Artificial intelligence: how it works and why it matters

Machine learning, deep learning, LLMs, economic disruption and ethical stakes

Introduction

Here is a contradiction worth sitting with: artificial intelligence is described in the same breath as the greatest technological leap since electricity and an existential threat to humanity, yet most people who have strong opinions about it cannot quite explain what it actually is. How can we meaningfully debate something we do not really understand?

The truth is more nuanced than either extreme. AI is neither a magical intelligence that thinks like a human being, nor an impenetrable black box accessible only to Silicon Valley engineers. It is a set of precise mathematical techniques, built on principles that any curious, non-technical person can grasp.

That is exactly what this piece gives you. You will come away understanding what machine learning is, why deep learning changed everything, what LLMs like GPT and Claude actually do, and how all of this is reshaping the global economy, while raising ethical questions none of us can afford to ignore.

By the end, you will be equipped to read an AI headline, cut through the noise, and understand what is really going on.

The global AI market, from $4.5 Bn in 2017 to over $800 Bn projected by 2030.

A history of AI, from the Dartmouth Conference (1956) to generative AI (2022–2024).

The Fundamentals: What AI Actually Is

A term that covers very different realities

The phrase ‘artificial intelligence’ was coined in 1956 at a summer workshop at Dartmouth College in New Hampshire. The goal was ambitious: build machines that could replicate human cognitive functions. Nearly seventy years later, that phrase covers an enormous range of realities, from an algorithm that recommends your next video to a system that detects cancer in a radiology scan.

To make sense of it all, think in layers, like a set of Russian nesting dolls. AI is the outermost doll, the broad concept. Inside sits machine learning. Inside that, deep learning. And at the very core, large language models, or LLMs, the technology behind ChatGPT, Claude, and Gemini.

“Classical” AI: when programmers write the rules

Before modern approaches took hold, AI systems worked on a straightforward principle: engineers wrote the rules explicitly. A chess program from the 1980s had hundreds of hand-coded instructions, if the king is under threat, move piece X. A fraud detection system from the 1990s had a checklist of suspicious conditions: if a transaction exceeds $10,000 and originates from an unusual country, flag it.

This approach works well for tightly defined problems with stable rules. But it hits a wall very quickly. How do you write the rules for recognising a cat in a photo? Or for understanding what a sentence means depending on its context? Tasks that a three-year-old handles effortlessly are nearly impossible to capture in explicit instructions.

That is precisely the problem machine learning was built to solve, in a radically different way.

How It Works in Practice

Machine learning: finding structure in data

The core idea behind machine learning is to let the system discover structure in the data on its own, rather than having a programmer define it. You are no longer writing rules, you are feeding the system thousands, millions, or billions of examples and letting it figure out the underlying patterns.

A concrete example: building a spam filter. The classical approach would be to write explicit rules, if the message contains certain suspicious words, block it. The problem is that spammers adapt. The machine learning approach: show the system 100,000 labelled emails (spam or not spam) and let it find the regularities on its own, not just the words, but also the structure, the sending time, the sender’s behaviour, and hundreds of other variables a human expert would never have thought to include.

Mathematically, a machine learning model is a function that takes data as input and produces a prediction as output. Training means adjusting the parameters of that function to minimise errors on the training examples. This process is called optimisation, and the most widely used algorithm is gradient descent, a technique that progressively searches for the combination of parameters that yields the best predictions, like finding the steepest downhill path on a hillside.

Deep learning: when neural networks change everything

Machine learning has existed since the 1950s, but its performance on complex tasks (image recognition, language understanding, machine translation) was long limited. The decisive turning point came in 2012, with a particular architecture called a deep neural network, or deep learning.

An artificial neural network is loosely inspired by the human brain. It is made up of layers of artificial neurons, in practice, simple mathematical operations. Each layer receives data, transforms it, and passes the result to the next layer. What makes it deep is the number of layers: where early networks had two or three, modern models have hundreds.

A useful analogy: picture a manufacturing line in a factory. Raw material (say, an image) enters at one end. Each workstation (layer) extracts progressively more abstract information: the first layer detects edges and contrast, the second simple geometric shapes, the third textures, the fourth partial objects, and so on until the system recognises the complete object. No programmer defined these steps, the network learned them entirely from examples.

In 2012, a team at the University of Toronto presented AlexNet, a deep neural network that slashed the error rate in image recognition from 26% to 15% in a single leap, after years of research had failed to move the needle by more than a fraction of a percent. Historians of technology mark that year as the beginning of the modern AI revolution.

LLMs: when language becomes computable

Large language models are a specific application of deep learning, designed to process natural language, text as it is actually written and spoken in the real world.

They are built on an architecture invented in 2017 by researchers at Google, called the Transformer. Without getting into the mathematics, the central innovation is an attention mechanism: the model learns to weigh the importance of each word based on its context within the text, not just the immediately surrounding words, but the entire passage at once.

This is what allows an LLM to resolve the ambiguities of natural language. Take the word ‘bank’: in ‘she withdrew cash from the bank’, the model connects ‘bank’ to ‘withdrew’ and ‘cash’ to infer a financial institution. In ‘he sat on the bank and watched the river’, it connects ‘bank’ to ‘river’ and ‘sat’ to understand it means a riverbank.

A human reader does this without thinking. For a computer system, it is the result of an attention calculation across the full text.

Training an LLM happens in two phases. First, pre-training: the model ingests astronomical quantities of text (hundreds of billions of words from books, websites, and scientific papers) and learns to predict the next word in a sequence. This deceptively simple task forces the model to develop an implicit understanding of syntax, semantics, logic, and even a form of factual reasoning. Second, human-guided fine-tuning (RLHF): human reviewers rate the model’s responses, and those ratings are used to steer its behaviour toward being more helpful, accurate, and safe.

The LLM parameter explosion, from 117 million (GPT-1, 2018) to trillions (2023–2025).

A parameter is simply a number. But in an AI model, those numbers do something remarkable. Picture an enormous neural network as a web made up of billions of connections. Each connection has a small numerical dial: that is the parameter. It determines how much weight that connection carries, in other words, how strongly one piece of information influences the final output.

Another image that helps: think of a mixing desk with billions of knobs. Each knob makes a tiny adjustment to the sound. One knob alone changes almost nothing. But together, they can produce extraordinarily complex music. Parameters work exactly like those knobs.

During training, the model adjusts these parameters over and over, across millions or billions of examples. Each time it makes an error, it nudges certain settings slightly to do better. Gradually, it learns to pick up on patterns, structures, and relationships in language.

By the end, all those numbers form something like a diffuse memory. The model does not store facts the way a database does. Instead, it encodes statistical regularities, the ways in which words, sentences, and ideas tend to relate to one another.

GPT-3, for example, has 175 billion parameters. That does not mean it knows 175 billion distinct facts. It means its ability to understand and generate text rests on 175 billion interconnected settings. When you ask it a question, the model does not pull up a knowledge card. It routes information through that vast network, where each parameter contributes a tiny amount. That collective dynamic is what produces the answer.

What LLMs are not

One common misconception deserves to be addressed head-on: LLMs do not understand in any human sense of the word. They have no awareness, no beliefs, no emotions. They do not know they exist. They are extraordinarily sophisticated statistical compression machines: they have learned to reproduce the structures of human language with such fidelity that they give the impression of understanding, but that is not what is happening under the hood.

This distinction is not merely academic. It explains why LLMs can produce false statements with the same confident tone as verified facts, a failure mode known as hallucination. It explains why they have no memory between separate conversations. And it explains why their performance depends critically on the quality and diversity of their training data.

The Economic Impact: A Structural Disruption

Investment at an unprecedented scale

Generative AI has become, in under three years, the focus of the largest concentration of capital in recent technology history. In 2024, the four biggest technology companies (Microsoft, Google (Alphabet), Amazon, and Meta) collectively invested more than $220 billion in AI infrastructure: data centres, chips, and networks. That figure exceeds the annual GDP of many national economies.

AI infrastructure investment (Capex), Microsoft, Google, Amazon, Meta (2023 vs. 2024).

Behind these numbers lies a race for computing power. Training the most capable models requires hundreds of thousands of GPUs, manufactured almost exclusively by Nvidia, whose market capitalisation surpassed $3 trillion in 2024, putting it in the same league as Apple and Microsoft. This dependence on a single hardware supplier is one of the structural vulnerabilities of the global AI ecosystem.

Sectors facing disruption

McKinsey analysed the potential impact of generative AI across 63 economies representing 80% of global GDP. The headline finding: generative AI could add between $2.6 trillion and $4.4 trillion in annual value to the world economy, the equivalent of adding the entire UK economy every year. But that value is not evenly distributed across sectors.

Estimated productivity gains by sector from generative AI (McKinsey, 2023).

The most exposed sectors are not the ones you might expect. High-value knowledge work (law, consulting, finance, medicine) is most directly in the frame, because AI excels precisely at processing and synthesising information at scale. A law firm can automate the drafting of standard contracts. A bank can automate risk analysis. A pharmaceutical group can shave years off the time it takes to identify promising drug candidates.

What this means in practice: McKinsey’s estimated productivity gains in legal and consulting reach 45%. That does not mean 45% of lawyers will be laid off tomorrow. It means that smaller teams will be able to produce the same output, which, over time, reshapes the structure of these professions, their training pipelines, and their pay scales.

The employment question: neither apocalypse nor business as usual

Employment is the issue that generates the most fear and the most confusion. It is worth starting with the data before reaching for the scenarios.

McKinsey’s analysis of 800 occupations across 63 economies distinguishes three categories of tasks: those where more than 70% of working time is automatable with current AI (around 5% of all jobs globally); those that are partially automatable (30 to 70% of time, affecting roughly 60% of jobs); and those with low automation exposure (less than 30% of working time).

Job exposure to AI automation, share of global working time affected (McKinsey, 2023).

What this tells us: nearly every existing job will be partially affected, but very few will disappear entirely in the short term. This is not an overnight revolution, it is a gradual but deep transformation of what jobs actually involve. An accountant will spend less time on data entry and reconciliation, more on strategic advice. A doctor will spend less time reading medical images, more time with patients.

Economic history shows that technological revolutions destroy some categories of work and create others that did not previously exist. The Industrial Revolution wiped out hand-weaving and created machine operators. The computerisation of the 1980s and 1990s eliminated most typing pools and created millions of developer and analyst roles. AI follows the same logic, but potentially at greater speed and scale.

The Ethical Stakes: Questions That Cannot Wait

The problem of algorithmic bias

An AI system learns from data produced by humans. If that data reflects historical inequalities or social biases, the model will reproduce them, and sometimes amplify them. This is not a theoretical concern: it is a matter of documented cases.

In 2018, Amazon scrapped an AI recruiting tool after discovering it was systematically downgrading applications from women for engineering roles. The reason: the model had been trained on ten years of Amazon’s historical hiring decisions, which reflected the under-representation of women in the tech industry. The system had learned that ‘strong candidate equals male’, not because a biased developer had programmed it that way, but because the historical data encoded that inequality.

Similarly, AI systems used for facial recognition have shown significantly higher error rates for darker-skinned individuals, a gap of up to 34 percentage points, according to a 2018 MIT Media Lab study by Joy Buolamwini and Timnit Gebru. The cause: training data that under-represented those populations.

The practical implication: in high-stakes applications (hiring, credit, criminal justice, healthcare) a biased AI system can institutionalise and automate discrimination that, when it came from a human being, was at least visible and contestable.

Transparency and the explainability problem

Deep learning systems are, by design, black boxes: no one (not even their creators) can read the weights of a neural network and explain precisely why it produced one output rather than another. This is known as the XAI problem, for Explainability of AI.

That opacity creates a fundamental ethical and legal problem in high-stakes contexts. If an AI rejects your loan application, you have a legal right (under the EU’s GDPR) to receive an explanation. If an AI assists a judge in a parole decision, how do you challenge that decision? If a self-driving car causes an accident, how is liability determined?

The field of explainable AI (XAI) is actively working on these questions, tools like LIME and SHAP can identify which elements of an input most influenced a given decision. But these are post-hoc approximations, not a direct window into the model. The underlying problem remains unsolved.

Power concentration and the governance challenge

Developing and operating frontier AI models requires resources that very few organisations in the world possess: billions of dollars in infrastructure, tens of thousands of GPUs, and teams of hundreds of world-class researchers. In practice, by the end of 2024, fewer than ten organisations worldwide were capable of building frontier models, and nearly all of them are based in the United States or China.

This concentration raises governance questions with no clear precedent. Who decides what values these systems encode? Who controls their deployment? Who is accountable for their failures? Entire regions of the world are using tools whose operating rules were defined unilaterally by a handful of private companies.

Several regulatory responses are taking shape. The European Union adopted the AI Act in 2024, the first comprehensive legal framework for AI, classifying systems by risk level and imposing transparency and audit requirements. The Bletchley Park AI Safety Summit (2023) and its successors brought together more than 28 countries in an attempt to coordinate a shared international approach. These efforts matter, but the pace of technological development continues to outrun regulatory processes by a considerable margin.

The environmental footprint

One dimension that often goes unmentioned: AI is energy-intensive. Training GPT-3 consumed the equivalent of roughly 1,300 MWh of electricity, as much CO₂ as five transatlantic return flights, or a small American town’s daily energy use. Training more recent models like GPT-4 is estimated to be orders of magnitude higher.

Inference (each time the model responds to a query) also draws power. A single ChatGPT request consumes roughly ten times more electricity than a standard Google search. With hundreds of millions of daily active users, the total energy footprint of generative AI is already substantial, and it is growing in step with adoption.

Significant efforts are underway to reduce this footprint: more efficient architectures (distilled models from 2024 approach the performance of much larger models at a fraction of the compute), increasing use of renewable energy in data centres, and inference optimisation. But the tension between capability and sustainability remains a structural challenge.

Key Takeaways

• AI covers a wide range of distinct techniques: classical AI hard-codes explicit rules; machine learning discovers rules from data; deep learning uses multi-layered neural networks; LLMs are the language-specific application of deep learning.
• An LLM is not an ‘intelligence’ in any human sense, it is an extraordinarily sophisticated statistical system that predicts text by drawing on structures learned from hundreds of billions of words.
• The 2012 revolution (AlexNet) was triggered by three converging factors: the availability of large datasets, the processing power of GPUs, and advances in algorithms. Without all three simultaneously, deep learning would not have taken off.
• The economic impact is structural: sectors with a high information-processing component (law, finance, medicine, consulting) are most directly exposed to productivity gains, and to workforce transformation.
• Algorithmic biases are not easily patched bugs, they reflect the inequalities embedded in training data, which requires active vigilance at every stage of design and deployment.
• The concentration of frontier model development within a small number of private organisations, primarily American and Chinese, creates an unprecedented global governance challenge.
• The energy footprint of AI is real and growing, though meaningful efficiency efforts are under way.
• AI will not replace all jobs, it transforms what jobs involve, eliminating certain tasks, amplifying others, and creating entirely new categories of work that did not exist a decade ago.

Real-World Applications

For an employee

A financial analyst at a commercial bank can use AI tools to automate the compilation of market data, the drafting of standard research summaries, and the first pass of portfolio analysis. Work that once took two days can now be done in two hours. The time freed up shifts toward higher-value tasks: contextual interpretation, client relationships, structuring complex scenarios. Understanding how LLMs work allows the analyst to write better prompts, spot the system’s limits (hallucinations on proprietary or recent data), and position their added value precisely where AI still falls short.

For an entrepreneur

A healthcare startup founder can use AI models to analyse clinical data, accelerate literature reviews, and generate functional user interfaces without a full development team. She will need to navigate three real risks, however: patient data confidentiality when submitting records to commercial LLMs (solution: locally deployed models or appropriate data processing agreements); potential model bias against populations under-represented in training data (solution: rigorous clinical validation before deployment); and technological dependence on vendors who can unilaterally change their terms or pricing (solution: a multi-vendor architecture and ongoing due diligence).

For an investor

A venture capital fund manager evaluating opportunities in the AI sector needs to distinguish three levels of the value chain: infrastructure (GPU manufacturers, data centre operators, massive barriers to entry, huge capital expenditure, high margins but long cycles); foundation models (prohibitive training costs for new entrants, global oligopoly, high regulatory risk); and vertical applications (low access cost via API, fragmented market, competitive advantage derived from proprietary data and sector-specific integration). The most accessible opportunities (and the ones least exposed to disruption by the technology giants) sit in specialised vertical applications, particularly those built on proprietary data in markets where global models underperform.

Conclusion

Generative AI is not a technology trend. It is a transformation of the underlying infrastructure of the information economy, comparable in scope to what the internet was in the 1990s, or electrification in the 1920s. Like those earlier revolutions, it will not replace humans. It will redefine what humans do, and where human judgment remains irreplaceable.

From here on, whenever an AI story breaks (a new model launch, a regulation, a funding round, a wave of layoffs, or a breakthrough application) you have the grounding to cut through it. Ask yourself five questions:

• Which layer of the hierarchy is this? (Infrastructure, foundation model, or application?)
• What data was this system trained on, and what biases might that introduce?
• What tasks does it actually automate, and what can it still not do?
• Who controls this system, and who is accountable when it fails?
• What is the economic interest behind this announcement, and what is it not saying?

These five questions will not answer everything. But they will make sure you are never again a passive spectator in a debate about AI, wherever in the world that debate is happening.