Artificial Intelligence (AI)
Definition: The broad field of computer science focused on creating systems that can perform tasks that typically require human intelligence, such as understanding language, recognizing images, or making decisions.
Discussion: AI is the umbrella term. Large language models (LLMs), chatbots, and generative AI are all specific types within this broader field. Not all AI creates new content – some AI is designed only to analyze, classify, or predict.
Bias
Definition: Systematic tendencies in AI outputs that can reflect or amplify patterns from the training data.
Discussion: Bias isn’t intentional – it comes from patterns in the data the model was trained on. For example, if most training text associates “CEO” with men, the model may reflect that. Be alert to stereotyping, framing, or exclusion and correct for it.
Context Window
Definition: The maximum amount of text (measured in tokens) that a model can take into account at one time.
Discussion: The context window explains why a model may “forget” parts of a long conversation or document. If the interaction is longer than the model’s limit, earlier text is dropped from consideration. Larger context windows allow the AI to handle bigger documents or longer conversations more effectively.
Customization
Definition: The process of adapting an AI system by adding persistent context or instructions that shape how it responds over time.
Discussion: Customization does not alter the underlying model. Instead, it layers user-defined instructions – such as preferred tone, spelling conventions, or reference materials – on top of the model. This makes outputs more consistent with user expectations across sessions. Customization is deliberate and user-driven, unlike personalization, which is often automatic.
Generative AI
Definition: A subset of AI that creates new content such as text, images, audio, or video, rather than just analyzing existing data.
Discussion: Tools like ChatGPT are generative AI because they produce new sentences or images based on patterns they’ve learned. This is different from older AI systems, which focused more on classification, prediction, or detection. Generative AI is especially useful for drafting, brainstorming, and creative work — but outputs still require review for accuracy and appropriateness.
GPT
Definition: Short for Generative Pre-trained Transformer, OpenAI’s family of large language models. GPTs can generate text and, in newer versions, also analyze data, interpret images, or process audio.
Discussion: GPT is specific to OpenAI. Other companies build transformer-based models under different names (e.g., Claude, Gemini, LLaMA). GPT outputs often resemble reasoning but are based on statistical pattern-matching. Treat it as a flexible assistant for drafting, analysis, or brainstorming, but remember it doesn’t verify facts or make real decisions.
Grounding
Definition: The practice of connecting a model’s response to specific, verifiable sources (for example, your docs, a database, or web search results) so the output is tied to real information rather than just the model’s internal patterns.
Discussion: Grounding doesn’t change the model itself; it adds trusted context at runtime and asks the model to base its answer on that context (often with citations). Many vendor docs describe grounding as linking outputs to authoritative data, frequently implemented via retrieval before generation. This reduces the odds of confident-but-wrong answers, especially for time-sensitive or domain-specific questions.
Hallucination
Definition: When an AI generates content that is wrong, misleading, or entirely made up, while presenting it as if true.
Discussion: Hallucinations occur because the model generates text by predicting the most statistically likely sequence of words, not by consulting facts or verifying information. This process can produce statements that sound confident but have no factual basis – such as invented statistics, misattributed quotes, or non-existent references.
Their likelihood depends on the task and the prompt: open-ended or speculative prompts often increase the risk. To avoid, use clear instructions, connect the model to reliable external data when possible (see Grounding), and independently verify any important details before relying on them.
For example, a tornado is very newsworthy in New Jersey because serious tornadoes are rare there—so when a tornado happens, there are a lot of news stories. LLMs spidered the web and indexed that news. Now, when you ask an LLM to help make a plan to prepare for climate change in New Jersey, it will tell you that tornadoes are one of your biggest problems and you should prepare for them. But…they probably aren’t.
https://mailchi.mp/aboard/zkd26k8jzm-10336916?e=be03de531c
LLM (Large Language Model)
Definition: An AI system built using the transformer architecture and trained on massive collections of text. It encodes patterns in language to predict and generate sequences of words, enabling it to draft, summarize, translate, or classify text.
Discussion: LLMs are the core technology behind GPT and similar tools. LLMs generate fluent text but do so through statistical prediction, not understanding.
https://uit.stanford.edu/service/techtraining/ai-demystified/llm
Model
Definition: A specific AI system trained to perform tasks, such as GPT-4, Claude, or Gemini.
Discussion: Different models have different strengths, weaknesses, and capabilities. Choosing the right model depends on the task – for example, some are stronger at reasoning, others at speed, cost efficiency, or handling multimedia.
Neural Network
Definition: A type of computer system inspired by how brains process information, made of layers of interconnected “nodes” that pass signals to one another.
Discussion: Transformers are a specialized kind of neural network designed for handling sequences like language. Neural networks don’t think like people – they calculate probabilities – but they’re the foundation of modern AI.
Personalization
Definition: The process of adapting AI outputs based on a person’s history, preferences, or past interactions.
Discussion: Personalization does not mean the model itself is learning or changing. The underlying model remains fixed. Instead, the system may add context – such as remembering recent conversations, saving user preferences, or applying organization-specific settings – before passing instructions to the model. This makes responses feel tailored without retraining the AI. Personalization can improve relevance and consistency, but it raises questions about privacy, transparency, and bias depending on what data is stored and used.
Prompt
Definition: The instruction or input you give the AI – a question, request, or set of directions.
Discussion: Prompts can be as short as a phrase or as detailed as a multi-step brief. The clarity and specificity of a prompt directly affect the quality of the response. Instead of a vague request like “help with our product launch,” more useful prompts are more specific like:
-
“Draft three LinkedIn posts announcing our new project-management app. Keep each under 100 words and use a professional but friendly tone.”
-
“Summarize this 2-page article into five bullet points for a team newsletter.”
-
“Suggest five subject lines for an email campaign. Example of the style we like: ‘Simplify Your Workday in Minutes.’”
-
“Explain this technical report in plain language for a non-expert audience.”
-
“Turn these meeting notes into an action-item list in table format.”
Prompts can ask the model to draft, rephrase, summarize, explain, structure, or suggest — and examples or formatting instructions help it match your needs more closely. Since the model doesn’t understand intent, prompting is how users communicate expectations and boundaries. Clear, specific prompts reduce the risk of irrelevant, off-tone, or inaccurate answers.
Prompt Engineering
Definition: The practice of designing and refining prompts to guide the AI toward more useful or reliable responses.
Discussion: Because the model generates text by predicting patterns, the way you ask matters. Prompt engineering is less about technical skill and more about being precise, structured, and iterative. For example:
-
Starting vague: “Make this clearer.”
-
Adding clarity: “Rewrite this paragraph so it’s easier to understand for a non-expert.”
-
Adding constraints: “Rewrite this paragraph in under 100 words, keeping the key point about sustainability.”
-
Adding examples: “Here’s the tone we want: friendly and inviting, like this sentence — ‘Join us for an evening of good food and conversation.’ Please adjust the paragraph to match that style.”
This process-oriented approach shows how prompts can shape outputs step by step — tightening clarity, adjusting tone, or creating variations — rather than expecting the AI to produce a finished product on its own. Small adjustments to prompts often lead to significantly better results. Tip: When you find a prompt that works especially well, save it. Reusing and adapting effective prompts saves time and creates more consistent outputs.
https://uit.stanford.edu/service/techtraining/ai-demystified/prompt-engineering
System Prompt
Definition: The hidden instruction that every AI interaction starts with. It sets the model’s role, tone, and boundaries before user input is considered.
Discussion: System prompts are often detailed documents, not just a single sentence. They can define the model’s role (“You are a helpful assistant”), tone (“Use plain language”), formatting rules (“Answer in bullet points”), and limits (“Do not provide medical, legal, or financial advice”). These guardrails exist because AI models can generate text that sounds confident but may be inaccurate or unsafe in sensitive areas. For example, a system prompt might include:
You are an AI assistant that helps professionals with writing and research.
- Be clear, concise, and polite.
- When explaining complex topics, use plain language and short examples.
- If the user asks for factual information, provide it accurately. If uncertain, say so.
- Do not provide medical, legal, or financial advice.
- Format lists with bullet points unless the user requests another style.
- Keep responses under 500 words unless otherwise requested.
Knowing that these hidden instructions exist helps explain why the model may consistently avoid certain requests or default to a particular style.
Token
Definition: The small units of text (words or pieces of words) that AI models use to process and generate language.
Discussion: Tokens affect the length and cost of interactions. You usually don’t manage them directly, but longer prompts and responses will consume more tokens, which can raise costs or hit length limits.
Training
Definition: The process of building an AI model by exposing it to large amounts of data and adjusting its internal parameters so it can predict patterns in language (or other inputs).
Discussion: Training is what makes a model capable in the first place. It’s extremely resource-intensive, involving powerful computers and vast datasets. Once trained, a model can be reused many times – but it doesn’t keep learning from everyday use unless it’s retrained or fine-tuned with new data.
Training Data
Definition: The large collections of text and information an AI model learns from during development.
Discussion: GPT does not pull live data from the internet when answering; it relies on what it was trained on. This means outputs may be out of date or incomplete. Fact-check and add current context where accuracy matters.
Transformer
Definition: A neural network architecture introduced by Google researchers in 2017. Its key innovation is the attention mechanism, which allows the model to evaluate relationships among all elements in a sequence (such as words in a sentence) simultaneously.
Discussion: This architecture makes it possible to process long passages of text efficiently and capture context more effectively than earlier approaches. Transformers are the foundation that nearly all modern language models are built on.
Attention Is All You Need, https://arxiv.org/abs/1706.03762