Generative AI: Get a Grip … and a Look Under the Hood
summary
Generative models rely on injected noise to create new outputs and the breakthrough is Transformer (2017) "Attention is All You Need", not vague hype. Training uses human-feedback RL with thousands of labelers, costs hundreds of millions (uploader notes OpenAI running costs ~7B, revenue ~3B, loss ~4B). Practical experiments run local 12B models (≈8 GB RAM but slow), try retrieval-augmented generation + vector DBs for recent docs—summaries work reliably, deeper synthesis and novel coding solutions remain hit-or-miss.
AI-written summary · about
chapters (18)
- 0:00Introduction
- 1:11What is generative AI
- 3:41Noise to new output
- 5:30Training LLMs and embeddings
- 8:40Transformer architecture and human feedback reinforcement learning
- 10:25Costs and challenges
- 11:51Can AI replace programmers?
- 13:40Retrieval augmented AI
- 15:26Testing local RAG systems
- 16:30Investments and energy limits
- 17:56Future models and reasoning
- 18:41Use cases and profitability
- 20:30AI for coding and content
- 21:15Q&A: Agent AI trends
- 22:47Q&A: Multi-source challenges
- 23:29Q&A: Risks of AI feedback
- 24:16Q&A: LLM pricing
- 24:59Q&A: Production projects
related talks
22mGenerative AI & its Transformation of Business, Research, and Society
17mGenerative AI: Keeping it real
19mLet's Talk GenAI
47mThe Rise of AI: A Reality Check on Generative AI
45mHow Generative AI Works: LLMs and RAG Under the Hood
53mGenerative AI: Overview and use cases
27mGenerative AI: What's Next?
42mFrom theory to practice: The rise of Generative AI