PODCAST

NVIDIA Generative AI

From language models to cameras, cars, and robots

What is actually inside the phrase generative AI, and why does every answer end up being a cost decision? NVIDIA Generative AI is a twenty-three-episode series from Cloudadorn Academy for people who want the real picture and are not already specialists.

Shelley asks the questions a curious adult would actually ask. Rob answers with something you can picture first and the name of the idea second: nesting dolls for the AI hierarchy, a group chat for attention, sand and statues for diffusion, a case conference for sensor fusion. Season one walks four acts. What these things are, from the hierarchy and the learning paradigms through attention, training, and the architecture zoo. How you make a model yours for the least money that works, from prompting to retrieval to LoRA to a full fine-tune, and which number tells you it worked. How a model gets eyes and ears, through shared embedding spaces, patches as tokens, diffusion, segmentation, and sensor fusion. Then the stack itself: precision formats and memory bandwidth, serving and packaging, the training pipeline, agents, guardrails, robots and world models, and why the ecosystem is hard to leave.

NVIDIA is the through-line because it sells at every one of those layers, so its product map doubles as a map of the field. Where a vendor number cannot be confirmed against a primary source, the hosts hedge it or drop it and say which one they did.

23 episodes · 5h 55m · Hosted by Shelley and Rob

Nothing plays on this site.

Episodes

  1. 01 Nesting Dolls: What Generative AI Actually Contains Runtime 16:25
  2. 02 Attention: How a Model Reads a Sentence Runtime 16:18
  3. 03 How a Model Learns: Forward Pass, Loss, and Backpropagation Runtime 16:14
  4. 04 When Training Goes Wrong: Overfitting, Bad Data, and Metrics That Lie Runtime 17:05
  5. 05 The Architecture Zoo: CNN, RNN, Transformer, GAN, VAE, Diffusion Runtime 14:25
  6. 06 The Cheapest Thing That Works: Prompting Before Fine-Tuning Runtime 15:21
  7. 07 Teaching an Old Model New Tricks: Transfer Learning, LoRA, and RLHF Runtime 15:48
  8. 08 Open Book: RAG Done Honestly Runtime 15:23
  9. 09 Which Number Means Good: BLEU, ROUGE, Perplexity, FID Runtime 15:44
  10. 10 More Than Text: Modalities, CLIP, and One Shared Space Runtime 14:36
  11. 11 Patches Are Tokens: The Vision Transformer Runtime 15:29
  12. 12 Sand and Statues: How Diffusion Makes a Picture Runtime 15:17
  13. 13 Pixel-Perfect: U-Net and the Shape of a Medical Problem Runtime 15:01
  14. 14 Fusion: Eyes, Ears, and Radar Runtime 15:37
  15. 15 Bolting an Eye onto a Language Model: How VLMs Are Built Runtime 14:37
  16. 16 The Metal: Cores, Bandwidth, Precision, and the Compiler Runtime 15:01
  17. 17 Serving It: Dynamo, Dynamo-Triton, and NIM Runtime 15:28
  18. 18 The Kitchen: NeMo from Raw Data to a Trained Model Runtime 14:59
  19. 19 Agents, and the Price of Thinking Longer Runtime 14:48
  20. 20 Trust: Bias, Guardrails, and Red Teams Runtime 15:29
  21. 21 Physical AI: Robots, World Models, and Simulated Worlds Runtime 15:35
  22. 22 The Flywheel: Why NVIDIA Is Hard to Leave Runtime 15:33
  23. 23 What Moved: The Words This Season Earned Runtime 14:28