01What generative AI actually is

AI, in one sentence

Artificial intelligence is software that learns patterns from large amounts of data and uses them to perform tasks that would otherwise require human judgement. Four phrases do the work in that sentence. Learns patterns: an AI system is not programmed step by step; it derives rules from examples. This has consequences: a model cannot be audited the way a conventional program can. Large amounts of data: no data, no AI. Data quality and quantity are the actual inputs, not compute. Performs tasks: AI is task-specific. There is no such thing as general-purpose AI in production in 2026, in spite of what the term "AGI" might suggest. And human judgement: the framing that sells, but a narrower technical reality.

The useful question is not whether AI will become conscious. It is which specific tasks it can perform reliably enough to be worth deploying in your organisation. That question has concrete answers, and we will come to them.

A short history

The field of artificial intelligence is older than commercial television. In 1943, Warren McCulloch and Walter Pitts published the first mathematical model of a neuron. In 1950, Alan Turing proposed the test that now carries his name. By 1958, Frank Rosenblatt had built the perceptron, the first neural network to physically exist. All of this happened before the invention of the integrated circuit.

The timeline matters for two reasons. The first is that it has already failed twice. In 1969, Marvin Minsky and Seymour Papert published Perceptrons, a book that mathematically demonstrated the limits of the then-dominant neural network architecture. Funding collapsed. The field entered what is now called the first AI winter, which lasted into the mid-1980s. A second winter followed in the early 1990s after the expert-systems boom failed to deliver on its promises. Both winters lasted years. Both followed a cycle of overpromising, under-delivering, and loss of credibility, a cycle that is worth remembering as we discuss the current one.

The second reason is that the modern breakthroughs happened in the stretch between 2006 and 2022, and even that period is sixteen years long. Deep learning emerged in 2006 when Geoffrey Hinton and colleagues showed that neural networks with many layers could be trained in practice. Transformers, the architecture behind every current large language model, were introduced in a 2017 paper from Google titled Attention Is All You Need. ChatGPT, the product that made the public aware that any of this had happened, was released at the end of 2022. There was no overnight revolution. There was a long, continuous accumulation of ideas, data, and compute that eventually crossed a visibility threshold.

1989: an AI that already shipped

One milestone on the timeline deserves to be told in full, because it corrects a common assumption about what AI costs and what AI needs. In 1989, Yann LeCun, then at Bell Labs, built a neural network called LeNet that could recognise handwritten digits. The National Cash Register Corporation deployed it to process bank checks. By the mid-1990s, LeNet-based systems were reading an estimated ten percent of all cheques written in the United States. This was not a research demonstration. It was a commercial system running in production, handling real money.

The hardware it ran on was the Intel 80486, a processor less powerful than a modern smartwatch. The model had roughly sixty thousand parameters. For comparison, a GPT-class model in 2026 has hundreds of billions of parameters and runs on data centres that consume the electricity of small cities. The ratio between the two is on the order of ten million. Both are AI. Both are useful. One is cheap, auditable, and runs on hardware a bank already owns. The other is expensive, opaque, and runs on hardware a bank must rent from a hyperscaler.

This matters strategically, not nostalgically. One of the mistakes most commonly made in current AI programmes is to assume that useful AI must look like ChatGPT: that it must be a large general-purpose model called through an API. For a well-defined task on a well-defined data domain, a small specialised model will often work better, cost a fraction of the price, run on your own infrastructure, and be subject to no vendor's pricing decisions. The term I will use for this is frugal AI. It is a strategic choice, not a compromise. We will return to it in Chapter 02 when we discuss the cost trajectory of the frontier APIs, and in Chapter 04 when we discuss the capability patterns to deploy.

The one-line summary: working AI has been running in production for thirty-five years. What changed in 2022 is the interface, not the intelligence.

The paradigm shift of 2022

What did change in 2022 is nonetheless substantial, and the word "paradigm shift" is not used lightly here. The distinction worth drawing is between two kinds of AI. One is discriminative; the other is generative. The first has been in production for decades. The second is what became mainstream in late 2022.

A discriminative model classifies, predicts, or decides. It looks at an input and produces a label. Is this credit-card transaction fraudulent? Does this chest X-ray show a tumour? Which of these ten search results is most relevant to the query? These are discriminative tasks, and they are mature. They power fraud detection, credit scoring, medical imaging, demand forecasting, search ranking, industrial quality control, and most of the systems that quietly make large organisations function. Billions of dollars of value have already been captured here, over the past two decades. Nothing about generative AI replaces this. Discriminative AI will continue to do most of the actual work.

A generative model produces new content. Given a prompt, it writes text, code, or an email; it drafts an image or a piece of audio; it summarises a document; it answers a question. The output does not exist before it is generated. This is a different kind of problem, and it is the kind of problem that could not be solved reliably at scale before 2022. The transformer architecture from 2017, trained on internet-scale text and then fine-tuned to follow instructions, is what crossed the threshold.

Conflating the two is the source of most misplaced expectations. Executives who have seen ChatGPT draft a convincing paragraph conclude that AI must now be able to do everything. It cannot. And teams that have been told AI will "transform" every process often try to apply generative models to tasks for which a discriminative model would be simpler, cheaper, more auditable, and more accurate. The distinction is small to state, important in practice, and often skipped by vendors with reasons to blur it.

Agents: what makes a language model actually useful

Agents are the single most consequential piece of jargon of 2025 and 2026. They are also the thing that turns generative AI from a curiosity into something a business can use. Left to themselves, large language models accept text and produce text, and that is all they do. No matter how good the reasoning gets, a model sitting alone cannot read your customer database, cannot send an email, cannot open a ticket, cannot touch anything in your business. Agents are what connects it to the world.

The clearest way to think about this is through a very old analogy. The model is the brain. An agent gives it sensors and actuators: the equivalent of eyes, ears, and fingertips on the input side, and arms and legs on the output side. On the input side, an agent lets the model read your documents, query your database, call your APIs, examine a webpage, listen to a phone call, watch a screen. On the output side, it lets the model send an email, write to a file, call an API, book a meeting, open a ticket, commit code. The language model sits in the middle, reasoning over what the sensors report and deciding which actuator to invoke next. The anatomical metaphor eventually breaks down (software agents are not literally embodied), but as a mental model it is the one that holds up longest in practice.

What makes an agent properly "agentic" is the loop between those two sides. The model observes the current state through its sensors. It decides what to do next. It calls one actuator. It reads the result. It re-plans. Then it does the next thing. One reasoning step on its own is easy. Chaining a hundred of them reliably: that is where the real engineering lives, and that is where most agents currently fail.

A robust production agent is closer in complexity to a space-shuttle control system than to a chatbot. Impressive demos are cheap. Reliability is not.

The honest picture is twofold. On the one hand, without agents, a language model is a brain in a jar, able to reason, unable to act, and therefore of limited value for most enterprise workflows. Agents are how the technology becomes genuinely useful. On the other hand, building an agent that works in production, one that handles edge cases, recovers from errors, respects a cost ceiling, escalates to a human when necessary, and does not quietly do the wrong thing a hundred times before anyone notices, is one of the hardest engineering problems in current AI. The gap between a demo agent and a production agent is enormous, and most of what is currently sold as "agentic" is closer to a demo than to a product.

The practical consequence, for a CEO or CIO evaluating a pitch: when a vendor proposes an agent, ask three questions. What sensors does it have: what can it actually read? What actuators does it have: what can it actually do? And what happens when its planning goes wrong: what are the error-handling, human-escalation, and cost-cap mechanisms? If the answers are not clear and concrete, the agent is a demo.

These are the two shapes in which generative AI becomes useful: either a smaller specialised model embedded into a specific workflow, or a large general model surrounded by agent machinery. Both are feasible. Both are hard. Neither is magic.