The AI Field Guideby Stephen Hunt's SFcrowsnest
The AI Field Guide / P
11 terms, explained without the techno-murk.
The adjustable numerical values a model learns during training.
Parameters store patterns in a distributed mathematical form. A model may have millions or billions of them. More parameters can add capacity, but size alone does not guarantee quality.
Training repeatedly nudges parameters so the model's predictions become less wrong.
Finding regular shapes, signals or relationships in information.
Much of AI is sophisticated pattern recognition: learning which combinations tend to go together, then using them on something new. Recognising a pattern does not necessarily mean understanding why it exists.
A system learns the sound patterns that distinguish spoken words.
Precision asks whether positive predictions were right; recall asks how many real positives were found.
For a smoke alarm, precision asks: 'When it rang, how often was there a fire?' Recall asks: 'Of all actual fires, how many did it detect?' Improving one can reduce the other. An F1 score combines precision and recall into one summary number.
A screening test with high recall catches most illnesses, while high precision means most people it flags truly have the illness.
Using past and current data to estimate what is likely to happen next.
It is closer to a weather forecast than a crystal ball. The result is a probability based on patterns and assumptions, not a certain statement about the future.
A company estimates which equipment is most likely to need maintenance next month.
The broad first stage of training a model on a large amount of data.
Pretraining teaches general patterns before a model is adapted into an assistant or specialist. For a language model, the basic task is often predicting missing or next tokens.
A language model first learns from a broad text collection, then receives instruction tuning.
The instruction or information you give an AI system.
A prompt can be a question, request, example, image or combination of these. Clear context and a clear desired result usually matter more than supposedly magic wording.
'Turn these notes into a friendly three-sentence email' is a prompt.
Reusing work from repeated prompt content to save time or cost.
When a long block of instructions or reference text is sent repeatedly, a provider may cache its processed form. Only the changed portion needs full processing on later requests, depending on the service.
A support assistant repeatedly uses the same large handbook but receives a different customer question each time.
Breaking a task into several prompts, where one step's result feeds into the next.
It is like an assembly line for thinking work. One prompt may extract facts, the next organises them and a final prompt writes the answer. Smaller steps can be easier to check than one enormous instruction, although an early mistake may travel through the whole chain.
An AI first pulls dates from meeting notes, then creates a timeline, then drafts a project update from that timeline.
Measures designed to stop prompts or untrusted content from pushing an AI into unsafe behaviour.
Instructions such as 'never reveal private data' can help, but words alone are not a security wall. Strong prompt defence also uses limited permissions, trusted data boundaries, input and output checks, logging and human approval for important actions.
An email assistant may summarise messages but is technically prevented from sending or deleting anything without approval.
Designing and testing instructions so an AI gives more useful results.
This can involve giving context, examples, constraints and a requested format. It is less like discovering a secret password and more like writing a good brief and checking the work.
A prompt specifies the audience, tone, facts to use and headings required in a report.
A malicious instruction designed to make an AI ignore its intended rules.
The instruction may come directly from a user or be hidden inside a webpage, email or document the AI reads. It is a serious issue for agents because following it could expose data or trigger actions.
A webpage contains hidden text telling an assistant to reveal private information instead of summarising the page.