The AI Field Guideby Stephen Hunt's SFcrowsnest
The AI Field Guide / G
9 terms, explained without the techno-murk.
A model's ability to use what it learned on new examples it has not seen before.
A student who understands multiplication can solve unfamiliar sums; one who only memorised the practice sheet cannot. Good generalisation means the model learned a useful pattern rather than the answers.
A handwriting model correctly reads writing styles that were not in its training set.
A system where one model creates examples and another tries to spot the fakes.
The generator and discriminator improve by competing. GANs became well known for realistic image generation, although diffusion models are now common for many image tasks.
A generator makes synthetic faces while a discriminator learns to distinguish them from real photos.
AI that creates new text, images, audio, video, code or other content.
Generative AI learns patterns from data and uses those patterns to produce a new result. It does not simply retrieve a stored copy, although its output can sometimes closely resemble training material.
A tool creates a product description from a few bullet points.
A family of language models that generate content after broad training on large amounts of data.
Generative means it creates output, pre-trained means it first learns broad patterns, and transformer names the underlying model design. GPT is a particular model family, not another name for every chatbot or LLM.
A GPT model can be placed inside a chat application, writing tool or coding assistant.
The part of a GAN that learns to create new examples resembling its training data.
The generator is like a forger practising against a detective. It starts with random input and creates an image or other sample, then learns from whether the discriminator spotted the fake. A generator in a GAN is a specific component, not a general name for every generative-AI tool.
A generator produces a synthetic face and adjusts itself after the discriminator judges how convincing it was.
A method that improves a model by repeatedly taking small steps toward a lower error.
Imagine standing on a foggy hillside and feeling which direction slopes downward. You take a small step, check again and repeat until reaching a low point. The 'hill' is the model's error.
Training uses gradient descent to adjust weights in directions that reduce prediction mistakes.
The trusted reference used as the correct answer when training or testing a model.
Ground truth is the answer key against which predictions are compared. The name can sound more certain than reality: human experts may disagree, records may be incomplete and the chosen reference can itself contain mistakes.
A specialist's confirmed diagnosis is used as ground truth when testing a medical-image model.
Connecting an AI answer to reliable evidence, data or the real world.
Grounding can mean giving a model trusted documents, search results, database records or sensor data. It can reduce made-up answers, but only if the source information is relevant and sound.
A policy assistant cites the exact staff handbook paragraphs behind its answer.
Rules and checks intended to keep an AI system within acceptable boundaries.
Guardrails can filter inputs, restrict tools, check outputs or require human approval. They lower risk but are not an unbreakable safety barrier.
A finance assistant is prevented from transferring money without a person's confirmation.