The AI Field Guideby Stephen Hunt's SFcrowsnest
The AI Field Guide / R
11 terms, explained without the techno-murk.
A prediction method that combines the answers from many decision trees.
One tree can be swayed by quirks in its training data. A random forest grows many slightly different trees and lets them vote, like asking a crowd of varied advisers instead of trusting one.
Hundreds of trees vote on whether a transaction looks fraudulent.
A model designed to spend extra processing on multi-step problems before answering.
These models can perform better on difficult maths, code and planning tasks by using more inference-time work. They can still make mistakes, and a longer answer is not proof of better reasoning.
A reasoning model explores several approaches to a logic puzzle before choosing an answer.
A system that suggests items a person may find useful or interesting.
Recommendations may use past behaviour, similarities between people or features of the items themselves. They can be convenient, but they also shape what people see and may repeatedly narrow choices toward more of the same.
A streaming service suggests films based on viewing history and patterns from similar viewers.
A neural network that carries information forward while processing a sequence.
Unlike treating every item separately, an RNN keeps a running note of what came before. This makes it useful for ordered data, although that note can fade across long sequences.
An RNN uses earlier words in a sentence when processing the next word.
Deliberately trying to make an AI system fail so weaknesses can be found.
Testers probe for harmful outputs, security gaps, bias and unexpected behaviour. Red teaming helps reveal problems but cannot prove that every possible failure has been found.
A team tests whether disguised instructions can trick an agent into sharing confidential files.
Predicting a number rather than choosing a category.
Classification answers 'which kind?'; regression answers 'how much?' or 'how many?'. It learns how input features tend to relate to a numerical result.
A model predicts a home's selling price from its size, location and condition.
Techniques that discourage a model from becoming too complicated or memorising its training data.
It is like grading a student on clear general rules, not rewarding them for memorising every answer on one worksheet. A small constraint can help the model perform better on new examples.
Training adds a penalty when model weights grow unnecessarily large.
Learning by trying actions and receiving rewards or penalties.
The system learns a strategy that tends to produce more reward over time. Designing the reward carefully matters because an AI may find shortcuts that score well without achieving the real goal.
A game-playing AI receives a positive reward for winning and learns which moves improve its chances.
Using human preferences as a signal to make a model's replies more helpful or acceptable.
People compare or rate model outputs, and that feedback helps train the system toward preferred behaviour. It reflects the examples and raters used, so it does not solve every alignment problem.
Reviewers choose which of two answers is clearer and safer, helping shape later responses.
Looking up relevant information and giving it to a generative model before it answers.
RAG helps a model use private or current sources without retraining. A search step finds useful passages, which are placed into the model's context. Poor retrieval can still lead to poor answers.
A workplace assistant searches company policies, then answers with links to the matching documents.
A physical machine that senses its surroundings and carries out actions.
Some robots follow fixed instructions, while others use AI to recognise objects, plan movements or adapt. AI does not need a robot body, and a robot does not necessarily use AI.
A warehouse robot moves shelves by following routes and avoiding obstacles.