn through multiple funding rounds waiting
• The next capability frontier isn't just accuracy but sustained autonomous operation over days, not minutes—persistence, recovery, coordination, and judgment across multi-day workflows

The author introduces Model-Market Fit (MMF) as a prerequisite layer beneath Andreessen's product-market fit framework. For AI startups, market demand alone can't pull a product out if the underlying model capabilities don't cross the threshold needed to actually perform the core task. This explains why certain AI verticals suddenly explode after years of dormancy—not because the market changed, but because model capability crossed a critical threshold.

The evidence is concrete: Legal AI struggled for years with classification-focused models that couldn't handle the generation and reasoning lawyers needed. The author founded Doctrine in 2016 and experienced this firsthand—investors saw legal AI as niche with limited upside. Then GPT-4 arrived in March 2023, and within 18 months the market minted more unicorns than the previous decade, including Thomson Reuters acquiring Casetext for $650 million. Similarly, Cursor existed before Claude Sonnet but was "meh"—the author installed and deleted it multiple times. After Sonnet dropped, it became impossible to work without. The product didn't change; the underlying model crossed the threshold.

The framework reveals a brutal strategic dilemma: build for current MMF (safe but limited markets) or anticipated MMF (massive markets but existential timing risk). The dangerous zone is 24-36 months out—close enough to seem imminent, far enough to burn multiple funding rounds. The author provides a practical test: can the model, given the same inputs a human expert receives, produce output customers would pay for without significant human correction? In regulated verticals, the gap between 80% and 99% accuracy is often infinite—a contract review tool that misses 20% of critical clauses isn't augmenting lawyers, it's creating liability. The next capability frontier isn't just accuracy but sustained autonomous operation—current MMF examples handle tasks measured in minutes, but the highest-value knowledge work requires agents that can maintain goals, recover from failures, and coordinate subtasks across days or weeks.