Three things, one name
When someone says AI today, they probably mean one of the three below. Most misunderstandings come from the other person meaning a different one.
1. The field of research. The branch that has asked since the 1950s how to make a machine do work that used to require a person. Everything belongs here, from chess programs to today's chatbots.
2. Predictive systems. These create nothing new: they classify, forecast and recommend. Your bank's fraud screening, a shop's recommendation engine, the spam filter, demand forecasting. They have been with us for twenty years — we just didn't call them AI. At plenty of companies this is what delivers the measurable return, while everyone talks about the third one.
3. Generative systems. These produce new content: text, images, audio, video, code. This is what broke through in 2022, and today it is almost always what AI means.
This blog is mostly about the third. But it is worth knowing the second exists — for a lot of business problems it is the cheaper and more reliable answer.
Bite-Sized AI · 001
Three things, one name
01
The research field
A branch of science running since the 1950s. It covers everything from chess programs to today's chatbots.
02
Predictive systems
They create nothing new: they classify, forecast, recommend. Fraud detection, recommendations, demand planning. Twenty years old.
03
Generative systems
They produce new content: text, images, audio, video, code. This is what took off in 2022 — and what we now call AI.
Most misunderstandings start when two people mean two different rows.
What AI is not
AI is not a thinking being, not a database and not a search engine. It does not know the answer, and by default it is not looking it up anywhere either. Rule those three out and what remains is surprisingly simple.
How it becomes a working system
The principle is the same across all three kinds: the machine derives the rule from examples instead of someone writing the rule in advance. That is machine learning. You do not program in "spam often contains the word prize" — you show it a hundred thousand emails and it finds the pattern.
The modern, many-layered version of this is deep learning, and it is what large language models — LLMs — are built from. Training here means that the model's internal numbers, the weights, get tuned until they predict what comes next more and more accurately.
The one sentence that explains everything
The model does not know the answer. It predicts what probably comes next.
That is it. And everything surprising about AI follows from that single sentence — in both directions.
The good direction: it handles tasks nobody trained it for specifically. The bad direction: it will confidently produce data that does not exist when it has no better guess. That is hallucination, and it is not a bug — it is the other side of the same mechanism. You cannot switch it off without taking the useful part with it.
An example that measures the difference
Two requests, same tool:
1. "Summarise this email in three sentences." The email is right there in front of it. The model does what it is good at: transforming text that already exists. Reliable.
2. "What was the euro exchange rate on 30 September last year?" The model did not look it up. It produced a plausible number. It might be right. You will not be able to tell by looking.
Same tool, same interface, two completely different levels of reliability. The difference is not the difficulty of the question — it is whether you gave it the material to work from.
Why now?
The method is not new: the idea of neural networks goes back to the sixties. Three things changed over the last decade.
- Data. The internet produced enough text to learn from.
- Compute. Graphics cards became cheap and powerful enough to train on.
- Architecture. The 2017 transformer let a model see a whole passage at once instead of crawling through it word by word.
Those three together carried the technology past the point where it stopped being a curiosity and became a tool.
Two common misconceptions
"It learns from me as we talk." By default, no. The model is frozen at the end of training. What you tell it inside a conversation it only knows for as long as that conversation lasts — that is the context window. Several tools have a separate memory feature, but that does not change the model either: it puts notes in front of it at the start of each conversation.
"If it can do this, it can do anything." No. What you use today is narrow AI: very good at certain patterns, with no goals, no experience and no accountability of its own. A general system, human-level at every intellectual task — AGI — is still a theoretical idea, whatever anyone is advertising.
Beyond answering
One important development: today's systems do not only answer, they act. They can search, write files, run programs, fill in a form, then check the result and try again. We call these AI agents, and commercially this is the real shift: a chatbot gives you advice, an agent does the job.
What to take away
Before you hand a task to AI, ask two questions:
- Do I have the material it can work from?
- Would I recognise a wrong answer?
If the answer to both is yes, it will probably work. If either is no, the correct answer and the plausible answer will be indistinguishable — and of the two, that is the more expensive mistake.