Question 1

What actually is Artificial Intelligence?

Accepted Answer

AI is software that learns patterns from data instead of following rules a human typed out line by line. Show it millions of examples and it builds a statistical model of how things usually go, then uses that model to answer, classify, predict or generate.

A cook who learned by tasting ten thousand dishes, not by reading one recipe.

Question 2

What is the difference between AI, Machine Learning and Deep Learning?

Accepted Answer

AI is the umbrella goal — machines doing things that look intelligent. Machine Learning is the main way we get there: learning from data. Deep Learning is a subset of ML that uses many-layered neural networks, and it is what powers today's language and image models.

Question 3

How does a Large Language Model work?

Accepted Answer

It converts text into tokens, then predicts the next token over and over, each time conditioned on everything before it. During training it saw enormous amounts of text and adjusted billions of internal weights to make good predictions. Fluency, reasoning and style all emerge from that one simple objective at scale.

Question 4

What is Generative AI?

Accepted Answer

AI that produces new artefacts — text, images, audio, video, code — rather than only labelling or scoring existing ones. Traditional AI answers 'is this spam?'; generative AI writes the email.

Question 5

What are tokens, and why do they matter?

Accepted Answer

Tokens are the chunks of text a model reads and writes — roughly three quarters of a word each in English. They matter because both price and memory are measured in tokens: your bill, the context window and the speed of an answer all scale with them.

Question 6

What is a context window?

Accepted Answer

The maximum amount of text a model can hold in view at once — prompt, documents, conversation history and its own answer. Go beyond it and the earliest content falls out of view. Bigger windows help, but relevance still beats volume.

Question 7

Why does AI make things up?

Accepted Answer

Because it is optimised to produce plausible continuations, not verified facts. When it lacks grounding it fills the gap with something that reads right. This is called hallucination and it is a property of the method, not a bug you can patch away.

Question 8

How do I stop hallucinations in my own use?

Accepted Answer

Give it the source material rather than trusting memory, ask for citations you can open, keep the task narrow, request 'say you don't know if it is not in the text', and verify anything consequential. Retrieval and structured output cut error rates dramatically.

Question 9

What makes a good prompt?

Accepted Answer

Context, role, task, constraints and format. Tell it who it is writing for, what success looks like, what to avoid, and how the answer should be shaped. Show one example if the format is unusual. Vague in, vague out.

Question 10

Should I use a bigger model or a better prompt?

Accepted Answer

Better prompt and better context first — it is free and instant. Move up a model tier only when quality plateaus on genuinely hard reasoning. Most disappointing outputs are underspecified inputs.

Question 11

What is temperature?

Accepted Answer

A randomness dial. Near zero the model picks the most likely next token every time, which is what you want for extraction, code and data. Higher values loosen the choice and produce more variety, which suits brainstorming and copywriting.

Question 12

Can AI reason, or is it just pattern matching?

Accepted Answer

Practically, it does something in between. Chain-of-thought and reasoning models genuinely improve on multi-step problems by generating intermediate steps. But it is reasoning learned from patterns, so it is strong where the shape of the problem is familiar and brittle where it is truly novel.

Question 13

What is RAG?

Accepted Answer

Retrieval-Augmented Generation. You store your documents as embeddings, retrieve the handful most relevant to the question, paste them into the prompt, and let the model answer from that evidence. It grounds answers in your own truth and makes citations possible.

Question 14

RAG or fine-tuning — which do I need?

Accepted Answer

RAG when the model needs to know *facts* that change: policies, pricing, tickets, docs. Fine-tuning when it needs to learn *behaviour*: a house tone, a rigid output format, a niche classification style. Most teams need RAG; a minority also need fine-tuning.

Question 15

What are embeddings?

Accepted Answer

Numerical fingerprints of meaning. Text is mapped into a high-dimensional space where similar meanings land close together, so you can search by concept rather than by keyword — 'cancel my plan' finds a document titled 'subscription termination'.

Question 16

What is an AI agent?

Accepted Answer

A model given tools, memory and a goal, allowed to loop: decide, act, observe the result, decide again. The difference from a chatbot is autonomy over multiple steps — it books the meeting rather than telling you how to book it.

Question 17

What is MCP?

Accepted Answer

Model Context Protocol — a common connector standard so any AI client can discover and use any tool or data source without a bespoke integration for each pairing.

Question 18

Do I need to know how to code to build with AI?

Accepted Answer

No longer a prerequisite. Prompting, no-code builders and AI-native app platforms let you ship real products. Understanding logic, data and how to specify a system clearly matters far more than syntax.

Question 19

How much does it cost to run AI in a product?

Accepted Answer

You pay per token in and out. A short support answer costs a fraction of a cent; a long document analysis can cost several cents. Costs are controlled by routing easy tasks to small models, caching, trimming context and capping output length.

Question 20

How do I evaluate whether my AI feature is any good?

Accepted Answer

Build a test set of real examples with known good answers, score every change against it, and track a small number of production signals — task completion, escalation rate, edit distance between draft and what the human sent. Vibes do not survive contact with scale.

Question 21

Is my data used to train the model?

Accepted Answer

It depends entirely on the plan. Consumer free tiers often train on inputs by default; business and API tiers typically do not and offer retention controls. Check the terms, turn training off where offered, and never paste secrets into a tool you have not cleared.

Question 22

Is AI biased?

Accepted Answer

Yes, in the way its training data and its tuning are biased. It reflects patterns in human-produced text, including unfair ones. Mitigation is process, not magic: diverse evaluation sets, measured outcomes by segment, and a human decision-maker wherever the stakes are real.

Question 23

What is prompt injection?

Accepted Answer

An attack where malicious instructions hidden in content the model reads — a webpage, a PDF, an email — hijack its behaviour. Defend by treating all retrieved content as untrusted data, restricting what tools an agent may call, and requiring confirmation for irreversible actions.

Question 24

Who owns AI-generated output?

Accepted Answer

Commercially, most providers assign you the output and many indemnify enterprise customers. Legally, purely machine-generated work has weak copyright protection in several jurisdictions. Substantial human authorship strengthens your claim, so keep a record of your contribution.

Question 25

Can AI be wrong in ways I cannot detect?

Accepted Answer

Yes, and that is the real risk. Confident, well-formatted, subtly wrong output is harder to catch than obvious nonsense. Design workflows so errors surface — citations, cross-checks, and a human owner for every consequential output.

Question 26

Will AI become conscious?

Accepted Answer

There is no evidence that predicting tokens produces inner experience, and no accepted test for it. The practical risk today is not machine consciousness but machine competence deployed without oversight.

Question 27

Will AI take my job?

Accepted Answer

It takes tasks before it takes jobs. Roles heavy in routine drafting, summarising, matching and first-pass analysis compress hardest; roles heavy in judgement, relationships and accountability expand. The reliable move is to become the person who directs AI in your domain.

Question 28

Which skills matter most now?

Accepted Answer

Problem framing, data literacy, verification instinct, systems thinking, and clear writing — because writing well is now literally how you operate the machine. Domain depth also rises in value: it is what lets you spot a wrong answer instantly.

Question 29

How should a company start with AI?

Accepted Answer

Pick one painful, high-volume, low-risk workflow. Measure it before. Ship a narrow assistant with a human in the loop. Measure it after. Then scale the pattern. Enterprise-wide platform programmes with no first win rarely survive their second budget cycle.

Question 30

What is AGI, and how close are we?

Accepted Answer

Artificial General Intelligence means a system matching human capability across essentially all cognitive tasks. Credible forecasts still range from under a decade to never, and the definition itself keeps moving. Plan your business on today's capability curve, not on an AGI date.

Question 31

Where should a complete beginner start?

Accepted Answer

Two courses and three tools. Do Andrew Ng's 'AI for Everyone' and 'Generative AI for Everyone', then use AI daily for a month on real work — a presentation, a research question, a small web page. Fluency comes from reps, not from reading about reps.

AI FAQs

The Basics

Using AI Well

Building With AI

Risk, Data and Trust

Work, Strategy and What's Next