Would you rely on ChatGPT to dial 911?
summary
Claims that production systems must mix deterministic rules with probabilistic LLM outputs and shows concrete tooling: rule‑first prefilter + ML classifier used at G Guardian to scan public GitHub (thousands of commits per second), DVC + CML for slice testing, Prodigy/Streamlit for annotation, and Nemo guard‑style guardrails. Empirical notes: 13 million tokens (secrets) leaked last year; hallucinations (fake cases, bomb instructions) motivate hard rules for 911; tradeoffs are latency versus cost and false positives versus coverage.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:51Risks of unreliable AI
- 3:40Why ML fails in production
- 5:45How AI generates text
- 7:30Training vs production randomness
- 9:14Uncontrolled outputs and legal risks
- 10:17Mitigation for AI reliability
- 13:22When rules are essential
- 14:44ML and deterministic rules
- 15:49Safeguarding with Nemo Guardrails
- 17:34Testing AI with data slices
- 19:43Monitoring and rollbacks
- 20:46Secret detection approach
- 22:52Closing and Q&A
related talks
18mDefending the unknown using AI and cybersecurity
50mIs ChatGPT (and Other AI) the Enemy of Diversity?
01h 04mMachines, Learning, and Machine Learning
41mThe AI boom: what and why?
29mSurvival of the Safest: Automating LLM Defense with Genetic Programming
50mPanel: Humanity in the Loop: Balancing AI, Data Science, and Human Ingenuity in the Cybersecurity..
55mLessons Learned From Implementing Generative AI Solutions
06mInterview with Spela Poklukar