Why AI? A director's foundation

Strip away the marketing and artificial intelligence is one thing: a machine that finds patterns in examples, then applies those patterns to something it has never seen before. That is the whole idea. Everything else is engineering around it.

An actual chip die under a macro lens. This is what “billions of parameters” physically sits on. Photo: Fritzchens Fritz, CC0.

This tells you what AI is for. Not a thinking machine, not a database: a pattern engine. The right question on any AI project never changes: is there a real pattern in this data, and do we have enough examples of it?

Half the failed AI projects in any company sit in the top-left box: an ordinary software problem, wearing an expensive costume.

Humans have always done this. A buyer with thirty years in a market can look at a supplier and tell you something is wrong without being able to say what. That is a pattern learned from thousands of examples, held below the level of language. AI does the same thing, on more examples than one lifetime allows, and with no ability to explain itself either.

The core idea is not new. In 1943 the neuroscientist Warren McCulloch and the logician Walter Pitts published a mathematical model of a single artificial neuron. By the 1980s, researchers including Geoffrey Hinton had working methods for training networks of them. Most of the mathematics behind what you are reading about was settled before the web existed. So the honest question is not what did they invent but what changed.

Four things have to be true at once for this to work. Until recently, no era had all four.

The fourth is the unglamorous one: infrastructure. Storing, cleaning, labelling and moving that volume of data is most of the real work, and it is where most corporate AI programmes actually fail. Not at the model. At the plumbing.

• 2006
  Cloud storage becomes a utility
  Keeping enormous datasets stops being a capital project and becomes a line item. Data that used to be discarded starts being kept.
• 2012
  AlexNet wins ImageNet, by a landslide
  Three Toronto researchers, Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton, enter the annual ImageNet contest with a large neural network. It beats the next-best method by roughly 10 percentage points. The field reorganises around the approach within two years.
• 2016
  Graphics chips become the engine
  Hardware built for rendering games turns out to be near-perfect for the arithmetic neural networks need. Training times collapse from months to days.
• 2017
  “Attention Is All You Need”
  Eight researchers at Google publish the transformer architecture under that title. It handles language and context far better than anything before it, and keeps improving as it is made bigger. Every system you have heard of since, GPT, Gemini, Claude, descends from this one paper.
• 2022–2026
  General-purpose systems reach the desk
  Models stop being one-task tools built by specialists and become general instruments any employee can use without permission, budget or training. This is the part that reached your organisation.

This is why treating AI as a passing enthusiasm is a category error. Enthusiasms fade because the thing underneath them was never real. Here the underlying conditions, cheap computation and abundant data, are not reverting. They compound.

You do not need to build one. You need to know roughly what is happening inside, because that is what lets you tell a sound plan from an expensive one.

A neural network is a long chain of simple arithmetic. Data enters at one end. It passes through layer after layer, each one multiplying and combining the numbers slightly differently. A result comes out the other end: a category, a score, a next word.

Training happens on rooms full of these. Photo: PDC, CC BY-SA.

Those parameters start as noise. Training is the process of correcting them, and it is far less clever than it sounds:

1. Show the system one example where you already know the right answer.
2. Let it guess. Early on the guess is close to random.
3. Measure how wrong it was. That measurement is called the loss.
4. Nudge every parameter a fraction in whichever direction would have made the guess less wrong.
5. Repeat several million times. There is no fifth step. That is the entire method.

Two consequences follow directly, and both are commercial rather than technical.

First, the system can only be as good as the examples you showed it. If your historical decisions were biased, the model learns the bias precisely and applies it consistently at scale. It does not average out human error. It industrialises it.

Second, nobody can point to where the knowledge sits. It is spread across billions of parameters with no labels. This is the real reason AI systems are hard to audit, and why any vendor promising full explainability deserves a second look.

Most of the authority gap in an AI review is vocabulary, not intelligence. These are the terms that come up, what they actually mean, and what each one should make you ask.

The last question is the most useful one you own. A good team will answer it precisely. If the answer is vague, the project is decoration.

The ten terms above are the foundations. These eighteen are what floated into the room since generative AI arrived — darker tiles are the ones where getting it wrong costs more than looking uninformed. Each links to one free (or clearly marked paid) course if you want the fuller version.

Knowing the limits is worth more than knowing the capabilities. The capabilities are advertised to you daily. The limits are where the money is lost.

The causation point is the expensive one, so it is worth an example.

Which is the honest summary of the whole technology. It finds what travels together. It has no idea what causes what. Supplying the why is the part that stayed human, and it is the part directors are paid for.

A technology becomes a management problem, rather than a technical one, at the moment it changes what a company can be asked to do. AI crossed that line when it became cheap enough for any employee to use without approval.

Three consequences reach you directly.

None of that requires a tool, a vendor or a budget. It requires understanding why the technology behaves as it does, which you now have.

This guide gets you through the door in ten minutes. Those two are the multi-week version, taught by the people who did the underlying research.