A GAN is a deep learning model that trains a generator, which makes fake data, against a discriminator, which tries to tell it apart from real data, until it can create new data that resembles the real thing.
Random numbers
Generator makes fake data
Discriminator judges real or fake
Both relearn from the verdict
The generator learns to be harder to catch and the discriminator learns to catch it more easily, and this repeats.
3 / 3
The generator builds data from random numbers, and the discriminator compares real data with the generator's output to judge which is which. Based on each verdict, the generator adjusts to become harder to catch, while the discriminator adjusts to catch it more easily. Through this repeated contest, the generator learns to produce data that is hard to tell from the real thing.
AnalogyPicture a forger and an art appraiser locked in a contest: the sharper the appraiser's eye becomes, the closer the forger's paintings get to the genuine article.
How to rememberThe letters stand for G = Generative, A = Adversarial, N = Network, so the three letters alone let you rebuild the full name.
Exam points
Mix-upThe discriminator (識別器) inside a GAN judges real versus fake, but it is only a part for training the generator. The purpose of the GAN as a whole is to generate new data.
Generative Adversarial Network (敵対的生成ネットワーク) is the full name of GAN, and the Japanese term is its official Japanese rendering.
"Adversarial" describes the rivalry in which the generator and discriminator each try to outwit the other, and "generative" describes the goal of creating new data. Each part of the name maps directly onto a key point of how the model works.
How to rememberIn English, Generative comes first, but the Japanese name puts "adversarial" (敵対的) first, so the word order is reversed.
Exam points
Mix-upIn this name, "adversarial" (敵対的) refers to the way the two networks inside a GAN learn by competing. It does not mean outside attacks or unauthorized access.
An LLM is a generative AI model that learns from an enormous amount of text and produces natural text that continues a given sentence.
Input sentence
Assign probabilities to next-word candidates
Pick one word and append it
Repeat to build a passage
It feeds the longer sentence back in and repeats the process until a full passage takes shape.
3 / 3
It is a language model, which predicts the probability of the next word in a sentence, built at a very large scale. It is first trained on vast data from a wide range of fields (pre-training) to gain general ability, and can then be adapted to a specific use through additional training (fine-tuning). A model that serves as the base for many uses in this way is called a foundation model.
AnalogyIt is like someone who has read countless books being shown an unfinished sentence and writing a perfectly plausible continuation.
How to rememberThe leading L is Large, the same L as on a large-size shirt.
Exam points
ExamAn LLM is the leading language example of a foundation model (基盤モデル), a model that gains general ability through pre-training (事前学習) and can be repurposed for many uses with additional training.
Large Language Model (大規模言語モデル) is the full name of LLM, and the Japanese term is its official Japanese rendering.
"Language model" means a model that expresses, as probabilities, how likely words are to follow one another, while "large" refers to the huge volume of training text and the huge number of internal values adjusted during training (parameters). Scaling up is what allowed it to handle a wide range of language tasks, such as answering questions, summarizing, and translating.
How to remember"Language model" is the name of a long-established technique; what is new is the "large" in front of it.
Exam points
Mix-up"Large" (大規模) refers to the size of the training data and the number of parameters. It does not mean the model handles many different languages.
Prompt engineering (プロンプトエンジニアリング) is the technique of refining how you write the instructions (prompts) given to a generative AI so that it produces output that fits your goal.
Write the instructions
Check the output
Revise the instructions
Rewrite the instructions to fill in what was missing, then hand them over again.
3 / 3
Even with the same model, output quality varies greatly depending on how specific the instructions are and how the background is presented. Typical techniques include stating up front what role the AI should play, breaking the task into small steps, and specifying that the answer come as a table or bullet points. People also often look at the output and revise the instructions over many rounds of trial and error.
AnalogyIt is like briefing someone you are working with for the first time: tell them the purpose, the deadline, and what the finished product should look like, and what you get back will be much closer to what you wanted.
How to rememberPrompt is the same word as the theater prompter who whispers cues to actors, and here it means a person handing the AI a cue to start writing.
Exam points
ExamPrompt engineering (プロンプトエンジニアリング) is the practice of polishing the questions and instructions given to an AI to draw out answers that match your intent.
Mix-upFine-tuning (ファインチューニング) changes the model's parameters through additional training. Prompt engineering leaves the model untouched and works on the instructions you feed in instead.
A GAN pits a generator against a discriminator and learns to create new data close to the real thing.
An LLM is a language model that has learned next-word prediction at a large scale, and it is a type of foundation model that can be adapted to specific uses with additional training after pre-training.
Prompt engineering raises the quality of generative AI output by refining how specific the instructions are and what form the answer should take.
The abbreviations map to full names as follows: GAN is Generative Adversarial Network (敵対的生成ネットワーク), and LLM is Large Language Model (大規模言語モデル).