We are the Guardians of our Future
summary
Shows concrete failures and fixes: a model claims 90% accuracy at detecting criminals using logistic regression/KNN and a biased CalGang ground truth that even lists infants; benchmarks like Labeled Faces (≈13k) and ImageNet (14M) skew results; LLaion‑5B (five billion) requires “781 years” to audit and overrepresents Pinterest/Shopify images; OpenAI CLIP and Google CLD3 choices bias selection; RAG systems compound errors to ~72% reliability, cause real harms (Air Canada payout), and remain vulnerable to prompt injection and hallucination.
AI-written summary · about
chapters (12)
- 0:00Introduction to AI risks
- 1:11Engineering responsibility
- 4:51Data bias and leaky markers
- 9:32Influential data benchmarks
- 16:14Generative AI and LAION-5B
- 26:09Embeddings and bias
- 33:12Automating historical bias
- 37:16Hallucination and reliability
- 40:44Probabilistic failure in RAG
- 43:07Prompt injection security
- 46:32Accountability and tooling
- 50:29Closing and Q&A
related talks
51mWe are the Guardians of our Future
34mDespicable machines: how computers can be assholes
42mIs your AI plotting to kill you? Why AI and Machine Learning is dangerous
38mHow to Test a learning system
34mNeural networks: insane in the membrane
29mResponsible AI Challenges and Recommendations
01h 03mThe Ethical Dilemma of AI: Save the World or Forsake It?
30mHow to Lie with AI: Understanding Bias, Ethics, and the Hidden Risks in ML