Learn to Train AI, Not Just Prompt It

In short

There is a line running through everything people call AI work, and it falls between using a system and building one. The UK government's introduction to AI assurance names six techniques for checking an AI system: risk assessment, algorithmic impact assessment, bias audit, compliance audit, conformity assessment and formal verification. All six assume a system with inputs you selected and outputs you can measure against a benchmark, which is exactly what a person who only prompts does not have. Learning to train models puts a learner on the side of the line where work can be inspected, measured and defended. We teach it with real data and small models, to learners from about twelve upwards and to adults, in groups of five to ten at one level or one to one. The first lesson is free, then USD 100 a month in a group or USD 150 one to one.

Start here

Three courses about building, not using

Each one ends with a model the learner trained, tested and can explain. Cards open the full syllabus.

TRAIN / 01

AI and Machine Learning for Teens

Data gathered, a model trained, evaluated against a held-out set, and then broken on purpose to find where it fails.

Open the syllabus →

TRAIN / 02

Data Science with Python

The part everyone skips: cleaning, questioning and understanding the data before a model ever sees it.

Open the syllabus →

TRAIN / 03

Generative AI, Properly

For adults and older students who need what is happening underneath rather than a tour of an interface.

Open the syllabus →

Six ways to check

What the government says you can do to an AI system

From the Department for Science, Innovation and Technology's Introduction to AI assurance, published 12 February 2024. These are its own definitions, shortened only where marked.

The six assurance techniques, as DSIT defines them
Technique	What the guidance says it does	What it needs to exist
Risk assessment	Used "to consider and identify a range of potential risks that might arise from the development and/or deployment of an AI product/system"	A system with a known development history
Algorithmic impact assessment	Used "to anticipate the wider effects of a system/product on the environment, equality, human rights, data protection, or other outcomes"	A stated purpose and a population it affects
Bias audit	"Assesses the inputs and outputs of algorithmic systems to determine if there is unfair bias in the input data, the outcome of a decision or classification made by the system"	Access to the input data. This is the decisive one
Compliance audit	"Involves reviewing adherence to internal policies, external regulations and, where relevant, legal requirements"	A record of how the system was built and operated
Conformity assessment	Demonstrates "whether a product or system meets relevant requirements, prior to being placed on the market. Often includes performance testing"	Requirements written down in advance
Formal verification	"Establishes whether a system satisfies specific requirements, often using formal mathematical methods and proofs"	A specification precise enough to be proved against

Read the third column down

Every entry needs something a person who only prompts does not possess: the training data, the build record, the stated requirements, the specification. The guidance describes a "toolbox of different products, services, and standards", and none of the tools in it fit around a text box.

The three verbs underneath

The guidance organises all of this under three headings. Measure means "gathering qualitative and quantitative data on how an AI system functions, to ensure that it performs as intended". Evaluate means assessing risks and impacts, including "against agreed benchmarks set out in standards and regulatory guidelines". Communicate means being able to say what you found, inside the organisation and outside it.

Measure, evaluate, communicate. A student who has trained a model can do all three about it. A student who has only prompted one can do none.

Benchmarks are the crux

The guidance is specific that teams should "build in rigorous software testing and performance assessment methodologies with comparisons to clear performance benchmarks".

A benchmark requires a held-out set of examples the model has never seen. Creating one is a ten-minute idea and the single most important habit in machine learning, because without it a model that has simply memorised looks identical to one that has learnt.

Source, read 20 September 2026: Introduction to AI assurance, Department for Science, Innovation and Technology, published 12 February 2024. The guide states that it "focuses on the underlying concepts of AI assurance rather than technical detail", and it is written for organisations rather than for learners; the application to learning here is ours.

Which side of the line

Using a system, or building one

Both are legitimate. They are not the same activity, they do not build the same knowledge, and only one of them leaves evidence.

The same word, two different jobs
	Prompting a system	Training a system
What you control	The wording of a request	The data, the structure, the objective and the stopping point
What you can measure	Whether this answer looked reasonable	Accuracy against examples the model has never seen
What you can audit	Nothing in the guidance applies	Inputs, outputs, bias, conformity, and the build record
What failure teaches	Try different words	Which part of the data or the design was wrong
What transfers	Familiarity with one product	The behaviour of the whole family of systems, product by product
What you can show	A conversation	A model, its data, its results and its limits

This is not snobbery

Prompting well is a real and useful ability, and we use these tools in lessons. The point is narrower: it is not a substitute for knowing what the system is doing, and it cannot be checked by anybody.

It is about evidence

A learner who trained something can hand over the data, the code and the results. That is what a coursework marker, an interviewer or a regulator can actually examine.

And about failure

A model that fails tells you where. A prompt that fails tells you to rephrase. Only one of those two is a lesson.

The bias audit row deserves a sentence on its own, because it is the one that most surprises students. The guidance defines it as assessing inputs and outputs to find unfair bias "in the input data, the outcome of a decision or classification made by the system". Input data first. A learner who has chosen a training set with their own hands understands immediately why that is the first place to look, and a learner who has only ever typed into a box has no way to reach the question at all.

What the wider evidence says about AI literacy points the same way, and the page on learning to think sets that out: in the most cited framework, nine of seventeen competencies are about how these systems work rather than how to operate them.

How training is taught

Data, model, benchmark, failure

Four stages that fit in a term and scale from a twelve-year-old to an adult professional. The only thing that changes is the dataset.

The first two

Data comes before anything. Where did these examples come from, what is missing from them, and who would be poorly served by a system trained on this? Doing this badly is the commonest failure in machine learning and the one nobody notices until much later.

Then the model. Small and simple to begin with, because a learner needs to be able to hold the whole thing in their head and say what each part is for. A big model that works is worth less to a student than a small one they can explain.

And the two that matter most

The benchmark. Split the data before training, hold some back, and only ever judge the model on examples it has never seen. This one rule separates measurement from wishful thinking, and it is what the guidance means by comparing against clear performance benchmarks.

Then failure, deliberately. Find the input that makes the model confidently wrong. Work out whether the fault was in the data, the design or the question. A learner who has done this stops believing confident outputs, permanently.

What a learner can demonstrate at each stage
Stage	What they produce	Which assurance idea it mirrors
Data	A described dataset, with what is missing written down	Bias audit, which begins with the input data
Model	A trained model they can explain part by part	The build record a compliance audit would ask for
Benchmark	A result on examples the model never saw	Performance testing against a clear benchmark
Failure	A documented case where it fails and why	Risk assessment, done at the scale of a classroom

That final column is not a claim that a school project is an audit. It is a claim that the thinking is the same shape, and that a learner who has done the small version recognises the large one when they meet it at work. Beginning in the right shape costs nothing extra.

Younger learners do all four stages with data they collect themselves, usually drawings, which is described in detail on the AI classes for kids page. Older students use published datasets, and adults usually bring something from their own work, which is where the questions get sharpest.

One honest limitation. We teach how these systems are built, trained and evaluated at a scale a learner can hold; we do not train large language models, and nobody in a classroom does. What transfers from the small version is the reasoning, not the hardware.

Progression

Four rungs into building rather than using

Placement follows what a learner has actually built, which is usually less than they expect and easy to fix.

What a learner can build at each rung
Rung	What they can do	What they still take on trust
1. Describe the data	Say where a dataset came from and what is missing from it	That the training process will do something sensible with it
2. Train something	Produce a working model from labelled examples	That the accuracy figure means what it appears to mean
3. Hold data back	Judge the model only on examples it has never seen	That the held-out set represents the real world
4. Break it deliberately	Find inputs that defeat it, and say why they do	Very little, which is the point of getting here

Rung three is where it becomes real

Until a learner holds data back, every result they produce is untrustworthy in a way they cannot see. It is a single habit and it changes everything downstream.

The whole topic sequence, including the Python that makes it possible, is on the coding roadmap.

Adults often start at rung one

Professionals who use AI daily frequently have never described a training set. That is not a gap in intelligence, it is a gap in exposure, and it closes fast.

Where anyone starts is settled in the free lesson rather than assumed from a job title or a school year.

The catalogue

Nine courses about building AI systems

Grouped by how much programming a learner has before they begin. Full syllabuses sit behind the cards.

I

Before the model

The language and the data

UK / TR1 / 01

Python in full

The language everything here is written in, taught by predicting output before running it.

Open the syllabus

UK / TR1 / 02

Data science with Python

Cleaning and questioning real datasets, which is most of the work and all of the risk.

Open the syllabus

UK / TR1 / 03

Databases and SQL

Where data actually lives before anybody trains anything on it.

Open the syllabus

II

Training models

The core of this page

UK / TR2 / 01

AI and machine learning for teens

Train, hold data back, measure honestly, then break it on purpose.

Open the syllabus

UK / TR2 / 02

Visual machine learning

Models built and compared visually first, so the ideas land before the notation does.

Open the syllabus

UK / TR2 / 03

AI literacy for kids

The same four stages for younger learners, with examples the child gathers themselves.

Open the syllabus

III

Going further

For students and adults heading somewhere

UK / TR3 / 01

Generative AI, properly

The architecture and the training process behind the tools, for people who need more than a feature list.

Open the syllabus

UK / TR3 / 02

Data science, complete

The full route for adults with a destination, from spreadsheets to deployed models.

Open the syllabus

UK / TR3 / 03

Python and AI automation

Putting trained systems to work, with the checks that catch them failing quietly.

Open the syllabus

How lessons work

Hold data back, then try to break it

India keeps one clock all year, which leaves the UK five and a half hours behind from October and four and a half from March. Pick a slot in UK time and it is yours.

After school

Where most school-age learners land.

Evening

Older students and working adults, after the day winds down.

Weekend morning

A longer block, which suits training a model from start to finish.

Data first, always

Where it came from and what is missing, before a single model is trained. Skipping this is how projects fail quietly.

Small models on purpose

A model a learner can explain part by part beats a bigger one that works for reasons nobody in the room can state.

A held-out set every time

Judged only on examples it has never seen. Without this, a memorised model is indistinguishable from a good one.

Deliberate failure

Every project ends with the learner hunting for the input that defeats their own model.

Written limits

Each finished model comes with a short note on what it cannot do, which is the habit the assurance guidance is built on.

Groups by level

Five to ten learners at one stage, comparing results and arguing about whose benchmark was fairer.

Student work

Models our learners trained

Four published projects. The student labs page holds more.

AI and ML

NutriLife

An AI nutrition coach that reads what you eat and works you toward a target.

by Bhavya · Open it

AI and ML

Misti

A chatbot that answers mathematics and programming questions, built and deployed by a student.

by Harshit · Open it

AI and ML

GuardianX

An assistant that helps a young person recognise unsafe situations online.

by Vivaan · Open it

Web app

SkyCast

A weather forecasting site with live conditions for any location.

by Krish

Fees

What it costs

One charge a month, set in US dollars for every country but India. Try the free lesson first; billing begins only when a course and a weekly time are fixed.

Free first class

USD 0

no card required

A genuine lesson, not a sales call
A straight view of the right rung
Only a phone number needed

Book it

Group batch

USD 100

a month, billed in US dollars

Five to ten learners at one rung
The same teacher for the course
Weekly written feedback on real work
A certificate at the end of a syllabus

Start here

One to one

USD 150

a month, billed in US dollars

One learner and one teacher
Built around a specific target or dataset
Suits adults bringing work from their own job

Enquire

What families say

Rated 4.9 across 547 Google reviews

Straight from Google, unedited, from families nobody prompted.

★★★★★

"The one step solution for my son. Modern Age Coders make learning coding so simple that kids love it. The teachers explain complex concepts clearly with practical exercises and interactive content."

Ria Mukherjee

Parent

★★★★★

"Modern Age Coders has been a game-changer for me. I struggled to grasp IT concepts and coding before joining, but their classes transformed everything. I can now confidently write complex programs with ease."

Samriddha Mondal

Student

★★★★★

"One of the most wonderful education centres out there. Education is not limited to school syllabus but focuses on skill development."

Vansh Agarwal

Student

★★★★★

"My child Dhairya is really enjoying the Modern Age Coders classes. This is his first online class and he eagerly looks forward to it. I can already see his improvement, and the teachers are very cooperative."

Sonam Oswal

Parent of Dhairya

★★★★★

"Modern Age Coders have wonderful teachers who teach in a clear, easy and practical way. The teacher boosts students' confidence and inspires them to learn without hesitation."

Sonu Goyal

Parent

★★★★★

"I highly recommend this computer coding class! The teachers are incredibly knowledgeable and passionate about coding."

Ritu Kedia

Parent

Questions about training AI

What learners and parents ask

What is the difference between prompting and training?

Prompting means wording a request to a system somebody else built. Training means choosing the data, building the model, measuring it against examples it has never seen, and being able to say where it fails.

Why does that difference matter?

Because the UK government's introduction to AI assurance names six techniques for checking an AI system, and every one of them needs something a prompt does not have: the input data, the build record, the stated requirements or a specification.

What are the six?

Risk assessment, algorithmic impact assessment, bias audit, compliance audit, conformity assessment and formal verification.

What is a bias audit?

The guidance defines it as assessing the inputs and outputs of algorithmic systems to determine if there is unfair bias in the input data or in the decision or classification the system makes. Input data comes first in that definition.

What age can a learner start training models?

About twelve for the full version with real datasets, and from around eight for a simpler form using examples the child collects, which is covered on our AI classes for kids page.

Do they need to be good at maths?

Not to begin. The first term is about data, evidence and honest measurement. The mathematics arrives gradually and much later.

Will they train a large language model?

No, and nor does anyone in a classroom. We teach the reasoning at a scale a learner can hold, and the reasoning is what transfers.

What is a held-out set?

Examples put aside before training and used only to judge the finished model. Without one, a model that has memorised the data looks exactly like a model that has learnt something.

Is prompting useless then?

No. It is useful and we use these tools in lessons. It is simply not a substitute for knowing what the system is doing, and it cannot be inspected by anybody.

What does it cost?

Nothing for the opening lesson. A place in a group runs at USD 100 each month afterwards, or USD 150 for a teacher to yourself, billed month by month and never up front.

Elsewhere on this site

Other pages on this

Each page takes one strand of this somewhere specific.

Learn to think, not just use AI tools

Five assumptions about software that AI breaks, from a DSIT-supported review.

AI classes for kids in the UK

The same four stages, with data a child collects themselves.

Problem-solving skills through coding

Why we will not tell you coding makes a child better at everything.

Python classes online in the UK

The language all of this is built in, and what the boards examine.

Coding classes in the UK

The national page, and the index of every UK page we publish.

Student labs

Work our learners built and published.

Start here

Train something in your first lesson

Leave a number and we will reply at a British hour you choose. The first lesson is real teaching, and for most learners it ends with a model they trained themselves.

Would rather read? Every course page lists its syllabus, how we teach explains the method and its limits, and the coding roadmap shows the whole sequence.

WhatsApp us · +91 91233 66161 · contact@modernagecoders.com

WhatsApp is quickest and free from a British mobile. The number is registered in India, stated here rather than discovered later.

Three courses about building, not using

AI and Machine Learning for Teens

Data Science with Python

Generative AI, Properly

What the government says you can do to an AI system

The three verbs underneath

Benchmarks are the crux

Using a system, or building one

This is not snobbery

It is about evidence

And about failure

Data, model, benchmark, failure

The first two

And the two that matter most

Four rungs into building rather than using

Rung three is where it becomes real

Adults often start at rung one

Nine courses about building AI systems

Before the model

Python in full

Data science with Python

Databases and SQL

Training models

AI and machine learning for teens

Visual machine learning

AI literacy for kids

Going further

Generative AI, properly

Data science, complete

Python and AI automation

Hold data back, then try to break it

Data first, always

Small models on purpose

A held-out set every time

Deliberate failure

Written limits

Groups by level

Models our learners trained

NutriLife

Misti

GuardianX

SkyCast

What it costs

Rated 4.9 across 547 Google reviews

What learners and parents ask

What is the difference between prompting and training?

Why does that difference matter?

What are the six?

What is a bias audit?

What age can a learner start training models?

Do they need to be good at maths?

Will they train a large language model?

What is a held-out set?

Is prompting useless then?

What does it cost?

Other pages on this

Learn to think, not just use AI tools

AI classes for kids in the UK

Problem-solving skills through coding

Python classes online in the UK

Coding classes in the UK

Student labs

Train something in your first lesson

Keep exploring Modern Age Coders

Related courses and guides

Learn more

Free resources

From the blog

Start here