TRANSCRIPT 🔗
NVIDIA Generative AI. What Moved: The Words This Season Earned.
Hosts: Shelley and Rob. Cloudadorn Academy.
This episode is narrated using AI voice technology. The content and script are original.
Shelley: I'm Shelley, I ask the questions. Rob builds these things for a living, which is why he gets the answers.
Rob: Last one.
Shelley: Three things, then. What actually moved this season. Which words changed meaning underneath us while we were talking. And which of them I have earned the right to say out loud.
Rob: All three. And the third is the reason we did the other twenty two the way we did.
Shelley: Start with the fact that anything moved at all.
Rob: This field does something no ordinary subject does. It keeps a standing list of things that used to be the right answer and are now the wrong one, and it rewrites it every year.
Shelley: Every field has those.
Rob: Not written up front, apologising for last year's answers. That list exists because this field ships faster than it names things.
Shelley: So it's a list of renames.
Rob: That is what it looks like. It is not what it is. Every line is a decision somebody made, and the label tells you least. I will give you the decision each time.
Shelley: How many of them?
Rob: Six that matter. And I am going first with the one where my own method fails, because that is the only honest place to put it.
Shelley: Go on, then.
Rob: The thing that holds a trained model and answers when people ask it questions. It had one famous name for years.
Shelley: And now it doesn't?
Rob: Now the product page says it was renamed, and calls the old name the former one.
Shelley: Renames happen all the time.
Rob: Then in the same spring, March 2025, a second product appears with almost the same name, and the company announcement calls that one the successor to the first.
Shelley: Successor. So the first one is finished.
Rob: That is what the word means everywhere else. So I went and read the code, because code cannot do marketing.
Shelley: And what did you find?
Rob: The old project is still there under its old name, shipping releases in the summer of 2026, and the new platform word is not on the front of it.
Shelley: And the new one?
Rob: Same on the front of that project. The sentence joining them lives on the pages where you buy things.
Shelley: And that's useful to me because?
Rob: Because both of them will keep arriving in front of you for years, and you will not lose an afternoon deciding which is obsolete. Neither is.
Shelley: And the two things themselves?
Rob: Different jobs. The older one serves any model from any framework. The newer sits above the engines and coordinates one enormous language model spread across many machines.
Shelley: So it isn't a replacement.
Rob: It is not. It conducts. Siblings in one family, not two words for one thing. And now the part where my method breaks.
Shelley: Break it, then.
Rob: All season I gave you the picture first and the name second, on the grounds that a picture survives a rename and a label does not.
Shelley: It has worked.
Rob: It works until a rename splits one picture into two products. The picture here is the thing that serves the model. Both are that. The picture cannot tell you which you want.
Shelley: So here you have to learn names.
Rob: You have to learn names, and that is the real cost of the rename, paid by everybody downstream. I would rather say so than pretend the rule holds everywhere.
Shelley: Give me the second thing.
Rob: How coarsely you are allowed to store a number. The old answer was eight bits, and that was already brave.
Shelley: And the new one?
Rob: Four. And native, meaning the silicon does it directly rather than pretending to.
Shelley: Show me the two levels.
Rob: A multiplier per block, and a second multiplier over the whole thing kept at full exactness.
Shelley: And the second level earns its keep.
Rob: The second level is what keeps four bits honest, and it is why this format is not the other four bit format.
Shelley: So what actually moved.
Rob: Not the format. Where it is allowed. Four bits started as a way of shrinking a model somebody had already built. It is now used for the building as well.
Shelley: Training in four bits.
Rob: Their own pretraining runs, hundreds of billions of pieces of text, published, and the accuracy lands where the wider format does.
Shelley: And the decision underneath it?
Rob: Somebody decided the accuracy you lose by being coarse can be bought back by being careful with the multipliers. That is an engineering bet, not a name.
Shelley: And the third?
Rob: The flagship. Two generations ago it was one chip everybody named.
Shelley: And now it isn't?
Rob: That chip is the old one. The current family is shipping, and the one after it went into full production in 2026.
Shelley: So the roadmap is faster.
Rob: The beat is yearly now and they say so. That is the whole line.
Shelley: And the fourth?
Rob: The word that used to cover everything they made. You would say it and mean the tools, the models, all of it.
Shelley: And now it's only the tools?
Rob: Now it is only the tools. The models were given their own names, and their open families page names five. The reasoning and agent one. The world models. The humanoid one. The driving one. And the biology one.
Shelley: And people still say the old word for a model.
Rob: Constantly. It is the commonest mistake in this material, and the naming created it. They called the model family something that rhymes with the framework.
Shelley: Say the rule again.
Rob: The framework is what you do to the thing. The families are the thing. If somebody asks you to name a model and you answer with the framework, you have named a kitchen when they asked for a dish.
Shelley: And the seeing ones?
Rob: Three separate research lines with three separate names, all folded into one family in January 2025. Their own repository carries the date and the sentence.
Shelley: Which is a rename that does help?
Rob: It is the good
Load full transcript
kind. Three names, one picture, and now one name. That is a rename collapsing a mess rather than making one.
Shelley: And the fifth?
Rob: What a large model is assumed to be inside. The old assumption was that every part of it works on every word.
Shelley: And what is it now?
Rob: Now it is a crowd of specialists and a router picking a few per word. Mixed with layers from the state space family that read long text cheaply.
Shelley: And that's normal now?
Rob: Their current large open models are built that way and the papers are public. I will say that rather than that the industry moved, because I have no page measuring the industry.
Shelley: And the picture making?
Rob: The same shift, one field over. The shape drawn like the letter U was the default backbone for years. At the front of the field it has largely gone, replaced by the block stack that reads sentences.
Shelley: And the reason given?
Rob: The original paper says it improves more predictably as you make it bigger. Not that it is cleverer. That improvement is easier to buy.
Shelley: Both of those are the same decision.
Rob: Both are the same decision. Predictable scaling beat the clever shape. Twice, in two different corners of the field.
Shelley: And the sixth?
Rob: Two shifts and they are the same shift. The first is the word for the whole area. It used to mean two or three kinds of input bolted together.
Shelley: And now one backbone?
Rob: One backbone handling words, pictures, video, sound and action, for understanding and for making.
Shelley: And what moved since.
Rob: Their world model family now ships as exactly that. One model, rather than seeing and making sitting apart.
Shelley: And the second shift?
Rob: Acting in the world went from a paragraph near the back to the middle of the page. The humanoid family, the driving family and the world models are named product lines now, not research directions.
Shelley: And underneath both?
Rob: Attention moved. It used to sit on chat and on looking things up. It sits now on things that hold a goal and use tools, and on paying a model to think longer.
Shelley: New thing. Why does this subject come in halves?
Rob: Because that is how the field is sliced, and the halves are not equal. One covers language, making a model yours, and the stack it runs on. The other covers eyes, ears, combining them, and making pictures.
Shelley: And the second assumes the first?
Rob: It has to. It skips the ground the first half covers and goes straight to what is unique to it.
Shelley: So there's an order.
Rob: There is, and it is the order we walked. What these things are. How you make one yours for the least money that works. Then more than text. Then the stack, and the world it touches.
Shelley: Was that your order or the field's?
Rob: The field's, mostly, and I will own that. We rearranged inside the acts. The spine was already there.
Shelley: And if I only ever learn one half?
Rob: The first. Not because the second is easier. Because the second is written on top of the first and will not make sense underneath it.
Shelley: Then hand them over. The third thing I'm owed.
Rob: Fast, and only the ones this season actually paid for. If we did not build the picture, you do not get the word.
Shelley: Go on, then.
Rob: The dolls. Artificial intelligence outside, machine learning inside it, deep learning inside that, generative inside that. Trained wide and reusable is a foundation model.
Shelley: The group chat.
Rob: Attention. Each word deciding how much every other word matters to it. Splitting text into pieces is tokenising. The number list carrying meaning is an embedding. Order is injected on purpose, because attention cannot see it.
Shelley: The whisper chain.
Rob: Forward pass, loss, then working backwards to find the share of the blame carried by each weight. That last part is backpropagation.
Shelley: And the dials?
Rob: What you choose before training is a hyperparameter. What the model works out is a parameter. One pass over everything is an epoch.
Shelley: Memorising the wrong thing.
Rob: Overfitting. Its opposite is underfitting, and the fix lists differ, which is the only reason to know both.
Shelley: Two models fighting.
Rob: Two set against each other is a GAN, and only a GAN can collapse into making one thing forever.
Shelley: Squeeze and rebuild.
Rob: An autoencoder. The version with a tidy middle you can wander around in is the VAE, and that middle is a latent space.
Shelley: Now do the ladder.
Rob: Ask well. Then look it up before answering, which is RAG. Then train a small piece, which is PEFT, and the famous small piece is LoRA.
Shelley: And the top rung?
Rob: Train the whole thing. Climb only when the rung below genuinely cannot do the job.
Shelley: And finishing school?
Rob: Training on human preference rather than on right answers. RLHF is the first of those and there are tidier ones since.
Shelley: Making it smaller.
Rob: Four levers and they are not synonyms. Coarser numbers is quantisation. Fewer parts is pruning. A small model copying a large one is distillation.
Shelley: And the fourth?
Rob: Fine tuning, which teaches skill and style, not facts.
Shelley: The numbers that grade it.
Rob: BLEU for translation, ROUGE for summaries. Perplexity for how surprised the model is, and lower is better. FID for whether pictures look right, and lower is better. WER for speech, lower again.
Shelley: Now the eyes.
Rob: Two encoders trained to agree, sharing one space, is CLIP. Cutting a picture into squares and feeding them in like words is ViT.
Shelley: And making one?
Rob: Adding noise then learning to take it away is diffusion. The old backbone was drawn like the letter U. The new one is DiT.
Shelley: Bolting them together.
Rob: A vision encoder, a small translating layer, and a language model behind it. The small layer is the projector and it is the quiet hero. Add actions to the end of that and you get VLA.
Shelley: Now the metal.
Rob: The programming layer everybody builds on is
CUDA. A format for moving models between tools is ONNX. The compiler that makes one fast on one chip is a different thing from the format, and people mix those two up constantly.
Shelley: And serving it?
Rob: Packaging a model with everything it needs to run behind a standard door is a NIM. Grouping arriving requests is dynamic batching. Splitting the reading stage from the writing stage across different machines is disaggregated serving.
Shelley: Then the agents.
Rob: A goal, a set of tools and a loop. One open standard for finding and calling tools is MCP, and MCP is not the only way, which is the most misused set of letters in the field.
Shelley: And then trust?
Rob: Rules at the edge of the application rather than inside the weights are guardrails. They are the answer whenever somebody says without retraining.
Shelley: And the attacks?
Rob: Attacking your own system on purpose is red teaming. Smuggling instructions in through the input is prompt injection.
Shelley: And the loop?
Rob: A loop where each turn makes the next turn cheaper is a flywheel. Anything else drawn in a circle is a circle.
Shelley: You skipped some. On purpose.
Rob: On purpose. The full list of words in this field runs to seventy or so.
Shelley: So how many did you give me?
Rob: About half.
Shelley: And the half you kept?
Rob: Names for things you already understand, where the word buys you nothing until you are in a room with somebody using it.
Shelley: Give me one.
Rob: How much a model can hold in front of it at once. The limit you hit when the document is too long.
Shelley: And the name for it?
Rob: Worth learning the day somebody bills you for it, and not before.
Shelley: That's a real answer.
Rob: It is the whole method. A word you cannot draw is one you will misuse confidently, and confident misuse is worse than not knowing.
Shelley: Give me three things.
Rob: One. A rename is a decision with a label stuck on the front. Read the decision.
Shelley: Give me the decisions.
Rob: Four bits reaching the training was a bet about multipliers. The crowd of specialists was a bet about not running everything. The letter U giving way was a bet about predictable growth.
Shelley: And the second?
Rob: Hold the picture, not the label, because pictures survive renames. And know the one case where that fails, which is a rename splitting one picture into two products. Then you learn the names.
Shelley: And the third?
Rob: The half I did not give you is not a gap. It is the part you collect from work, on the afternoon somebody uses the word at you and you can see what they are pointing at.
Shelley: And the season?
Rob: Twenty three conversations. The product map of one company used as a map of a field.
Shelley: Give me the sentence I can say back to somebody.
Rob: A generative model is a machine that learned the shape of data. Everything after that was a cost decision.
Shelley: Which decisions, specifically?
Rob: Which rung you climbed to. Which precision you stored it in. Which layer you served it from. And who paid for it to think longer.
Shelley: And underneath all of it?
Rob: Still multiplying grids of numbers on a chip built for video games. That part has not moved once.