Pelajaran 1 dari 47 menit
What generative AI is
Generative models produce new text, images or code by learning patterns from huge amounts of examples. Here is what that means, and what it does not.
Pelajaran ini hanya tersedia dalam bahasa Inggris. Antarmuka halaman tetap dalam bahasa Anda.
Dalam pelajaran ini
- Distinguish generative models from other kinds of AI
- Describe training and inference at a high level
- Name realistic uses and poor fits
From recognising to generating#
Much of the AI you have used for years classifies: is this email spam, is this photo a cat, which product will this customer like? A generative model instead produces something new: a paragraph, an image, a function, a summary.
Predictive models
- Output a label or a number
- "Is this transaction fraud?"
- Easy to measure: right or wrong
Generative models
- Output new content
- "Draft a reply to this customer"
- Quality is a judgement: helpful, accurate, on tone
Large language models (LLMs) are generative models for text, and the ones behind most chat assistants and coding tools. The rest of this course focuses on them.
How a language model learns#
An LLM is trained on a very large collection of text (web pages, books, code) to do one deceptively simple task: predict the next piece of text given what came before. To do that well across billions of examples, the model has to pick up grammar, facts, styles and reasoning patterns. All of it is stored as numbers (its parameters), not as a searchable copy of the documents.
Most assistants then go through further training on examples of helpful conversations and on human feedback, so they follow instructions instead of just continuing text.
Catatan
The model does not look anything up while it answers unless a tool gives it access (search, a database, files). Everything else comes from patterns learned during training, which stopped at a knowledge cutoff date.
Training and inference#
| Phase | What happens | Who does it |
|---|---|---|
| Training | The model's parameters are adjusted on huge datasets; costs a lot of compute | Model providers |
| Inference | The trained model generates a response to your input | Every time you use it |
As a developer you almost always work at inference time: sending input to a model through an app or an API and using the output.
What it is good at, and what it is not#
Generative models are strong at language-shaped work: drafting, rewriting, summarising, translating, explaining code, extracting structure from messy text, and brainstorming. They are weak where exactness matters and nothing checks the output.
Useful, but not an oracle
A model's answer is a fluent prediction, not a verified fact. It can be correct, subtly wrong or entirely made up, and it sounds equally confident each time. Use models where a person or a program can check the result, or where "good enough and fast" beats "perfect and slow".
LatihanGood fit or poor fit?
For each task, decide whether a language model is a good fit and why: (1) summarising a long meeting transcript, (2) calculating payroll tax, (3) suggesting names for a new product, (4) deciding whether to approve a loan.
Tampilkan solusi
(1) Good fit: language-shaped and easy for a person to check. (2) Poor fit: exact arithmetic and rules; use ordinary code and let a model at most explain the result. (3) Good fit: creative, many acceptable answers. (4) Poor fit as the decision-maker: high stakes, needs explainable and fair rules; at most a model might help summarise documents for a human reviewer.
Recap#
- Generative models create content; LLMs do it for text by predicting what comes next.
- Training builds the model once; inference uses it every time you send input.
- Treat outputs as drafts to check, not facts.