jona Andrew Ng: Building Faster with AI
Guidance for applying Andrew Ng's principles on startup execution speed and AI-accelerated development, based on AI Fund's experience building approximately one startup per month.
Core Thesis
Execution speed is the strongest predictor of startup success. AI coding tools enable 10x faster prototyping, shifting bottlenecks from implementation to product management and user feedback.
AI Stack Layers
Understand where opportunities exist in the AI technology stack:
Layer 5: Applications ← Biggest opportunities (must generate revenue for all layers below)
Layer 4: Orchestration ← New layer enabling easier app building (agentic coordination)
Layer 3: Foundation Models ← High PR/hype but dependent on app layer revenue
Layer 2: Cloud Infrastructure
Layer 1: Semiconductors
Key insight: Media focuses on technology layers, but by definition, the application layer must generate the most value to sustain the entire stack.
Agentic AI Workflows
Why Agentic Approaches Matter
Traditional LLM usage forces linear generation (first word to last, no backspace). Humans don't write well this way—neither does AI.
Agentic workflow pattern:
- Generate outline/plan
- Conduct research (web fetches, document retrieval)
- Write first draft
- Self-critique the draft
- Revise based on critique
- Iterate until quality threshold met
This iterative loop is slower but produces dramatically better results.
When to Use Agentic Workflows
Apply agentic approaches for:
- Complex compliance document analysis
- Medical diagnosis reasoning
- Legal document interpretation
- Any task requiring research + synthesis + revision
Implementing Agentic Workflows
┌─────────────────────────────────────────────┐
│ Agentic Workflow Loop │
├─────────────────────────────────────────────┤
│ │
│ ┌──────────┐ │
│ │ PLAN │ ← Define task, create │
│ └────┬─────┘ outline or approach │
│ │ │
│ ▼ │
│ ┌──────────┐ │
│ │ RESEARCH │ ← Fetch context, retrieve │
│ └────┬─────┘ documents, web search │
│ │ │
│ ▼ │
│ ┌──────────┐ │
│ │ DRAFT │ ← Generate initial output │
│ └────┬─────┘ │
│ │ │
│ ▼ │
│ ┌──────────┐ │
│ │ CRITIQUE │ ← Self-evaluate against │
│ └────┬─────┘ requirements │
│ │ │
│ ▼ │
│ ┌──────────┐ No │
│ │ GOOD │────────┐ │
│ │ ENOUGH? │ │ │
│ └────┬─────┘ │ │
│ │ Yes │ │
│ ▼ │ │
│ ┌──────────┐ ┌────┴─────┐ │
│ │ OUTPUT │ │ REVISE │──────┐ │
│ └──────────┘ └──────────┘ │ │
│ ▲ │ │
│ └────────────┘ │
│ │
└─────────────────────────────────────────────┘
Startup Execution Speed Principles
Speed as Success Predictor
High-velocity execution correlates with startup success. AI Fund observes this across their portfolio of monthly startup launches.
AI-Enabled Acceleration
Current AI coding tools enable approximately 10x faster prototyping:
- Rapid MVP development
- Faster iteration cycles
- Reduced time from idea to testable product
Shifted Bottlenecks
When implementation accelerates 10x, new bottlenecks emerge:
- Product management decisions - What to build becomes more important than how fast you can build
- User feedback collection - Getting real user input becomes the limiting factor
- Idea validation - Testing concrete ideas matters more than coding speed
Best Practices for Speed
- Start with concrete ideas - Vague concepts slow execution; specific features accelerate it
- Use AI coding assistants - Leverage tools for rapid prototyping
- Iterate in 2-3 month cycles - Best practices change frequently; stay current
- Focus on customer conversations - Speed includes talking to users, not just writing code
- Design features and pricing early - AI Fund teams do this alongside coding
Evaluating AI Startup Opportunities
Application Layer Focus
When evaluating where to build:
| Layer | Opportunity Size | Competition | Barrier |
|---|---|---|---|
| Applications | Largest by definition | Growing | Domain expertise |
| Orchestration | Medium, emerging | Moderate | Technical depth |
| Foundation Models | Large but concentrated | Intense | Capital-intensive |
Opportunity Identification Framework
- Find existing workflows that can be improved with agentic AI
- Identify domains where iterative reasoning adds significant value
- Look for tasks currently requiring expensive human expertise
- Target applications that can generate revenue to justify AI costs
Validated Use Cases from AI Fund
Domains where agentic workflows have proven valuable:
- Compliance document processing
- Medical diagnosis support
- Legal document reasoning
- Any domain requiring research → synthesis → revision
Quick Reference: Building Faster
Before Starting a Project
- Define concrete, specific features (not vague concepts)
- Identify the workflow to implement or improve
- Determine if agentic approach is needed (complex reasoning, research, revision)
- Set up AI coding assistant for rapid prototyping
During Development
- Use iterative agentic loops for quality-critical outputs
- Prototype quickly, then gather user feedback
- Let product decisions (not coding speed) drive priorities
- Revisit approach every 2-3 months as tools improve
Evaluating Progress
- Measure execution velocity, not just output quality
- Track time from idea to testable prototype
- Monitor where bottlenecks actually occur
- Adjust based on real user feedback, not assumptions