Skip to content
Artwork for NVIDIA Generative AI
TechnologyEducationCourses

NVIDIA Generative AI

Cloudadorn Academy

Every vendor wants to sell Ana a model. Ana runs Kestrel, a company that sells shelf-scanning cameras to warehouses, and both she and Kestrel are made up. She isn't technical, so she asks her friend Shelley what to believe, and Shelley brings each brochure, quote and strange email to Rob, who explains the machinery underneath. Over twenty-three episodes they open the nesting dolls of generative AI, follow a sentence through attention, watch a model learn and fail, and climb the ladder of ways to make a model yours, from asking differently to RAG and LoRA. Then the model gets eyes and ears, through CLIP, diffusion, a hospital scan, a car braking in the drizzle and a demo that describes a crushed box. The last stretch walks NVIDIA's stack: the chips and CUDA, TensorRT, serving with Dynamo, NeMo, agents and the price of thinking longer, guardrails, robots trained in warehouses that don't exist, why NVIDIA is hard to leave, and how to check whether a product was really replaced. Where the story lives elsewhere, other invented people step in, like Joan and Denise. No technical background needed. More at cloudadorn.com.

Play
  • 20 episodes
  • Avg 8 min
  • English
  • S1 · E23
    August 25 · 6 min

    S1E23. What Moved: The Words This Season Earned

    Triton has been replaced. You are on borrowed time. That email, from a consultant offering to move Ana onto the successor for a fee, arrives with NVIDIA's own press release attached. Ana and her shelf-camera company Kestrel are a pair made up, and her box-spotting program runs on Triton. Rob checks the claim three ways: the March 2025 press release that calls Dynamo the successor, NVIDIA's product page that says Triton was renamed Dynamo Triton and works alongside the new Dynamo, and Triton's own page on GitHub, still putting out new versions in the summer of 2026. You'll understand why Dynamo is a conductor for one enormous language model spread across many machines, which isn't Ana's problem, and why the check sometimes comes back yes, as with VILA, now part of Cosmos Nemotron. More at cloudadorn.com.

    • Transcript
  • S1 · E22
    August 25 · 8 min

    S1E22. The Flywheel: Why NVIDIA Is Hard to Leave

    Ana's engineer turned down a rival's cheaper machines before she'd finished the sentence, because the machine is cheaper but the switch isn't. Ana and her shelf-camera company Kestrel are made up, though the offer is the kind real companies get. A board member keeps saying NVIDIA has a flywheel, so Rob explains the heavy disc, tests it on Amazon, then pokes the step where footage supposedly flows back to the chip company. You'll hear what a Jaguar Land Rover privacy notice and a General Motors page about Super Cruise actually say, why CUDA, with six million developers by NVIDIA's count, is the floor everyone stands on, how Tesla left NVIDIA's platform in the car yet still buys training machines, and why Google's own chips don't end the argument. Last, a consultant's email telling Ana she is on borrowed time.

    • Transcript
  • S1 · E21
    August 25 · 7 min

    S1E21. Physical AI: Robots, World Models, and Simulated Worlds

    Trained in a warehouse that doesn't exist, says the brochure for a robot arm, and Ana read the line out twice. Ana and her shelf-camera company Kestrel are made up, and she started reading brochures after a customer's top shelf gave way overnight. Rob explains why the deadline of a falling shelf makes physical AI a different problem from a chatbot, and walks through three computers: one for training, one riding on the robot, and the flight simulator in between. You'll hear why free, perfect outlines may matter more to Kestrel than the safety story, how NVIDIA's Omniverse builds a careful copy of a real place while Cosmos makes video of what happens next, and why a box must come out from behind a pillar. Then NeMo versus Nemotron, and Figure's humanoid robot on a real car plant line. Next, an engineer who turns down cheaper chips mid-sentence.

    • Transcript
  • S1 · E20
    August 25 · 8 min

    S1E20. Trust: Bias, Guardrails, and Red Teams

    Two maps of the town, laid side by side, show a credit union's loan system saying no a lot more often on the east side, though nobody ever wrote a rule about it. Denise, who runs the loan desk, and her credit union are made up, so don't go looking for either. Rob explains how a system trained on past decisions can copy an old unfairness, or be confidently wrong about people it barely met, and why deleting columns about the applicant fails while the postcode stays. Hear how the Shapley method, borrowed from splitting a prize, explains one man's no, what a model card should admit, and how NeMo Guardrails puts rails in front of a chatbot rather than inside it. Then prompt injection, why the defence is limiting what it can reach, and red teaming. Up next, a robot arm trained in a warehouse that doesn't exist.

    • Transcript
  • S1 · E19
    August 25 · 9 min

    S1E19. Agents, and the Price of Thinking Longer

    The twelfth looks like one delivery was logged twice. That note was waiting at Kestrel one morning, written overnight by an agent on trial, which had also ordered a pallet of dog food nobody asked for. Ana and her camera firm Kestrel are made up. Rob describes an agent as a chatbot allowed to go round again and touch things, and pictures a model locked in a room passing notes under the door. You'll understand MCP, the Model Context Protocol, and why doing without it doesn't make an agent worse, why you ask for the trace and a spending ceiling, and what NVIDIA's NeMo Agent Toolkit adds. Then the bill's second line, thinking mode, which NVIDIA's own estimate puts at more than a hundred times the work on a hard question, paid every time. Next, a loan system that says no far more often on one side of town.

    • Transcript
  • S1 · E18
    August 25 · 7 min

    S1E18. The Kitchen: NeMo from Raw Data to a Trained Model

    A vendor wants to build Kestrel its own model for spotting damaged boxes, trained on years of shelf photos, and keeps saying it will be built on NeMo. Ana and Kestrel are imaginary, but her two worries are real: weeks of giant computers, and a retail customer that won't let a single photo leave its warehouses. Rob pictures NeMo, NVIDIA's toolkit, as a kitchen rather than a dish, where most of the time goes on shopping and washing. You'll hear why Curator hunts near-copies of a quiet shelf photographed at three in the morning and again at four, why it scrubs personal details, when synthetic data helps, and how big training runs survive a machine that quietly hangs. Then TAO, for teaching a model that already sees, and the question that settles her retailer: where the machines are. Coming up, a pallet of dog food nobody asked for.

    • Transcript
  • S1 · E17
    August 25 · 8 min

    S1E17. Serving It: Dynamo, Dynamo-Triton, and NIM

    A queue of people waiting, and the chip sitting idle at the same moment. That's what Ana's dashboard showed when every warehouse opened and her team's new assistant crawled. Ana and Kestrel exist only because Shelley and Rob made them up, but the mistake is common: a chip built to do an enormous number of small things at once was handed one request at a time. Rob explains what a serving layer is and isn't, why a compiler like TensorRT comes first, and how dynamic batching works like a lift holding its doors. Learn how Dynamo Triton differs from Dynamo, which splits reading a question from writing the answer, and what goes inside an NVIDIA NIM package, with a sober look at NVIDIA's own comparison of about twice as much per second. Next, a vendor pitching Ana a model of her own, trained on years of shelf photos.

    • Transcript
  • S1 · E16
    August 25 · 8 min

    S1E16. The Metal: Cores, Bandwidth, Precision, and the Compiler

    A four-bit model, compiled with TensorRT, running on a Jetson in every aisle. Ana understood about every fourth word of that quote, and the other quote she got sends every photo to a data centre. Ana is pretend, and so is Kestrel, her shelf-camera company. Rob starts with a fence, a mile of flat boards that wants a thousand painters rather than one master, to show why a GPU suits a model's dull multiplication and why NVIDIA's CUDA is the foreman. You'll understand why the chip spends most of its life waiting at a narrow gate, which is what memory bandwidth measures, how quantization stores the same numbers more roughly, and why ONNX is a format while TensorRT is a compiler tied to one chip. Then running at the edge, and the one aisle Ana signs for first. Up next, a queue of people waiting while the chip sits idle.

    • Transcript
  • S1 · E15
    August 25 · 8 min

    S1E15. Bolting an Eye onto a Language Model: How VLMs Are Built

    Crushed at one corner, label facing the wall. That sentence came back from a single shelf photo at a demo Ana sat through, and now she wonders whether to rip out her cameras. Ana and her shelf-camera company Kestrel are made up. Rob explains that a vision language model like the demo is usually two existing experts, an eye and a talker, bolted together with a small adaptor called the projector. You'll hear how the 2023 paper Visual Instruction Tuning trained it with both ends nailed down, why a paper titled Vision language models are blind caught them failing on circles that touch, why before and after across a video is weaker still, and how NVIDIA's NVILA scales then compresses to fit places like cameras and robots. Ana's second demo settles the counting question. Next, a vendor's quote in which Ana understood about one word in four.

    • Transcript
  • S1 · E14
    August 25 · 9 min

    S1E14. Fusion: Eyes, Ears, and Radar

    At dusk, in drizzle, a friend's car braked hard before Shelley had seen anything, and then a van shot out of a side road. The story is made up, though an everyday one, and now she's buying a car while one salesman says cameras are all it needs and another praises the laser in the roof. Rob sets out what radar, lidar and cameras are each blind to. A hospital with three specialists explains early fusion and late fusion, and a fire drill explains hiding one sense at random in training. Hear about a 2020 study where programs that watched and listened did worse than one sense alone, where Tesla, Waymo and NVIDIA's sensor kit stand, and what the US National Transportation Safety Board report on the May 2016 Florida crash actually says. Coming up, one shelf photo described back in a sentence.

    • Transcript
  • S1 · E13
    August 25 · 9 min

    S1E13. Pixel-Perfect: U-Net and the Shape of a Medical Problem

    Joan, a friend of Shelley's made up, went back for a second scan, and her doctor put the two pictures side by side, each with a thin coloured line drawn round the thing they're watching. Joan wants to know who drew that line, and how sure they are. Rob explains why asking a computer where is so much harder than asking what, which is the job called segmentation. Shrinking a picture teaches the program what it's seeing and wipes out where, so the fix is to keep the drafts: skip connections that hand each copy straight across, drawing the capital U that names the U-Net. You'll also hear how federated learning lets hospitals train together while the scans stay put, with NVIDIA's tool for it, and why a 2024 re-test found the U-Net still did best. Next, a car that brakes before anyone has seen the van.

    • Transcript
  • S1 · E12
    August 25 · 9 min

    S1E12. Sand and Statues: How Diffusion Makes a Picture

    No skylights, no rain, and that camera has never been photographed at night, yet there it is in a gorgeous brochure picture. Kestrel's marketing lead typed one sentence into an image generator, and Kestrel, its boss Ana and he are all made up. Rob explains that the picture started as pure static, then pictures a statue buried by a sandstorm and software trained only to guess how much sand is on each frame. That is how diffusion models make something from nothing. Learn why the sentence is fed in at every stroke, why pushing the guidance scale all the way up made everything garish, what a negative prompt really does, why the seed is the number to write down, and how shrinking the picture leaves sixty-four times fewer places to work on. Up next, a friend's hospital scan and the thin coloured line someone drew round it.

    • Transcript
  • S1 · E11
    August 25 · 8 min

    S1E11. Patches Are Tokens: The Vision Transformer

    A vendor's brochure says its new camera model reads photos the way chatbots read sentences, and that Ana's current one is a generation behind. Ana and her shelf-scanning camera company, Kestrel, are both made up. Rob starts with the CNN, the magnifying-glass kind that sees local detail first, and why it struggles to check a box in the middle against the shelf label along the bottom edge. Then the Vision Transformer, which cuts the photo into squares like a bar of chocolate, turns each into the same kind of list of numbers a chatbot makes from words, and lets every square look at every other. You'll understand patches and tokens, why built-in assumptions win on a small pile of photos, what Touvron and colleagues showed on a single computer, and why smaller squares cost very roughly sixteen times the work. Next, a gorgeous picture of a warehouse that never existed.

    • Transcript
  • S1 · E10
    August 25 · 8 min

    S1E10. More Than Text: Modalities, CLIP, and One Shared Space

    Oh, brilliant, said a warehouse manager on a support call, in the flattest voice you've ever heard, after a camera marked his whole aisle empty again. The summary said the customer was very pleased. He'd rung Kestrel, the shelf-camera company Ana runs, and Shelley and Rob made up both of them. Rob explains that the tone died when the call became a transcript, and what multimodal really means: one model holding the voice and the words together, not two models passing a note. Then CLIP, trained on four hundred million picture-and-caption pairs from the web, which lets Ana write a sentence like a photo of a crushed pallet instead of retraining her cameras. You'll also learn what CLIP can't do, from counting boxes to word order, and why its original release isn't for shipping. Coming up, a camera that reads photos the way chatbots read sentences.

    • Transcript
  • S1 · E9
    August 25 · 8 min

    S1E09. Which Number Means Good: BLEU, ROUGE, Perplexity, FID

    Clear winner, in red, with a circle round one vendor's name. Kestrel's operations lead made the sheet, comparing two tools that turn warehouse staff's voice notes into damage reports and summaries. Kestrel, Ana and everyone who works there are made up. The column he read, word error rate, counts mistakes, so the smaller number wins. Rob asks two things of every score: what it counts, and whether bigger is good or bad. Hear why a trial recorded in a quiet office says little about a floor of reversing forklifts, and how ROUGE rewards covering the points but marks down a good rephrasing. Then the trick of having another AI model mark the summaries, which agreed with people more than eight times in ten in the best-known study but can flip with the order and favours its own writing. Next, a sarcastic support call filed as a happy one.

    • Transcript
  • S1 · E8
    August 25 · 11 min

    S1E08. Open Book: RAG Done Honestly

    Asked whether the bracket warranty covers a camera in the freezer aisle, Kestrel's support assistant said fully covered, quoting the manual. One line further down, the manual says it doesn't apply in freezer aisles. Ana and Kestrel are imaginary, and her team wants a bigger model. Rob says hold the money, because retrieval augmented generation, or RAG, is two machines bolted together, one that fetches pages and one that writes. You'll understand embeddings as addresses for meaning, how chunking can cut an exception off its rule, and the silent trap of mixing embedding models. The papers Fine-Tuning or Retrieval and Lost in the Middle show why looking things up reduces hallucination without ending it, and two checks, faithfulness and whether the right pieces came back near the top, point at which machine failed. Up next, a comparison sheet with a red circle round the wrong vendor.

    • Transcript
  • S1 · E7
    August 25 · 10 min

    S1E07. Teaching an Old Model New Tricks: Transfer Learning, LoRA, and RLHF

    Please complete section four of the damage report form. That was the whole reply Kestrel's support assistant gave when a warehouse manager asked for a short list for his night shift. Ana and her camera company are made up, and a vendor had just retrained an entire model on everything Kestrel ever wrote. Rob uses hiring for a café to explain why starting from a lab's model makes sense, then decodes the delivery note: fine tuning, RLHF learning from people's clicks, as in InstructGPT, and a simpler method that skips the judge. Learn about catastrophic forgetting, why whole-model retraining needs so much memory, and how LoRA leaves the base model frozen and puts the change in a small file. A study called Fine-Tuning or Retrieval explains why the catalogue still came out wrong. Next, an assistant that looks things up and still gets the freezer aisle wrong.

    • Transcript
  • S1 · E6
    August 25 · 8 min

    S1E06. The Cheapest Thing That Works: Prompting Before Fine-Tuning

    A vendor has quoted Ana months of work and a number with a lot of zeros to fine-tune a model on every document Kestrel has ever written. Ana and Kestrel, her warehouse camera firm, are as imaginary as each other, but her three wishes for a support assistant are ordinary: answer questions about a returns policy that keeps changing, write in Kestrel's friendly, dry voice, and fill in a ticket form exactly. Rob lays out a ladder of ways to make a general model do your job, in price order, and says climb only when the rung below has failed. You'll understand why the form needs a stencil, why RAG suits a policy that moves, where LoRA earns its place for the voice, and what full fine-tuning and pretraining are for. Temperature isn't training at all. Coming up, an expensive retrain that can no longer write a simple list.

    • Transcript
  • S1 · E5
    August 25 · 7 min

    S1E05. The Architecture Zoo: CNN, RNN, Transformer, GAN, VAE, Diffusion

    A cat is asleep on a warehouse shelf, and the shelf camera reported everything fine, because it was only ever taught to look for dented boxes. The camera is one of Kestrel's, the company Ana runs, and both she and Kestrel are imaginary. Rob treats the model names two vendors pitch as shapes of machine, each built for one kind of problem. Hear how a CNN slides a tiny window across a photo, how an autoencoder trained only on normal shelves finds the odd one out by rebuilding it badly, and why latent space is the recipe rather than the cake. For fake dented boxes, he weighs a GAN, including NVIDIA's StyleGAN and the risk of mode collapse, against diffusion, citing the 2021 paper Diffusion Models Beat GANs and a 2020 measurement by Song, Meng and Ermon. Next, a vendor's quote with a lot of zeros.

    • Transcript
  • S1 · E4
    August 25 · 8 min

    S1E04. When Training Goes Wrong: Overfitting, Bad Data, and Metrics That Lie

    On day two at a client's warehouse, a new checker built to spot dented boxes waved through an entire pallet of crushed ones, even though its report said it was right nearly every time. Ana and her shelf-camera company Kestrel were made up. Rob explains why accuracy hides the failure you care about when the thing you're hunting is rare, and how precision and recall point at false alarms and misses. A pair of learner drivers, one who only knows the road to work and one who quit after a lesson, separate overfitting from underfitting, and a locked drawer of photos tells them apart. You'll also hear about early stopping, a yellow damage sticker that may have given the answer away, and a 2021 study led by Curtis Northcutt that found wrong labels in ten standard test collections. Up next, a cat asleep on a warehouse shelf.

    • Transcript
Showing 1–20 of 20 episodes