ChatGPT is lying, how can we fix it?
summary
Claims that ChatGPT produces hallucination—confident but false facts—because pretraining stops at a cutoff (September 2021) and models only predict next tokens. Proposes Retrieval-Augmented Generation using embeddings, semantic search and a vector database (example: Qdrant) to fetch K most-relevant chunks, then pass those snippets to the LLM; notes limits (token windows, cost — ~100k tokens ≈ $1 per call), chunking, prompt engineering, verifier models and human-in-the-loop as practical mitigations.
AI-written summary · about
chapters (12)
- 0:00Introduction to LLMs
- 1:04Hallucinations in models
- 3:32Training process limitations
- 9:07Frozen knowledge cutoff
- 12:01Language tasks vs facts
- 13:26Extending context for accuracy
- 18:06Retrieval augmented models
- 22:42Building knowledge base
- 24:52Prompt engineering challenges
- 28:47Turning knowledge tasks into language tasks
- 30:55Tools and maintenance
- 32:03Q&A session
related talks
46mConversations with ChatGPT: illusion or reality?
01h 01mHow ChatGPT Works
06mInterview with Spela Poklukar
02h 00mChatGPT & Prompt Engineering
41mThe Unreliable Computer Revolution - Embracing the Double-Edged Sword of GenAI
01h 16mChatGPT - and How we Can Benefit from it
01minteresting experience with ChatGPT
15mHarnessing the Power of Generative AI for Intelligent Applications