Filter articles
The basics of artificial intelligence (AI)
12 Jun 2025, 12:59 PMIf you're new to artificial intelligence (AI), or want to revisit some of the basics about this technology, this article is for you.
At Catalyst, we have many experts and enthusiasts who have been involved in machine learning and artificial intelligence in some shape or form for decades. Now AI has become more mainstream, we want to help people adopt it practically, safely and with intention. That’s why we’re launching a series of articles on AI so you can better understand what it is, how it works, where it’s useful, and the principles that underpin our approach to AI at Catalyst.
First, let’s focus on the basics, cover a bit of AI history, and a few common terms.
AI didn’t just appear. It evolved in waves.
Did you know, we’re living through the third major wave of AI?
The origins of artificial intelligence trace back to the 1950s when British mathematician Alan Turing, the same brilliant mind who helped crack the Enigma code during World War II, asked a deceptively simple question:
"Can machines think?"
That question, posed in his paper Computing Machinery and Intelligence, essentially launched the field of artificial intelligence. This was the beginning of the first wave of excitement about AI. Early projects like ELIZA (developed at MIT in the 1960s) also began to experiment with human-like conversation, and researchers started exploring symbolic reasoning and systems.
The second wave arrived in the 1980s, bringing us neural networks and the machine learning techniques that still underpin much of today's technology.
The new wave of AI
We're now in the third wave. What's different this time isn't the core idea — that hasn't changed since Turing's day. Instead, three major factors brought this third wave of excitement:
- Massive data availability: Every search, purchase, and post generates training data for AI systems.
- Advanced computing power: Graphics Processing Units (GPUs) and cloud infrastructure enabled models to process more data, faster.
- Breakthrough use cases, such as DeepMind’s AlphaGo beating human champions or OpenAI’s ChatGPT generating human-like text, captured global attention.
By 2022–2023, AI was no longer an abstract concept. It was popping up in your workplace, your browser, and possibly your daily routine.
Understanding the basics: What is AI?
Defining AI isn't as straightforward as you might expect. That's because AI means different things in different contexts. To compound that, many people now use it as a “catch-all” term.
For organisations, the legal and regulatory definition of AI is what we believe counts in practice. In New Zealand's case, the OECD definition is currently the closest we have to an official reference, though it's worth noting that regulatory frameworks are still developing.
For our purposes, we'll focus on how AI is understood in computer science, since that's where most of the practical development happens, and this aligns with the OECD definition.
The computer science lens
At its core, AI refers to computer systems that can do things we typically think of as requiring intelligence. This includes recognising what's in a photo, understanding what you're saying, making decisions, and solving complex problems.
Most of this functionality comes from Data Science and Machine Learning (ML), where systems learn patterns from data instead of following manually programmed rules.
Think of it this way:
- You show the system tons of data (like thousands of photos of cats and dogs).
- The system spots patterns in that data (maybe cats have pointy ears and dogs have floppy ones, among hundreds of other subtle patterns).
- When you show it a new photo, it uses those learned patterns to make predictions (this new photo probably shows a cat).
- Later, the AI model might be updated further. For example, accounting for seasons by adding pictures with snow or adding rabbits to the data.
Without good quality data and careful training, even the most sophisticated AI system won't work well.
And this is exactly why we need to talk about ethical and responsible AI development, because the patterns these systems learn come directly from the data we feed them, including any biases (conscious or subconscious) or limitations that data might contain.
The three main types of AI models
1. Predictive models
These forecast future outcomes using historical data.
Examples: Forecasting stock trends, suggesting products, and predicting travel delays.
2. Discriminative models
These classify data or distinguish between categories.
Examples: Identifying spam emails, detecting faces in photos, and identifying songs.
3. Generative models (GenAI)
These create new content like text, images, or music, based on learned patterns.
Examples: Writing emails, generating images, and creating product descriptions.
Modern AI systems are increasingly combining multiple capabilities to create more helpful experiences.
Let's take travel planning as an example. An AI assistant might:
- analyse your preferences and past choices,
- predict destinations and activities you'd enjoy,
- then generate personalised itinerary options for you to choose from.
This integrated approach makes AI more useful by drawing on different strengths to solve complex, real-world problems.
Large language models (LLMs): the engine behind generative AI
One of the most prominent types of generative AI is the large language model (LLM).
LLMs are trained on vast volumes of text — books, websites, articles — and learn to predict the next word in a sentence. That simple mechanism, at scale, enables them to generate everything from emails to code to customer support replies.
A few LLMs you may have heard of:
- OpenAI’s GPT series.
- Anthropic’s Claude.
- Google’s Gemini.
- Mistral’s Mixtral.
- High-Flyer’s DeepSeek.
Many different LLMs are now embedded into everything from search engines to design tools, democratising access to advanced language generation.
Open-source vs proprietary models
As with other software and platform developments, there are two approaches that organisations use to build these models: open vs closed.
Proprietary (closed) models
These are developed and controlled by companies like OpenAI, Google, or Anthropic. The inner workings and training data are typically not made public. Ultimately, that means it’s hard to understand what training data was used and the quality of it, any potential biases that may be replicated, the considerations of cultural considerations or training that may have been implemented.
Open source and open-weight models
Examples of these models include OLMo, Mixtral or DeepSeek, to name a few. These models are shared more freely. Developers can inspect, adapt, and even deploy them privately, making them attractive for enterprises needing transparency or control. However, access to the training data is only available in open-source models and not open-weight models.
This distinction impacts everything from innovation speed to data privacy to compliance. For organisations, choosing between open and proprietary AI is a strategic decision and not a one-size-fits-all.
Our team prefers helping our clients adopt open-weight and open-source AI. These models enable greater security and transparency. For example, how they respond and what data they were trained on, so we can minimise the impact of issues like bias.
Why Catalyst is committed to responsible AI
At Catalyst, we don’t use AI just because it’s available; we use it because it serves a purpose, and we focus on enabling others to adopt it with intention.
We’re committed to:
- Transparency: understanding how AI tools work and where their limits lie.
- Ethical use: respecting data, people, and outcomes.
- Long-term thinking: avoiding shortcuts in favour of sustainable innovation.
This is the lens through which we evaluate and deploy AI in our business and our client work. You can learn more about our approach to artificial intelligence in our AI principles.
What’s next
We want everyone to be able to interact with AI safely, responsibly and ethically. That’s why we created the AI Discovery Lab, where you can learn about key considerations when using LLMs.
We'll be updating our website with even more articles and educational content on AI over the coming months. If you don't want to miss these, follow us on LinkedIn or subscribe to our mailing list.
If you’re already exploring AI in your organisation, we’d love to hear where it’s working or where it isn’t.
Related projects
The curious users ultimate guide to asking better AI questions
This guide covers seven essential questions to ask when using GenAI, though most considerations apply to other AI systems as well.
How can student data be used ethically in AI and analytics?
Discover how we can continue to protect student data and ensure it’s used ethically with AI and LMS analytics.