What the work is actually typed into. Titles are on the spines; the notes underneath are the three tellings, in short — for a ten-year-old, for the architect, and for the boardroom.
6 concepts on this shelf · Beginner → Intermediate · about 7 min read
Kitchen table — This is the language people use to write instructions for the machine, and it was designed to read almost like English — 'if the picture is a cat, put it in the cat folder'. It is not the fastest language, but it is one of the friendliest, which is why scientists who never planned to be programmers use it every day. Think of it as the handwriting of AI: nearly everything on the shelves above was scribbled in it first.
Whiteboard — Slow in the loop, unbeatable as the glue. The entire scientific stack assumes it, and the heavy lifting happens in C, C++ and CUDA underneath — your Python is mostly orchestrating compiled kernels, which is why 'Python is slow' is true and rarely matters. The version-and-environment story remains a mild embarrassment (venvs, lockfiles, the annual packaging drama), so treat environment pinning as seriously as you treat code review. The language is the easy part; the ecosystem hygiene is the job.
Boardroom — The default skill on any AI team — if you are hiring for AI, you are hiring Python, and the libraries matter far more than the language itself. Its dominance is a network effect a decade deep, which makes it a safe standardisation bet: tooling, talent and vendor support all point the same way. The hidden cost is not licence fees (there are none) but dependency governance — knowing what open-source code your systems actually run and who keeps it patched.
Python, programming language, data science, scripting
Kitchen table — Imagine a calculator that works on a whole spreadsheet at once. Instead of adding up one row at a time, you say 'double every number in this column' and it happens in a blink, even with a million rows. That is NumPy. pandas sits on top and gives the columns names, so you can say 'show me every sale from March' like you are talking about the data, not marching through it cell by cell.
Whiteboard — Vectorised arrays and dataframes: the discipline is thinking in whole-array operations instead of row loops, knowing where copies happen versus views, and learning groupby properly instead of writing the same join four ways. Chained indexing and silent dtype coercion are where the bugs hide. When the data outgrows one machine there are heirs — Polars, Arrow, DuckDB — but the mental model they inherit is this one, so it is still the one worth learning deeply.
Boardroom — This is the spreadsheet layer of AI work — where data gets cleaned, joined and reshaped before anything intelligent happens to it, and where your analysts actually spend their days. It is free, universal and battle-tested, which means the investment case is not about buying tools but about training: a team fluent here answers questions in minutes that a clumsy one answers in days. Faster successors exist for heavy workloads, but the skills transfer — nothing learned here is wasted.
NumPy, pandas, dataframe, array, vectorisation, data wrangling
Kitchen table — A toolbox full of ready-made learning machines. You do not have to build the machine yourself — you pick one off the shelf, feed it your examples, and say 'fit'. It learns. Then you say 'predict' and it guesses on new things. Every machine in the box works the same way, so once you have learned to use one, you can use them all. It is like a chemistry set where all the experiments follow the same three steps.
Whiteboard — One consistent interface — fit, predict, transform — over most of classical ML, plus pipelines that keep preprocessing honest between training and serving, which is a quieter achievement than it sounds: half of all deployment bugs are a transform applied differently in production than in the notebook. The estimators are solid, the documentation is a model for the industry, and the design has been stable for a decade. For tabular business data, gradient-boosted trees from this box still embarrass deep learning regularly.
Boardroom — For a great many business problems — churn, credit risk, demand forecasting, fraud flags — this free library, not a large language model, is still the right and dramatically cheaper answer. The costs are pennies, the behaviour is stable and explainable, and the talent pool is deep. A useful board question when someone proposes an LLM for a prediction problem: did we try the boring thing first? Very often the boring thing wins on accuracy and costs a thousandth as much to run.
scikit-learn, sklearn, classical machine learning, pipelines, regression, classification
Kitchen table — If scikit-learn is a box of ready-made machines, these are the workshops where you build your own from parts. You snap together the layers like giant Lego, and the workshop does the two hard bits for you: it remembers every calculation you made, and it works out backwards which parts to adjust when the answer was wrong. The big brains — the ones that talk and see — all came out of workshops like these.
Whiteboard — Tensors, autograd and the training loop: define the forward pass, and the framework records the computation graph and differentiates it for you — that is the whole magic. PyTorch won the research mindshare with eager execution that debugs like ordinary Python; TensorFlow persists in production estates and on-device. Everything above them — the fine-tuning libraries, the serving stacks, the agent frameworks — is a convenience layer over these primitives, so when an abstraction leaks, this is the level you drop down to.
Boardroom — This is where deep learning models are actually built and trained, and for most organisations the strategic relevance is simple: these frameworks decide your hardware bill, because they are the reason the GPUs are on order. Unless you are training models yourself, your exposure is indirect — but if you are, know that the framework choice is also a hiring choice and a vendor-ecosystem choice, and that the scarce resource is not the software, which is free, but the people and the silicon.
PyTorch, TensorFlow, deep learning framework, GPU, tensors, autograd, training loop
Kitchen table — Scientists keep lab diaries: what I tried, what happened, what I changed. Machine-learning people need the same thing, because they run hundreds of experiments and the good one is easy to lose. A notebook is where you think out loud with code — a bit of writing, a bit of maths, a chart, all on one page. The tracker is the diary that remembers every run, so when someone asks 'which settings made the good one?', you can actually answer.
Whiteboard — Jupyter for exploration, MLflow or Weights & Biases for runs, parameters, metrics and artefacts. The rule that keeps teams sane: notebooks are for thinking, not for production — they hide state, defeat diff tools and encourage out-of-order execution, so the path from promising notebook to reviewed, tested module has to be a deliberate step, not a copy-paste. Log everything about a run automatically; the experiment you cannot reproduce is an anecdote wearing a lab coat.
Boardroom — This shelf-corner is about one word: reproducibility. If nobody can say which data and which settings produced last quarter's model, you do not have an asset, you have an anecdote — and when a regulator, auditor or acquirer asks how a decision-making system was built, 'we are not sure' is an expensive answer. The tooling is cheap; the habit is the investment. Insist that model results presented to you come with their lineage attached.
Jupyter notebook, MLflow, Weights and Biases, experiment tracking, reproducibility
Kitchen table — Water reaches your tap through pipes you never see, cleaned and pumped along the way. Information travels the same way inside a company — from the shop tills and the website to the place where the learning machines drink. The pipes have to run every night without anyone watching, and when one leaks or clogs, everything downstream goes wrong at breakfast. The machines get blamed, but it is usually the plumbing.
Whiteboard — Batch and streaming, orchestrators and retries, schema drift, late-arriving data, backfills that rewrite history, and idempotency so a rerun does not double-count the world. The unfashionable truth: pipeline failures cause more AI incidents than models do, and the worst ones are silent — the job that succeeds while quietly processing garbage. Data contracts and freshness checks at the boundaries buy more reliability per hour of effort than anything you will do to the model itself.
Boardroom — This is the plumbing, and it obeys plumbing economics: invisible when it works, catastrophic when it does not, and never celebrated at the launch party. Underfund it and every AI initiative downstream inherits the mess — stale dashboards, models trained on wrong numbers, decisions made on data that quietly stopped arriving in March. When AI budgets are set, the unglamorous line item for data engineering is the one protecting all the glamorous ones.
data pipeline, ETL, ELT, orchestration, streaming, data engineering, schema drift