3 Steps to Create Trustworthy Generative AI Products
summary
Three concrete stages for finance chatbots: a Prototype that enforces intent classification, data chunking and custom prompts plus a patent-pending quality scoring engine that checks context, math and sensibility; a Scaling phase that optimizes architecture for cost-per-query, robustness and token limits; and an Evolution phase using RLHF and fine-tuning techniques like LoRA quantization. Examples: GPT-3 (176B parameters trained on ~45 TB), compressing a 65 GB model to ~3 GB, and a 3-day hackathon that built an automation engine to summarize and prioritize ServiceNow tickets successfully.
AI-written summary · about
chapters (14)
- 0:00Introduction to trustworthy generative AI products
- 1:04What is generative AI
- 4:13LLM evolution and safety
- 7:26Working with large language models techniques
- 10:17GenAI use cases
- 12:43GenAI product roadmap
- 15:28Prompt engineering
- 21:08Choosing the right large language model
- 23:14Trustworthiness challenges
- 24:57Minimizing hallucinations
- 27:22Response quality scoring
- 29:08Reinforcement learning
- 30:56GenAI new directions
- 34:49Q&A
related talks
17mGenerative AI: Keeping it real
47mThe Rise of AI: A Reality Check on Generative AI
52mBeyond the AI Hype: What's Real, What's Next
39mBuilding Your Own Bot with GPT: Harnessing the Power of AI...
47mOperationalization of AI/ML models and use cases for ChatGPT Q&A
01h 06mAI in the Trenches: Engineering Real-World Systems
19mLet's Talk GenAI
27mSafeguarding Generative AI Applications: Navigating Security Challenges and Best Pract...