Do we need a language model for this problem?

Only if the problem involves language or judgement about unstructured material. Predicting demand, spotting an anomalous transaction or ranking a list are all better served by smaller established approaches over your own structured data, which are cheaper to run, faster, easier to evaluate and considerably easier to explain to an auditor.

How do we test something that learnt from examples?

By measuring it over many cases rather than reasoning about it. Hold back a set of examples it never saw, run them, and count what it gets wrong and in which direction. That is the whole method, and it is why evaluation is a standing activity rather than a stage that finishes.

Why does it fail in ways that seem arbitrary?

Because there is no rule to violate. The behaviour came from patterns in the examples, so a case unlike anything it was shown produces whatever the nearest pattern suggests, and that can look capricious from outside while being entirely consistent from within. Unusual inputs are where this shows up first.

Is the older half worth understanding?

Yes, particularly for anyone buying. A large share of what is marketed with the newer vocabulary is a well-established technique with a fresh interface, which is not a criticism, and knowing that changes what you should pay and what you should expect it to be able to explain about its own decisions.

Glossary

Machine learning (ML)

Machine learning is software that derives its behaviour from examples rather than from rules somebody wrote, which makes the examples the specification and the specification much harder to read.

In plain terms

Software that works things out from examples instead of being told the rules. Show it enough invoices marked as duplicates and it starts spotting duplicates, without anybody writing down what makes one. That is powerful and it moves the specification from something you can read into a pile of examples nobody can inspect.

Why it matters

Because it changes what it means for software to be correct. Conventional software does what it was told and can be checked against the instruction; a system that learnt from examples does what the examples implied, which nobody wrote down and nobody can read back. Testing therefore becomes measurement over cases rather than reasoning about rules.

How it works

The examples are the specification, which is the single most useful thing to understand about the whole field. Whatever pattern is present in them becomes the behaviour, including patterns nobody intended and nobody noticed, so the quality of the outcome is bounded by the quality and the representativeness of what it was shown.

Correctness becomes statistical rather than logical. There is no line to point at that decides a case, so a system is right most of the time and wrong sometimes in ways that do not follow a rule, which is why evaluation is a measurement over many cases rather than an argument about a specification.

The older half of the field is still doing most of the quiet work. Classification, forecasting, ranking and recommendation predate the current wave and run inside pricing, logistics, fraud detection and stock control in almost every large organisation, usually without anybody calling it AI.

Language models belong to this field rather than replacing it, which matters when choosing a tool. A great deal of ordinary business prediction is better served by a small, well-understood, cheap model over your own tabular data than by a general assistant, and reaching for the assistant because it is the thing in the news is a common and expensive mistake.

Where the specification lives

The trade here is worth stating without either enthusiasm or alarm, because it is the whole reason this field feels different to work with. Rules are limited and legible: you can only encode what somebody can articulate, and in exchange anybody can read what was encoded. Examples are the reverse. They capture things nobody could have articulated, which is why this approach solved problems that resisted decades of rule-writing, and the price is that the resulting behaviour is not written down anywhere. Every practice that has grown up around these systems follows from that price. Evaluation over held-back cases exists because you cannot inspect the rule. Measuring for bias exists because a pattern in the examples becomes a pattern in the behaviour silently. Monitoring in production exists because there is no line to review when something changes. None of that is ceremony; it is what replaces being able to read the specification.

Seen in the wild

The hub at the centre of open machine learning, where the models, the datasets and the tooling the ecosystem shares all live together.
Hugging Face
A conversational data analyst that runs real statistical methods, from regression to time-series forecasting, with the code behind them visible.
Julius AI
A database platform whose AI classifies records and pulls details out of the documents and operational data a workspace collects.
Airtable AI

Common misconceptions

People assume

It means the same thing as AI.

In fact

It is the part of the wider field that learns from examples, and it is where essentially everything currently sold as AI actually comes from. The distinction matters less in conversation than in procurement, where a great many products described with the broader word are doing something quite specific and well established.

People assume

It is all language models now.

In fact

Most of it is not, and most of what runs a business is not. Forecasting demand, scoring a transaction and ranking a list are all handled by smaller, older, cheaper approaches over structured data, and reaching for a general assistant to do them is usually slower, dearer and less accurate.

People assume

You can inspect what it learnt.

In fact

Not in the way you can read a rule. The behaviour lives in numbers derived from the examples rather than in statements anybody wrote, so understanding why a particular case went the way it did is a research exercise rather than a matter of finding the relevant line.

Telling them apart

Machine learning vs Deep learning

Machine learning

The whole field of learning from examples.

Deep learning

One family within it, built from many layers.

All deep learning is machine learning and most machine learning running in a business is not deep learning, which is the opposite of the impression the news gives.

Questions

Do we need a language model for this problem?: Only if the problem involves language or judgement about unstructured material. Predicting demand, spotting an anomalous transaction or ranking a list are all better served by smaller established approaches over your own structured data, which are cheaper to run, faster, easier to evaluate and considerably easier to explain to an auditor.
How do we test something that learnt from examples?: By measuring it over many cases rather than reasoning about it. Hold back a set of examples it never saw, run them, and count what it gets wrong and in which direction. That is the whole method, and it is why evaluation is a standing activity rather than a stage that finishes.
Why does it fail in ways that seem arbitrary?: Because there is no rule to violate. The behaviour came from patterns in the examples, so a case unlike anything it was shown produces whatever the nearest pattern suggests, and that can look capricious from outside while being entirely consistent from within. Unusual inputs are where this shows up first.
Is the older half worth understanding?: Yes, particularly for anyone buying. A large share of what is marketed with the newer vocabulary is a well-established technique with a fresh interface, which is not a criticism, and knowing that changes what you should pay and what you should expect it to be able to explain about its own decisions.

Key takeaways

The examples are the specification, and the specification cannot be read back.
Correctness becomes a measurement over cases rather than an argument about rules.
Most machine learning running a business is not the kind that is in the news.
Language models are part of this field, not a replacement for it.