AI Engineering Anti-patterns
summary
Talk lists concrete anti‑patterns and shows what actually fails: claims of “30% energy savings” collapse on user behavior and simpler convex optimization would solve costlier parts; retrieval needs careful chunking and metadata enrichment and caches leak PII; routers and scorers are on‑premise classical models, not fancy LLMs; adopting agents or vector DBs too early hides complexity; demos hit “80% in a weekend” then stall in production; AI judges drift so regular human evaluation with clear guidelines stays essential.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:06AI vs ML engineering
- 3:58Traditional ML process
- 5:22AI process and org shifts
- 7:53Prerequisites for AI engineering
- 9:38Hybrid AI applications
- 10:40Robust AI systems with guardrails
- 13:30AI in modern software engineering
- 14:56Anti-pattern: AI when unnecessary
- 18:04Anti-pattern: Bad product vs bad AI
- 21:56Anti-pattern: Overcomplicating RAG
- 24:24Anti-pattern: Agentic frameworks too early
- 27:55Anti-pattern: Solely AI judges
- 31:04Anti-pattern: Overindexing on early success
- 31:48Q&A
related talks
44mAI Engineering Anti-patterns
01h 02mArchitecture Patterns for AI-Powered Applications
43mThe Evolution of Design Patterns in the Age of AI
09mLessons Learned in AI Engineering
39mBack to Basics: Cutting Through the AI-Generated Code Noise
36mThe Best AI Strategy? Don’t Build AI 🤯 Build AI-Ready Software
39m10 Things I Hate About AI
14mHow We Turned AI Hype into Real Value