Fine-tuning

What fine-tuning is, how continuing to train a model on your own examples differs from prompting and from retrieval, what it suits and what it does not, and what it costs in data and testing.

Free to watch, no sign-in, captions.

More shorts

3 min 14 Agents What an AI agent is: a model given a goal, tools it can call and a loop of deciding, acting and observing; how that differs from a single reply, where the loop goes wrong, what a person should decide before letting one act, and why people disagree about the word.
2 min 47 Reasoning What a reasoning model is, why writing out intermediate steps tends to help on multi-step problems, what the visible steps are and are not, when the extra steps cost more than they give, and how to check a shown chain of steps.
2 min 54 Transformers How a transformer, the neural network design behind many large language models, turns text into tokens and vectors, lets each token weigh the others through attention, stacks layers to end in a probability for the next token, and why the cost of attention bounds how much text a model can take in at once.
3 min 10 Tokens What a token is in a language model, how a tokenizer cuts text into pieces drawn from a fixed vocabulary, why the same sentence takes different numbers of tokens in different languages, why context length and usage are counted in tokens, and how tokenization contributes to mistakes with spelling and arithmetic.
3 min 13 Embeddings What an embedding is, how a piece of text becomes a point in a space where nearby points have similar meaning, how vector search finds the points nearest to a question, and why a close match in meaning is not the same as a correct one.
3 min 17 Deepfakes What a deepfake is (synthetic or altered audio, image or video of a real person made by a generative model), why one is now cheap to make, the three common harms, why looking harder at the pixels is the wrong instinct, and the checks that work instead: call back on a known number, a family code word, tracing the source, reverse image search and content credentials.
3 min 15 Context What the context window of a language model is, what counts toward it, why its size is limited, what happens when a conversation outgrows it, why fitting in the window does not mean every part is used equally well, and what a careful user does about it.
2 min 59 Training What training a language model involves: pre-training on a very large body of text by predicting the next token, then post-training on examples and preferences that shape how it responds, how the character of the data shapes the output, what training does and does not store, and why using people's writing to train models is contested.
2 min 53 Distillation What distillation is: how a small AI model is trained to reproduce the outputs of a large one, what the small model gains and cannot exceed, and why a provider's terms of service can matter to how it is trained.
3 min 19 Quantization How quantization stores a model's parameters with fewer bits per number to save memory and time, what rounding costs and where it matters more, and why the result should be tested on your own task.
2 min 44 Superintelligence What the word means, the three kinds one philosopher distinguishes, why goals and capability are separate questions, and why the idea itself is argued over.
2 min 55 Foom What 'foom' means: the idea that an AI able to improve itself could go from human level to far beyond it very quickly, and the long-running disagreement about whether such a takeoff would be fast or slow.
2 min 40 RAG What retrieval-augmented generation is: how an assistant answers from your own documents by searching them first, why that lowers errors without removing them, and what to check in a cited answer.
3 min 11 Diffusion How many image and video generators make a picture: noise is added to images step by step in training, then removed step by step to generate, steered by a text description.
3 min 8 Hallucination Why an AI assistant can state something false with complete fluency, what is happening inside the model when it does, and what reduces it without removing it.
2 min 59 Jailbreaking What it means to get an AI assistant past its own rules, why that is hard to prevent, how it differs from prompt injection, and why it matters to anyone who puts an assistant in a product.
3 min Emergence What 'emergent abilities' means for large models, the disagreement about whether abilities jump or the way we score them does, and what to ask when you read the word.
3 min 8 AGI What the term artificial general intelligence means, the different ways people define and measure it, and why the question 'is this AGI?' cannot be answered until a definition is stated.
2 min 33 ChatGPT What the assistant most people meet first actually is, what it is good at, where it gets things wrong, and what it costs to start.