The AI Field Guideby Stephen Hunt's SFcrowsnest
The AI Field Guide / C
23 terms, explained without the techno-murk.
Intermediate reasoning steps used by a model while working toward an answer.
Asking for step-by-step reasoning can sometimes help, but a written explanation is not a guaranteed window into the model's true internal process. Some systems keep private reasoning hidden and provide a shorter summary instead.
A model may break a word problem into smaller calculations before giving the final number.
Software designed to hold a conversation by text or voice.
Some chatbots follow fixed scripts. Newer ones often use language models, which makes them more flexible but also less predictable and more likely to invent an answer.
A shop's chatbot might answer delivery questions or help find a product.
OpenAI's conversational AI product for working with text, images, files and other information.
ChatGPT is the application people interact with, rather than one permanent model. A useful comparison is a web browser: the visible product provides the conversation and tools, while different AI models may do the work underneath. Like other generative AI, it can still misunderstand requests or produce incorrect information.
Someone uses ChatGPT to explain a document, brainstorm ideas or draft a first version of an email.
A saved version of a model at a particular point in training.
Training can take a long time, so developers save checkpoints along the way. A checkpoint contains learned weights and can be tested, shared or used as the starting point for more training.
If training crashes on day ten, the team can restart from yesterday's checkpoint rather than from zero.
When one category appears far more often than another in a dataset.
Imagine a thousand medical tests with only ten true cases of a disease. A model that always says 'no disease' would be 99 percent accurate but completely useless at finding the disease. Rare categories need suitable data and evaluation measures.
Fraudulent payments may be vastly outnumbered by legitimate payments in fraud-training data.
Giving an item one of several labels or categories.
Classification is like sorting post into labelled trays. The model studies examples, then decides which tray a new item most likely belongs in.
An email filter classifies a message as spam or not spam.
The cut-off score at which a model changes from one classification to another.
A smoke alarm must decide how much evidence counts as smoke. Lowering the threshold catches more possible fires but also creates more false alarms; raising it reduces false alarms but may miss real fires.
A transaction is blocked when its fraud score rises above the chosen threshold.
An OpenAI model that learned to connect pictures with written descriptions.
CLIP places matching images and text near each other in a shared mathematical space. Think of it as teaching a system that a photograph of a bicycle and the words 'a bicycle' belong together, without giving it a fixed list of every visual category it may ever encounter.
CLIP can compare a photograph with several descriptions and estimate which description fits best.
Automatically grouping items that seem similar, without being given the group names first.
Imagine tipping a mixed box of buttons onto a table and arranging them into piles by shape, colour or size. A clustering method finds its own useful piles from patterns in the data.
A retailer groups customers with similar shopping habits without deciding the customer types beforehand.
Combining several AI methods so each handles the part it is best suited to.
Rather than expecting one model to do everything, a composite system may use rules for firm boundaries, machine learning for prediction, search for evidence and a language model for explanation. It resembles a mixed team of specialists rather than one employee attempting every job.
A support tool searches approved documents, applies account rules and then uses an LLM to draft a clear response.
The processing power used to train or run an AI system.
Compute usually means the work done by specialised computer chips and data centres. More compute can allow larger experiments, but it also costs money and energy and does not automatically produce better AI.
Generating a short reply uses far less compute than training a large model from scratch.
The field of teaching computers to work with images and video.
Computer vision systems can identify objects, read text, track movement, inspect damage or describe a scene. They do not see as humans do; they analyse learned patterns in visual data.
A phone uses computer vision to recognise a face or scan a document.
When the real-world pattern a model learned changes over time.
It is like using an old map after the roads have changed. The model may still work exactly as built, but its predictions become less reliable because today's world no longer resembles its training data.
A fraud model trained on last year's scams misses a new style of fraudulent transaction.
A table showing the different ways a classification model was right or wrong.
Despite the name, it is a clarity tool. It separates correct positive and negative predictions from false alarms and missed cases, giving a more useful picture than one overall accuracy number.
A medical confusion matrix shows how many illnesses were found, missed, falsely reported or correctly ruled out.
The information an AI can use while producing its current answer.
Context can include your prompt, earlier messages, uploaded files, retrieved notes and hidden instructions. It is closer to a temporary workbench than a permanent memory.
Uploading a menu gives the model context it can use to answer questions about dishes.
Designing everything an AI receives so it has the right information and tools for the task.
Prompt engineering focuses mainly on the instruction. Context engineering arranges the whole workbench: system rules, conversation history, retrieved documents, examples, tool results and available actions. A beautifully written prompt cannot compensate for missing, stale or badly organised context.
A legal assistant receives the user's question, current case files, relevant policies, citation rules and access to an approved search tool.
The maximum amount of information a model can consider in one go.
The window is measured in tokens and includes both what you send and, usually, what the model produces. A large window can hold long documents, but important details may still be missed.
When a conversation becomes too long, its earliest messages may fall outside the context window.
Technology that lets people interact with a computer using ordinary conversation.
Conversational AI combines language or speech technology with software that decides how to respond or act. Some systems follow narrow scripts; others use language models. Human-sounding conversation does not guarantee that the system understands, remembers accurately or tells the truth.
A customer says, 'Where is my order?' and a voice assistant looks up the delivery and answers aloud.
A neural network designed to spot local patterns, especially in images.
It moves small pattern detectors across an image, rather like sliding a magnifying window over it. Early layers may notice edges; later layers combine them into shapes and objects.
A CNN learns to identify visual clues associated with cracks in a bridge photograph.
A general label for an AI assistant that helps a person do a task.
The word suggests that the human remains in charge while the AI drafts, suggests or handles routine work. It is also used in several product names, so its exact meaning depends on context.
A coding copilot suggests the next lines of a program while a developer reviews them.
Customer relationship software that uses generative AI to create or summarise material.
A CRM keeps a company's customer information and interactions organised. Adding generative AI can help draft emails, summarise support cases or suggest next steps using that information. Access controls matter because customer records may contain private or commercially sensitive data.
A sales representative receives a draft follow-up email based on approved notes from a customer meeting.
Testing a model several times using different portions of the available data.
The data is divided into sections. The model trains on most sections and is checked on the section left out, then the process rotates. It is like judging a student with several practice exams instead of trusting one unusually easy or difficult paper.
Five-fold cross-validation runs five trials, with a different fifth of the data used for checking each time.
Protecting computer systems, devices and data from attack, damage or unauthorised access.
Cybersecurity aims to keep information confidential, accurate and available to the right people. AI systems create familiar security concerns and newer ones, such as hidden prompt injections, stolen model access and agents being tricked into using tools improperly.
A company limits an AI assistant's file permissions and monitors unusual attempts to access confidential records.