Reality is not an end-to-end prediction problem
summary
Argues that prompts alone fail for production systems and shows a concrete pipeline: use spaCy llm and Prodigy for LLM-assisted annotation, transfer-learn small task components that run in-house, and measure against a baseline; a case with S&P Global produces six megabytes components, speeds data creation 10x faster to about 15 hours, and yields higher accuracy and modular, privacy-safe deploys.
AI-written summary · about
chapters (10)
- 0:00Programming computers evolution
- 2:07Large generative models challenges
- 3:52Modular pipelines for tasks
- 5:58S&P Global case study
- 9:11Application context importance
- 10:37GitLab ticket insights
- 12:46Window knocking machine test
- 14:12Personal assistant automation
- 15:57Abstractions and product decisions
- 16:59Key takeaways for generative AI
related talks
40mBuilding new NLP solutions with spaCy and Prodigy
55mLessons Learned From Implementing Generative AI Solutions
29mAnnotating data the right way
30m🤯 No data? No problem! Synthetic data to the rescue
30mProductionizing an unsupervised machine learning model to...
26mA Practical Introduction to Productionizing NLP Models
15mChallenges in moving Gen AI apps to production
01h 16mGeneralization in natural language processing