How Generative AI Works: LLMs and RAG Under the Hood
summary
Describes how large language models use Transformer decoder stacks (about a dozen repeated layers) with tokenization, positional encodings and masked self-attention to predict next tokens, giving concrete scale numbers (e.g., GPT‑3 ≈175B parameters, ~250B training tokens, very high GPU costs and GPT‑4 costing tens–hundreds of millions). Explains Retrieval‑Augmented Generation (RAG) using semantic embeddings and nearest‑neighbor retrieval (instead of BM25), prompt augmentation and ranking plus RLHF/fine‑tuning to reduce hallucinations, and outlines image generation via CLIP plus diffusion models (Stable Diffusion / DALL·E) that denoise from noise guided by text.
AI-written summary · about
chapters (9)
- 0:00Introduction to generative AI
- 1:03Language model history
- 8:56Transformer architecture
- 13:30Tokenization and embeddings
- 27:17LLM limitations
- 30:50RAG: how it works
- 36:03RAG use cases and trends
- 38:32AI image generation
- 44:54Conclusion & Q&A
related talks
47mThe Rise of AI: A Reality Check on Generative AI
26mGenerative AI: Get a Grip … and a Look Under the Hood
53mGenerative AI: Overview and use cases
52mUnleash the Power of Large Language Models (LLMs)
43mBeyond the Hype: A Realistic Look at Large Language Models
32mHow generative AI is Changing the way people interact with computers
01h 00mInside GPT – Large Language Models Demystified
17mUnlock the power of Generative AI: From understanding its principles to utilizing its power