Skip to content

Chapter 12: Staying Ahead — Modularity and What's Next — Resources

Curated resources for deeper exploration of topics covered in this chapter.

Frameworks from This Chapter

  • 10 Principles of AI-First — Ten enduring principles that transcend specific technologies: build for agents, routing is strategy, give AI superpowers with guardrails, own your domain share your foundation, build to add not to replace, and more.

Tools & Platforms

  • Vercel AI SDK — Model abstraction layer with a single generateText call working identically regardless of provider; switching providers means changing a model string parameter (referenced in Section 2: Building for Evolution)
  • Stripe API Versioning — 13-year backwards compatibility record; date-based versions with automatic pinning; transformation modules walk responses "back in time" (referenced in Sections 2 and 4)
  • AWS SageMaker — Launched 2017 with modular components; models can register with Bedrock through simple UI workflow while maintaining existing infrastructure (referenced in Sections 2 and 5)
  • AWS Bedrock — Arrived 2023 as a parallel service to SageMaker, not a replacement; demonstrates addition-without-replacement pattern (referenced in Sections 2 and 5)
  • Amazon Q — 2024 addition creating integration points rather than forced migrations (referenced in Section 5: Amazon and Tesla Examples)
  • Jasper AI — Model-agnostic AI engine routing different content types to optimal models without touching application logic (referenced in Section 1: Why Modularity Matters)
  • Intercom Fin AI Agent — Expanded from chat to email support by adding a new component; processed over 1M emails in first month with AI providing answers to 81% of conversations (referenced in Section 1: Why Modularity Matters)
  • Microsoft Azure Shadow Mode — New models process requests but don't serve responses, logging predictions for offline comparison before gradual traffic rollout (referenced in Section 2: Building for Evolution)
  • Thoughtworks Technology Radar — Published biannually since 2010; April 2025 edition featured 48 AI-related items; four-ring framework: Adopt, Trial, Assess, Hold (referenced in Section 3: Monitoring Emerging Tech)
  • LangChain — Referenced as a cautionary example; companies reported abandoning it when abstractions became limiting; teams that kept usage isolated to specific modules survived (referenced in Section 6: What's Next)
  • Tesla FSD Shadow Mode — Runs silently on every vehicle making hypothetical decisions and comparing to driver choices; disagreements become training data across 2M+ vehicles (referenced in Section 5: Amazon and Tesla Examples)
  • DeepSeek R1 — Placed in "Assess" ring of Technology Radar despite enormous hype; technical innovation evaluated separately from headlines (referenced in Section 3: Monitoring Emerging Tech)

Further Reading

  • Netflix Deployment Architecture — 4,000+ daily deploys across 200+ independent microservices; demonstrates modular velocity at scale
  • Uber's Domain-Oriented Architecture — Feature integration time dropped from 3 days to 3 hours; training speed improved 1.5-4x; 100,000+ deployments per week across thousands of services
  • Stripe Backwards Compatibility — All internal code runs on latest version; transformation modules convert responses for each customer's pinned version; nearly 100 breaking changes absorbed
  • Salesforce + AWS Bedrock Integration — Used strangler fig pattern to integrate Bedrock Custom Model Import while maintaining existing SageMaker endpoints; 30% faster deployments, 40% cost savings
  • Tesla FSD v11 to v12 Transformation — Collapsed 300,000 lines of C++ for decision-making into 2,000-3,000 lines of end-to-end neural networks, delivered via OTA update
  • Anthropic Claude Explains Experiment — Acquired 24 websites linking to posts in one month; killed it anyway due to reputational risk from AI-generated content for a company whose credibility depends on AI accuracy
  • OpenAI Operator Deprecation — Launched January 2025 as standalone agent interface; deprecated by July 2025; capabilities integrated directly into ChatGPT after users found switching too high-friction
  • Linus Torvalds on AI Hype — "It is currently 90% marketing and 10% reality"; cautionary perspective for technology evaluation

Research & Data

  • Gartner Prediction — 40% of enterprise applications will embed agentic AI by 2026, up from less than 5% today
  • AI Inference Cost Trends — Costs decreased 10x annually 2022-2025; high-end models saw 900x reduction at GPT-4o level
  • Context Window Expansion — Claude 3.5 Sonnet at 200K tokens; Gemini 1.5 Pro supports 1-2 million tokens; enables entire codebases in single context
  • RAG vs Fine-Tuning Adoption — RAG adoption jumped from 31% to 51% in one year; fine-tuning stayed at 9% (Menlo Ventures 2024 report)
  • Enterprise Build vs Buy Shift — Enterprises went from 47% build/53% buy in 2024 to 76% buy in 2025 (Menlo VC 2025 report)
  • AI Pilot Failure Rate — 88% of AI pilots never reach production; time-boxed exploration (Janea Systems) saves $50-90K per failed experiment
  • Enterprise Microservices Consolidation — 2025 data shows enterprises consolidating microservices back into modular monoliths in some cases; not because modularity failed but because they over-architected too early
  • IBM 2026 Predictions — "Smaller reasoning models that are multimodal and easier to tune"; reasoning bifurcates from conversational
  • OpenAI Function Calling Disruption (October 2024) — Broke production systems with minimal warning; models returned function responses as regular messages; demonstrates gap between stated deprecation policy and operational reality

Community & Learning

  • Thoughtworks Technology Radar — Biannual publication tracking AI and other technology trends; provides Adopt/Trial/Assess/Hold categorization framework
  • Menlo Ventures State of GenAI in the Enterprise — Annual report tracking enterprise AI adoption, build-vs-buy trends, and RAG/fine-tuning adoption rates
  • Janea Systems Rapid Prototyping Framework — Staged approach: 2 hours (basic prototype), few days (feature-rich), 2-4 weeks (quick win sprint); each gate saves $50-90K vs traditional development
  • Strangler Fig Pattern — Named after the tropical tree; wrap, route, replace, repeat approach to gradual modernization while both old and new systems coexist
  • Feature Flags for AI — Progressive delivery pattern for AI: gradual rollouts (1% to 10% to 100%), targeting by segment, kill switches, A/B tests with statistical rigor

Companies Referenced in This Chapter

Netflix, Stripe, Tesla, Amazon (AWS), Uber, Intercom, Jasper, Salesforce, Microsoft (Azure), Vercel, Anthropic, OpenAI, Thoughtworks, Janea Systems, Yirifi, DeepSeek, IBM, Google (Gemini)