Twain: (re)building AI writing assistant
summary
Twain starts as rule-based NLP with spaCy and moves to fine-tuned BART models for detection+correction, then adopts GPT‑4 in March 2023 which immediately improves generation quality and enables GPT-4 subject-line in two days shipped to production. A hybrid architecture routes cheap detection to rules or GPT‑3, heavy generation to GPT‑4, and uses prompt-management tooling, AB testing, streaming and smart caching to cut latency and cost. The stack is built and iterated by a 10-person team with linguists using low‑code prompt tools to deploy prompts without engineers.
AI-written summary · about
chapters (9)
- 0:00Introduction to Twain
- 2:13Building an AI assistant
- 4:26Rule-based NLP foundations
- 5:49Evolution to internal models
- 6:58Integrating GPT-4
- 8:25Hybrid AI strategy
- 10:25Prompt engineering tools
- 12:51Prompt optimization loop
- 16:25Future and conclusion
related talks
45mHow to use GPT in your projects: A Case Study
25mFrom Treatise to Tech: The Development of ChatSOC, an AI Research Assistant
02h 00mChatGPT & Prompt Engineering
01h 08mIn Prompts We Trust
16mUnveiling the practical prowess of OpenAI's ChatGPT
36mUnleashing Generative AI: Improving Developer Productivity in SRE
45mWeb Development in the Age of AI
47mThe Art of Prompt Engineering: Best Practices for Effective AI Interactions