Chapter 2: The AI-First Mindset -- Resources¶
Curated resources for deeper exploration of topics covered in this chapter.
Frameworks from This Chapter¶
- The 7 Mental Models of AI-First Thinking -- Agent-first design, probabilistic thinking, data as product, permission spectrum, compound iteration, build vs buy inversion, and human-AI collaboration.
- Probabilistic AI -- Designing for uncertainty as a feature rather than a bug.
- Build vs Buy Calculus -- MIT's "Buy, Boost, or Build" framework and the three questions that determine your path.
- Human-AI Collaboration -- MIT's finding that human-AI combinations don't automatically outperform, and the three conditions where collaboration succeeds.
Tools & Platforms¶
- Anthropic Model Context Protocol (MCP) -- Standard for agent-tool communication; automatically connects, retrieves tools, and executes without human intervention.
- GitHub Copilot -- AI coding assistant demonstrating probabilistic confidence filtering; 78% task completion rate; ~30% suggestion acceptance rate; 55% faster task completion.
- Cursor -- AI coding tool implementing explicit PermissionOptions with allowlists, denylists, and "YOLO mode" for maximum speed.
- Replit -- Offers "Max Autonomy" mode with 200 minutes of continuous AI operation versus standard mode with frequent checkpoints.
- Vercel v0 -- Iterates on prompts "almost daily" using automated evaluations; each edge case becomes a test case.
- Braintrust -- AI evaluation platform; crystallized iteration into three interconnected loops (Exploration, Evaluation, Execution).
- Autobound -- Achieved 20x acceleration in LLM iteration cycle using live customer data testing.
- Adobe Firefly -- Creative AI tool illustrating the "creative director" model of human-AI collaboration.
- Grammarly -- Shows multiple suggestions instead of one "right answer," targeting 95% user-generated accuracy.
Further Reading¶
- MIT Sloan: Successful AI Implementation -- Build, Buy, or Boost -- The framework showing 65% of total software costs occur after deployment; Buy, Boost, or Build decision model.
- Klarna AI Assistant Handles Two-Thirds of Customer Service Chats -- 2.3 million conversations, 81% reduction in resolution time, projected $40M savings.
- Sequoia Capital Podcast: Training Data -- Dust -- Dust co-founder Stanislas Polu (ex-OpenAI) on why one AI model will never rule them all.
- Shopify AI-First Hiring Policy -- CEO Tobi Lutke's mandatory AI policy: demonstrate why AI cannot handle the task before requesting headcount.
- Figma Design Survey 2025 -- 84% of designers collaborate with developers weekly as AI handles execution; fewer than half feel AI makes them better.
- How We Develop v0 -- Vercel's evaluation-driven development with code-based, human, and LLM-based grading.
- Stripe Payments Intelligence Suite -- Stripe's fraud detection improvement from 59% to 97% through compounding capability investment.
Research & Data¶
- Human-AI Collaboration Review (Nature Human Behaviour) -- MIT Center for Collective Intelligence meta-analysis of 100+ studies finding human-AI combinations don't automatically outperform the best single performer.
- Carnegie Mellon COHUMAIN Framework -- Framework for treating AI as "partner under direction, not teammate."
- Postman State of API Report 2025 -- 82% of organizations adopted API-first strategies.
- McKinsey Agentic AI Survey 2024-2025 -- 23% of organizations scaling agentic AI systems, 39% actively experimenting.
- BCG: How Agentic AI is Transforming Enterprise Platforms -- One-third of global IT leaders have implemented agentic AI systems.
- MarketsandMarkets AI Agents Market Report -- AI agent spending projected to reach $50B by 2030.
- Bloomberg Engineering: BloombergGPT -- $3.5-8M training investment for a 50B parameter financial model.
- Data-Centric AI Movement -- Research consensus that marginal returns from model tweaking are diminishing; data improvement is where gains are.
Community & Learning¶
- Canva Magic Studio -- 220+ million monthly users; Magic Write feature with 8 billion uses since launch, demonstrating compound iteration at scale.
- Duolingo Engineering Blog -- Birdbrain AI processes 1.25 billion exercises daily with real-time feedback loops.
- Spotify Engineering -- Processes 500 billion daily events from 678 million users; dedicated "Algorithmic Responsibility" research team.
Companies Referenced in This Chapter¶
| Company | Key Contribution | Data Point |
|---|---|---|
| Harvey | Capability thinking | $100M ARR; legal reasoning as platform capability |
| Dust | Multi-agent infrastructure | 80,000 agents, 12 million conversations in 2025 |
| Glean | Capability pivot | Evolved from enterprise search to agentic reasoning |
| Stripe | Data flywheel | Fraud detection improved from 59% to 97% |
| Klarna | Buy path | $40M savings using OpenAI for customer service |
| Bloomberg | Build path | $3.5-8M training BloombergGPT for data isolation |
| Morgan Stanley | Boost path | GPT-4 + 70,000 proprietary documents; 98% advisor adoption |
| NVIDIA | Data flywheel | Decades of chip design data powering AI-assisted GPU design |
| Spotify | Data as product | 500 billion daily events; algorithmic responsibility team |
| Duolingo | Fast feedback loops | 1.25 billion exercises daily with real-time learning |
| Shopify | Augmentation mindset | AI-first hiring policy; demonstrate AI can't do task before hiring |