Q&A #5
summary
Argues that prompt engineering starts prototypes while fine-tuning becomes best practice in production, with fine-tunes often needing 100–250 examples but sometimes getting 90%+ from 20 examples for big cost savings. Recommends logging UX signals (copy button, thumbs up) as training labels, using LLMs to augment datasets (a Data synthesis tool), and combining automated checks with human review. Lists fallbacks: retries, default content, or routing to a live agent. Cites HelloFresh results: employee retention +22% and satisfaction +18% after automating simple tasks.
AI-written summary · about
chapters (12)
- 0:00Introduction and setup
- 1:25Prompt engineering vs fine-tuning
- 3:11Persona chatbots and models
- 5:00Fine-tuning data requirements
- 7:28Prompt engineering in enterprises
- 9:38Collaboration between engineers and linguists
- 12:31Evaluating prompts and models
- 17:07Guardrails and fallbacks in production
- 19:16Refining models in production
- 21:01Customer service automation
- 25:18Future trends in AI
- 28:12Closing remarks
related talks
39mOptimizing a Prompt for Production
27mEnhancing AI-Driven Quality Assurance
15mPractical tactics to build reliable AI apps
03mInterview
21mDesigning in the age of the 10x engineer
13mChallenges, Considerations, and Best Practices
15mPrompt Engineering 2.0
26mMastering Large Language Models: Prompt Engineering for Precision & Performance