AI Engineering Anti-patterns
summary
Generative models often solve the wrong problem: a claimed 30% electricity savings comes without a baseline check and greedy scheduling or classic optimization often matches or beats the ML approach. Real systems pair handfuls of classifiers (safety, FAQ hit, PI detector) with a LLM like Gemini or clot behind retrievalaugmented generation; performance jumps come from preprocessing and Q&A-style chunking not swapping vector DBs; evaluation needs humans because AI-as-judge is non-deterministic and brittle.
AI-written summary · about
chapters (11)
- 0:00AI engineering anti-patterns intro
- 1:07AI vs ML engineering
- 10:05Chatbot system architecture
- 15:00GenAI when unnecessary
- 17:29Bad product vs bad AI
- 19:37Overcomplicating system design
- 24:21Skipping human evaluation
- 28:10Evaluating with AI judges
- 36:58Overindexing on early success
- 40:09Skills for AI engineers
- 40:51Q&A
related talks
32mAI Engineering Anti-patterns
01h 00mAI: It AI-n't what you think!
25mThe (mis)use of GenAI in Data Analytics
43mAI Engineer Ethics: Shaping Responsible Generative AI
49mAn AI with an Agenda
39mBack to Basics: Cutting Through the AI-Generated Code Noise
41mGenerative AI: bliss or miss?
55mAI Showdown: The Good, The Glitchy, and The Risky