The Learning Machine A plain-language guide to artificial intelligence

How to learn AI, and what the jobs are

A six-stage path from spreadsheets to shipped models, plus the roles that hire at each stage.

Start with data, not with models

The common mistake is to begin with a deep learning course, follow it to the end, and find that no employer wants a person who can fine-tune a transformer but cannot clean a spreadsheet. The order below is deliberate, and the early stages are the ones that get people hired.

Timings assume five to eight hours a week. Halve them if you are studying full time, and expect to revisit earlier stages constantly.

Six stages

Each one produces something you can show someone.

  1. Get comfortable with data

    4–6 weeks

    Spreadsheets, then SQL. Learn to load a messy file, clean it, group it and plot it. Most real AI work is this, and people who skip it build models on top of broken data.

    • Spreadsheets
    • SQL
    • Basic statistics
    • Charts that do not mislead

  2. Learn Python properly

    6–10 weeks

    Not the whole language — lists, dictionaries, functions, files, and the three libraries everything sits on: NumPy, pandas and Matplotlib. Write small scripts that do real chores for you.

    • Python
    • NumPy
    • pandas
    • Jupyter notebooks

  3. Understand the maths you will actually use

    8–12 weeks

    Linear algebra for how data moves through layers, calculus for why gradients point downhill, probability for why a model is uncertain. Aim for intuition and notation, not exam technique.

    • Vectors and matrices
    • Derivatives and gradients
    • Probability
    • Distributions

  4. Train classical models first

    8–12 weeks

    Linear and logistic regression, decision trees, random forests, gradient boosting. Learn train/test splits, cross-validation, overfitting and the metrics that matter for imbalanced data. Boosted trees still win on tabular problems.

    • scikit-learn
    • Feature engineering
    • Model evaluation
    • Handling imbalance

  5. Move into deep learning

    3–6 months

    One framework, learned well — PyTorch is the common choice. Build a small image classifier and a small text classifier from scratch, then learn to fine-tune a pretrained model, which is what most jobs require.

    • PyTorch
    • Convolutional networks
    • Transformers
    • Fine-tuning

  6. Ship something to real users

    Ongoing

    A model in a notebook helps nobody. Wrap it in an API, put it behind a simple interface, log what it does, and watch it drift as the world changes. This stage is what separates a portfolio from a certificate.

    • APIs
    • Docker
    • Monitoring
    • Version control

The roles these lead to

Titles vary wildly between companies. The work described here does not.

Role What the job actually involves
Data analyst Turns records into answers a manager can act on. The most common way in, and the fastest to hire for.
Data scientist Frames the question, builds the model, explains the uncertainty. Sits between the business and the maths.
Machine learning engineer Takes models to production and keeps them healthy. Stronger software engineering than the title suggests.
Data engineer Builds the pipelines everything else depends on. Quietly the highest-leverage role on most teams.
MLOps engineer Deployment, versioning, monitoring, rollback. The discipline that stops a good model rotting in production.
AI policy and safety Evaluation, risk assessment, regulation, audit. Growing fast, and it needs people who can read both law and code.
Prompt and product designer Designs how people actually use a model, including what it should refuse to do.
Research scientist Invents the methods. Usually needs a doctorate and a publication record.

Where to study without paying

Nearly all of the material worth having is free. Universities publish their courses, the major libraries have thorough documentation, and the reference textbooks for the field are online with their authors' blessing.

  • Open courseware from universities covers the maths and the classical methods properly.
  • The documentation for pandas, scikit-learn and PyTorch includes tutorials better than most paid courses.
  • Public datasets and competition platforms give you real, messy data to practise on.
  • Reading one paper a fortnight, slowly and with the code, beats any certificate.

What to build first

Pick a question you actually care about in a domain you already understand — your football league, your shop's sales, your county's rainfall. Domain knowledge is what makes a portfolio project look like judgement rather than a tutorial that was followed.

BackBias, risk and hard questions