The Unstuck Scaling Framework
Overview
A systematic approach to improving AI reliability by treating "getting stuck" as the primary bottleneck. Instead of broad improvements, painstakingly identify specific failure modes and create tight feedback loops.
Core principle: Address specific bottlenecks, not general intelligence.
The Cycle
┌─────────────────────────────────────────────────────────────────┐
│ │
│ ┌───────────────────┐ │
│ │ IDENTIFY │ │
│ │ 'Stuck' Points │ │
│ │ (auth, payments) │ │
│ └─────────┬─────────┘ │
│ │ │
│ ▼ │
│ ┌───────────────────┐ │
│ │ ADDRESS │ │
│ │ Specific │ │
│ │ Bottlenecks │ │
│ └─────────┬─────────┘ │
│ │ │
│ ▼ │
│ ┌───────────────────┐ │
│ │ QUANTITATIVELY │ │
│ │ Tune System │ │
│ │ (pass/fail rate) │ │
│ └─────────┬─────────┘ │
│ │ │
│ ▼ │
│ ┌───────────────────┐ │
│ │ FAST FEEDBACK │─────────────────────────┐ │
│ │ Loop │ │ │
│ └───────────────────┘ │ │
│ ▲ │ │
│ └───────────────────────────────────┘ │
│ │
└─────────────────────────────────────────────────────────────────┘
Key Principles
| Principle | Description |
|---|---|
| Specific blockers | Identify exact points where AI fails |
| Quantitative tuning | Measure stuck rates, not vibes |
| Fast feedback | Rapid iteration on fixes |
| Bottleneck focus | Specific roadblocks > general intelligence |
Common Mistakes
- Focusing on general model improvements
- Failing to measure "stuck" rates quantitatively
- Slow feedback loops preventing rapid iteration
Source: Anton Osika (Lovable, GPT Engineer) via Lenny's Podcast