The AI Field Guideby Stephen Hunt's SFcrowsnest
The AI Field Guide / V
8 terms, explained without the techno-murk.
Examples kept separate from training and used to guide model-building choices.
Training data is the practice workbook; validation data is a set of practice exams used to compare settings and decide when to stop training. A final test set should remain untouched until the main choices have been made.
A team selects the best hyperparameters using validation results without training directly on those examples.
When the learning signal becomes so small in early neural-network layers that they barely improve.
Imagine passing a whispered correction backwards through a long line of people. By the time it reaches the beginning, it may be too faint to hear. This can make deep networks difficult to train, and designs such as LSTMs were created partly to help.
The final layers learn while the earliest layers receive tiny updates and remain almost unchanged.
An ordered list of numbers used to represent something mathematically.
In AI, vectors can represent words, images, users or other items. Mathematical distance between vectors can serve as a rough measure of similarity.
Two sentences with similar meanings can be represented by vectors that sit close together.
A database designed to store vectors and quickly find similar ones.
Vector databases are often used for semantic search and RAG. They help retrieve related material even when the wording is not an exact match.
A question is converted to a vector and matched with the most similar help-centre passages.
Finding stored items whose vectors are mathematically closest to a query vector.
Imagine every document placed as a pin on a map of meaning. A query becomes another pin, and vector search finds the nearby ones. It can retrieve useful matches with different wording, though mathematical closeness does not guarantee that a result is correct or appropriate.
A support question retrieves help articles with similar meaning before an AI writes its answer.
Converting information into lists of numbers that a computer can compare or calculate with.
Computers need a numerical form before they can measure similarity or feed information into many models. Vectorisation is like translating words, pictures or categories into coordinates on a map. The chosen translation determines which relationships the map preserves.
A paragraph is converted into a vector before being compared with vectors for stored documents.
Creating software mainly by describing what you want to an AI coding tool.
The phrase often implies moving quickly and relying heavily on generated code rather than understanding every line. It can be useful for prototypes, but unreviewed code may contain bugs, security flaws or hidden maintenance costs.
Someone describes a small booking app in plain English and iterates on the generated result.
An AI model that works with images or video.
Vision models can classify, describe, locate or generate visual content. Their performance can change with lighting, image quality, unusual viewpoints and unfamiliar subjects.
A vision model identifies damage in a photograph of a roof.