What Is Artificial Intelligence?

10 Fundamental Questions About How AI Works

1. Are AI neural networks really modelled on the human brain, or is the similarity misleading?

The inspiration comes from the brain, but AI is a mathematical system - not a digital replica of the human brain.

The fundamental difference is that AI neural networks are mathematical systems that learn patterns from data (basically combining complex non-linear and linear calculations) whereas the human brain learns through perception, physical experience, memory, emotion, reasoning and interaction with the world. The learning process is as different as it can be.

A human child can see a dog twice, hear it bark and interact with it to begin understanding what a dog is. An AI system may need to process thousands or millions of examples to learn to recognise dogs reliably across different sizes, breeds, poses and environments.

The human brain also does not learn everything from scratch. We can draw on what we already know, make connections and apply lessons to unfamiliar situations. AI can also generalise and transfer what it has learned but it needs to be there in the training data. Humans do it themselves anyways.

So, where does the similarity between AI and the human brain come from?

The similarity lies in the broad idea of interconnected units processing information in stages- not in how the brain and AI actually work.

A biological neuron is a nerve cell that receives signals from other neurons, processes them and passes signals onwards. Early AI researchers took inspiration from this basic concept and created simplified mathematical models of neurons.

2. What are the different components of an AI system?

An AI system consists of several components that work together to deliver an answer or perform a task.

• AI model: The engine that learns patterns from data and generates answers. Examples include GPT, Claude and Gemini. The Models can be LLM (large language model) or SLM (small language model).

• Data: The information used to train the model or provide it with additional knowledge.

• Computing infrastructure: Chips such as GPUs and data centres that provide the processing power to run AI.

• Application: The product users interact with, such as ChatGPT.

• Tools: Connections to the internet, databases or other software that let AI retrieve information and perform tasks.

• Agent: A system that uses an AI model and tools to complete tasks through multiple steps.

3. What are AI data centres, GPUs and AI chips and why are they so important?

AI models require enormous computing power to learn from data and generate answers. This computing power comes from specialised chips housed in data centres.

• AI data centres: Large facilities containing thousands of servers, specialised chips, storage systems and cooling equipment. They provide the infrastructure needed to train and run AI models.

• GPUs (Graphics Processing Units): Chips originally developed for graphics processing that are particularly good at performing many mathematical calculations simultaneously. This makes them highly effective for AI training and inference.

• AI chips: A broader category that includes GPUs and specialised processors such as Google's TPUs, designed to accelerate AI workloads.

• CPUs (Central Processing Units): General-purpose processors that coordinate tasks and run much of the conventional software in a computer or server.

Training a large AI model can require thousands of specialised chips working together for weeks or months. Even after training, the model needs computing power whenever someone submits a question and requests an answer. This process is called inference. This is why companies such as Nvidia, Microsoft, Amazon and Google are investing heavily in AI chips and data centres. The race to develop more capable AI is also a race to secure computing power, electricity, cooling capacity and high-speed networks. In simple terms, the AI model is the engine, the chips provide the computing power, and the data centre houses the infrastructure that makes the system work.

4. What is an LLM, and how does it differ from an SLM? Why do both exist?

The main difference is size, and capability - and therefore cost.

An LLM (Large Language Model) is an AI model trained on large amounts of data to understand and generate language. Examples include GPT, Claude and Gemini. An SLM (Small Language Model) is a smaller model designed to perform similar tasks with fewer computing resources. LLMs generally have greater capacity to handle a wide range of complex tasks, while SLMs can be faster, cheaper to run and suitable for specific tasks, sometimes even on a laptop or phone.

For example, a company might use an LLM to analyse a complex business strategy but use an SLM to classify customer support requests or summarise routine documents. As a smaller model is not necessarily less accurate for every task; a well-trained SLM can outperform a larger model on a narrowly defined job.

Why do both exist?

Not every task requires the most powerful model. Companies choose between them based on the complexity of the task, accuracy requirements, speed, privacy and cost. In simple terms, an LLM is like a versatile generalist, while an SLM is like a more resource-efficient specialist.

5. What is the difference between an AI model, an AI application and an AI agent?

Think of AI as having three different parts.

• AI model — the brain: The model is the technology that has learned from data and can answer questions, write stories or solve problems. GPT, Claude and Gemini are examples of AI models.

• AI application — the product you use: An application makes the AI model available to you through a screen or interface. For example, ChatGPT is an application that lets you interact with AI models. You type a question, and it displays the answer.

• AI agent — the task-doer:  An AI agent can go beyond answering questions. You give it a goal, and it can work through several steps to achieve it, using tools when needed.

For example, if you ask for a list of the top five competitors of a company:

• A model can explain how to identify competitors.

• An application lets you ask the question and read the answer.

• An AI agent can search the internet, identify competitors, compare their products and prepare a report.

The simplest way to remember it: The model provides the intelligence, the application lets you use it, and the agent uses it to get things done.

6. Where does AI get its information?

AI gets information from several sources, depending on how it has been built and how you use it.

• Training data: Before an AI model is released, it is trained on large amounts of text, code, images and other data. It learns patterns and relationships from this information. It does not store and retrieve every training document like a library.

• Internet searches: If the AI application has web search enabled, it can search for current information and use the results to answer your question. Without this capability, it cannot automatically look up the latest information.

• Uploaded documents: You can upload a PDF, spreadsheet or other file. The AI can analyse its contents and use them to answer questions about that specific material.

• Databases and company information: Businesses can connect AI systems to internal documents, customer records, product catalogues and other databases, subject to access permissions.

• Retrieval-Augmented Generation (RAG): This technique allows an AI system to retrieve relevant information from an external source (such as a company document library) and provide it to the model as context when generating an answer.

For example, a company could use RAG to let employees ask questions about its policies. The system retrieves relevant sections from the policy documents and gives them to the AI model to formulate an answer.

The key takeaway: Training gives the model what it has learned; searches, uploaded files and connected databases can supply additional information when needed. RAG is one way of bringing relevant external information into the answer-generation process.

7. Why do different AI models give different answers to the same question?

Think of AI models like car engines. Some are optimised for high horsepower (BHP), others for torque, fuel efficiency or smooth performance. Similarly, some AI models are designed to excel at complex reasoning, others at coding, speed, lower costs or handling large amounts of information. No single model is necessarily best at everything. Even when given the same question, different models may emphasise different factors, interpret ambiguity differently or produce different recommendations.

AI models such as GPT (OpenAI), Gemini (Google), Claude (Anthropic) and Llama (Meta) are built by different organisations. They use different training data, model designs, training methods and instructions, which influence how they answer questions.

What makes the AI models different?

1. Training data: Each model may be trained on a different mix of text, code, images and other information.

2. Size and design: Models differ in their number of parameters, architecture and computing requirements. Larger models may handle a broader range of complex tasks, while smaller models can be faster and cheaper.

3. Training and tuning: Developers train models to follow instructions and may optimise them for particular tasks, such as coding, reasoning, writing or research.

4. Tools and information access: Some AI products can search the internet, analyse uploaded files or connect to databases. Others may not have these capabilities, or may use them differently.

5. Speed and cost: Some models are designed for demanding tasks; others prioritise quick responses and lower operating costs.

For example, if you ask several models to assess a startup, one may focus on market size, another on competition, and another on the founder's ability to execute. The differences may reflect their analytical approach, not simply the quality of their intelligence.

8. How are companies actually using AI today?

Companies are using AI to automate routine work, analyse information, improve decisions and help employees work faster.

• Customer service: Answering customer queries, resolving common problems and assisting support teams.

• Software development: Writing and testing code, finding errors and helping developers build software faster.

• Marketing and sales: Creating content, analysing customer behaviour, personalising campaigns and identifying sales opportunities.

• Research and analysis: Summarising documents, researching markets, analysing competitors and extracting insights from large datasets.

• Finance: Detecting fraud, analysing transactions, forecasting financial performance and supporting risk assessment.

• Operations: Managing inventory, forecasting demand, improving supply chains and automating administrative tasks.

Companies are also integrating AI into existing products and developing AI agents that can perform multi-step tasks using software and data.

The key takeaway: Businesses are moving beyond using AI simply to generate text. They are integrating it into everyday workflows to reduce costs, improve productivity and create new products and services. The greatest value comes when AI improves a business process, not merely when a company adopts AI.

9. What happens to Google Search when AI can answer questions directly?

Traditionally, Google Search gives you a list of websites where you can find answers. AI can summarise information from multiple sources and provide an answer directly, reducing the need to visit individual websites.

This changes how people find information in three ways:

• AI-powered search: Search engines combine traditional links with AI-generated summaries and answers.

• AI chatbots: People increasingly ask tools such as ChatGPT to explain concepts, compare products or research topics without starting with a traditional search engine.

• AI agents: In the future, people may delegate tasks such as researching products, comparing prices or finding service providers to AI systems that search and act on their behalf.

For websites, this creates both a challenge and an opportunity. If AI answers a question without requiring a click, the website that supplied the information may receive less traffic. However, websites that provide original research, specialist expertise, useful tools and credible information may still attract visitors through citations, recommendations and deeper engagement.

The key takeaway: Search is evolving from finding websites to finding answers and completing tasks. For businesses and publishers, visibility will increasingly depend not only on ranking in Google but also on being useful and credible enough to be discovered and referenced by AI systems.

10. Can We Trust AI? What Are AI Hallucinations, and Why Does AI Give Wrong Answers?

AI can be useful for answering questions, analysing information and helping people make decisions. But AI can also give answers that sound convincing but are incorrect. This is one of the most important things to understand about artificial intelligence.

What are AI hallucinations?

An AI hallucination occurs when an AI system generates information that appears plausible but is false, inaccurate or unsupported by reliable evidence.

An example: Suppose you ask an AI model for the revenue of a private startup that does not publish its financial results. Instead of saying that the information is unavailable, the model might generate a plausible revenue figure based on patterns it has learned about similar companies. The answer may sound authoritative, but the number could be entirely fabricated. These mistakes are called hallucinations because the AI produces information that is not real.

Why does AI give wrong answers?

AI models learn patterns from training data and use those patterns to generate responses. They do not automatically verify every statement against a reliable source before answering.

As a result, AI can make mistakes for several reasons:

• Incomplete or inaccurate information:  The model may have learned incorrect information or lack the knowledge needed to answer a question.

• Outdated information: Its training may not include recent developments, unless it can access current information through search or other tools.

• Ambiguous questions: The model may misunderstand what you are asking or make assumptions about missing details.

• Plausible but incorrect responses: The model may generate an answer that fits familiar patterns even when the underlying facts are wrong.

• Faulty reasoning or calculations: AI can make errors when solving problems, interpreting evidence or performing calculations.

Can we trust AI?

Yes, but the level of trust should depend on the task and the consequences of getting it wrong.

AI can be useful for explaining concepts, summarising documents, generating ideas and helping with routine work. However, important facts, financial decisions, medical information, legal interpretations and other high-stakes outputs require independent verification.

AI tools with internet search or access to reliable databases can improve factual accuracy, but they can still misinterpret sources or draw incorrect conclusions.

Want to understand how technology creates business opportunities?

AI is changing how companies build products, compete and grow. Understanding these changes is increasingly valuable for anyone interested in startups and venture capital. The VC Academy helps aspiring venture capital professionals understand how investors evaluate startups, assess business models and make investment decisions.