OpenAI is in the middle of twelve days of announcements, one per day. Google launched Gemini 2.0 last week. Every morning, our inbox receives a new "AI-powered" product announcement. Yet when we ask people who use these tools at work what machine learning actually is, most of them are unable to answer.
This is not a criticism. No one has explained it to them.
It has also been a significant year for the field. In October, the Nobel Prize in Physics was awarded to John Hopfield and Geoffrey Hinton for foundational work on machine learning with artificial neural networks. The following day, half of the chemistry prize went to Demis Hassabis and John Jumper for AlphaFold, a machine learning system that predicts the shape of proteins. Two Nobel prizes were awarded in two days. This article offers a plain-language explanation, without mathematics, for business owners who would like to understand how these tools work.
What is machine learning?
Machine learning is learning from examples. Conventional software follows rules written by a person: if an invoice is more than 30 days overdue, send a reminder. Such software is clear and predictable, and exactly as capable as the person who wrote the rule.
Machine learning reverses this approach. Instead of writing the rules, you show the computer a large number of examples and let it identify the patterns itself.
The classic example is a spam filter. No one can write a rule for every type of spam. A system is therefore shown thousands upon thousands of emails, each labelled "spam" or "not spam", and it gradually determines which words, senders, links and patterns tend to be associated with spam. It is then given an email it has never seen, and it makes a prediction.
In short: examples go in, patterns are learned, and predictions come out.
What do the key terms mean?
- Training: showing the system examples and letting it adjust itself until its predictions improve.
- Model: the result of training. A large file of numbers that contains the patterns the system has identified. When people refer to "the model", this is what they mean.
- Data: the examples, such as emails, photos, sales records or sentences.
- Neural network: a widely used method of building a model, loosely inspired by the way brain cells connect. This year's physics prize recognized work in this area.
Where do ChatGPT and other language models fit?
ChatGPT, Claude and Gemini are also machine learning systems. They operate at an enormous scale, but rest on the same basic principle.
They were trained on vast amounts of text with one main task: predicting the next word, repeated billions of times. To perform this task well, a model must acquire grammar, facts, writing styles and the way an argument usually unfolds. The companies then carry out additional training with human feedback to make the models helpful and courteous in conversation.
This explains much of their behaviour. They excel at producing text that sounds correct, because that is precisely what they were trained to do. Sounding correct and being correct are, however, different things. When a model invents a statistic or a source, it is not deliberately lying. It is producing the kind of sentence that usually appears in that position.
We have been using these models since GPT-3, and we published a beginner's guide to ChatGPT last year. If there is one concept we recommend remembering, it is this principle of "predicting the next word". It explains both why these models are so useful and why their output must be verified.
Why does poor data produce poor results?
A model knows only what its examples have taught it. If the examples are incorrect, incomplete or unbalanced, the model will be as well, and it has no way of recognizing the problem.
A well-known case illustrates this: Amazon reportedly abandoned an experimental hiring tool a few years ago after finding that it penalized résumés from women. The tool had been trained on the company's past hiring, which consisted mostly of men. The model learned the pattern it was shown, and that pattern was the problem.
Smaller versions of this situation occur constantly. A translation tool trained mostly on European French produces text that sounds European. A tool trained mostly on data from large urban businesses may give advice that does not suit a village of 1,000 people. A sales forecast built on the last few years, including the pandemic, will learn some unusual patterns.
This is largely what is meant by bias in AI. It is less a matter of bad intentions than of unbalanced examples.
Does your own data matter?
Yes. If you use tools that learn from your business data, such as booking predictions, ad targeting or product recommendations, the same principle applies. Disorganized and incomplete records produce unreliable predictions. Cleaning up your data is not a particularly engaging task, but it is often the most useful AI project a small business can undertake.
How can you tell what an AI tool was trained to do?
This is the practical aspect of the subject.
Every AI tool was trained for a specific purpose. A language model was trained to write plausible text and hold a conversation; an image tool, to create images from descriptions; a fraud detector, on past fraud cases. These tools perform well at what they were trained for, and often poorly at anything else.
Before entrusting a task to a tool, we recommend asking three simple questions:
- What was this tool trained to do?
- What kind of examples was it trained on, and do they resemble my situation?
- How would I know if it is wrong?
A language model used for accounting is working outside its training. It will produce figures that resemble accounting, and those figures must be verified. A model drafting a courteous reply to a customer is working well within its strengths. Even so, the reply should be reviewed before it is sent.
How can you learn more over the holidays?
If the holidays offer a few quiet evenings, Google's free course is a good next step. It is designed for people who want to understand the concepts, and the early modules require little mathematics: Google's Machine Learning Crash Course.
Some people learn more effectively with their own tools and their own business in front of them. A small team can benefit considerably from a few hours of hands-on AI training in the language in which it works.
A new generation of tools with new names is expected next year, possibly as early as February. Rather than attempting to follow each one, we apply the same three questions to all of them.