Machine Learning, Data Science, Deep Learning: What the AI Vocabulary Actually Means

On this page

1. A Concrete Starting Point: Two Bakeries
2. Machine Learning, Formally
3. What “learning” actually means, mechanically
4. What actually gets used in practice
5. Data Science, Formally
6. The Boundaries Really Are Fuzzy
7. Neural Networks and Deep Learning
8. The brain analogy: handle with care
9. Transformers, Foundation Models and Generative AI
10. Start with the phone keyboard
11. Where the training data came from
12. The result: foundation models
13. Three Developments That Should Change How You Plan
14. 1. Reasoning models, and paying for thinking time
15. 2. Agents, and the gap between demo and deployment
16. 3. RAG, and the difference between memory and reference
17. Further Terminology You Will Encounter
18. The Tools That Do Not Learn Anything
19. How It All Fits Together
20. Why the Vocabulary Has Business Consequences
21. Glossary
22. The Short Version
23. References
24. Footnotes

Almost everyone in business now hears the words machine learning, data science, neural networks, deep learning and generative AI on a weekly basis. Most of these terms are used loosely. Some are used interchangeably when they should not be. A few are used simply because they sound impressive in a board deck.

That confusion has a cost. Teams buy the wrong tool, hire the wrong person, or ask an engineer for a slide deck and an analyst for a production system.

This article untangles the core vocabulary, shows how the pieces fit together, and updates the classical definitions with what has changed in the last few years. The technical landscape of 2020 and the landscape of today are not the same field.

Two projects, one dataset

A Concrete Starting Point: Two Bakeries#

Imagine a small chain with two bakeries. Three years of till data sits in a spreadsheet: date, store, weather, day of the week, what was baked, what sold, what went in the bin at closing.

There are two fundamentally different things you can do with that table, and they map almost perfectly onto the two most commonly confused terms in the industry.

Option one: build something that runs. You want the ordering system to decide, every night at ten, how many of each pastry to prepare for tomorrow. The input, meaning tomorrow’s date, store, forecast weather and the recent sales history, is conventionally called A. The output, meaning the quantity to bake, is called B.

A system that learns this A-to-B mapping from historical examples is a machine learning system. The deliverable is software. It is a service that can respond at 3 a.m. on a Sunday without a human in the loop, and it has to keep responding, correctly, next month and next year.

Option two: build understanding. You want someone to examine the same data and surface insights. They may find that the second store throws away nearly twice as much as the first on weekday mornings. They may find that opening two hours earlier on Saturdays would pay for itself, and that the same change on Tuesdays would not.

Those findings inform decisions. Should the Saturday shift start at six? Is the second store’s waste a baking problem or a forecasting problem? This is a data science project. The deliverable is not running software. The deliverable is a better decision.

That distinction is the first important line to remember: machine learning often produces systems that run; data science often produces insights that humans act on.

If the word data itself is doing a lot of unexamined work in your organisation, that is a separate and prior problem, and Understanding Data is the piece on it.

The older definition

Machine Learning, Formally#

The definition most often quoted in slide decks comes from Arthur Samuel, and it is worth knowing that the famous wording, the field of study that gives computers the ability to learn without being explicitly programmed, does not actually appear in the 1959 paper it is credited to. It is a later paraphrase that hardened into a quotation. What Samuel did write, in the abstract of that paper, is better anyway: “Enough work has been done to verify the fact that a computer can be programmed so that it will learn to play a better game of checkers than can be played by the person who wrote the program.”¹

Samuel was an IBM researcher whose checkers-playing program, developed through the 1950s and 60s, eventually played the game better than he did. That detail matters because it captures the central idea of the field: the system became competent in a way its creator could not fully write down as a list of rules.

The practical signature of machine learning is that you specify the goal and provide examples, rather than specifying the full procedure.² Nobody writes a complete set of if-then rules for “this photo contains a tumor” or “this transaction is fraudulent.” Instead, you supply labelled examples and the algorithm learns a pattern.

What “learning” actually means, mechanically#

This is worth pinning down because it demystifies almost everything else.

A model starts out useless. Its internal numbers are basically random, and its first guess at tomorrow’s croissant count will be wildly wrong. The training process then does something simple, over and over:

1. Show the model one example, such as a single day from the sales history.
2. Let it guess a quantity.
3. Measure how wrong the guess was compared to what actually sold.
4. Adjust its internal numbers slightly so the next guess would be less wrong.
5. Repeat this process millions or billions of times.

That is the basic idea. Step 4 has a formal name, gradient descent. The method for working out which internal numbers should move, and in which direction, is called backpropagation. But the plain-English version is enough: guess, measure the error, adjust, try again.

There is no ghost in the machine. The system is just adjusting a huge number of internal settings until the answers improve. The surprising discovery of modern AI is that when this process happens at enormous scale, the result can look a lot like competence.

What actually gets used in practice#

The most commercially dominant variety of machine learning remains supervised learning, which means learning A-to-B mappings from labelled examples.³ Despite years of hype around more exotic approaches, much of the value captured by deployed AI in insurance, logistics, credit, advertising, manufacturing inspection and medical imaging still comes from supervised learning on tabular, image or text inputs.

One useful fact often gets lost in the hype: for structured business data, the kind that lives in spreadsheets and databases, the best-performing tool is often not a neural network. It is a family of methods called gradient-boosted decision trees. You may see names like XGBoost, LightGBM and CatBoost.

The idea is intuitive. Build a simple flowchart of yes/no questions that predicts the answer roughly. Then build a second flowchart that specialises in correcting the first one’s mistakes. Then build a third to correct what remains. Repeat this for hundreds of rounds.

On the kind of table most companies actually have, meaning tens of thousands of rows, mixed types and a lot of uninformative columns, this approach still beats deep learning, and the benchmark that established it has not been overturned for that regime.⁴ It has been contested at the small end: a tabular foundation model published in 2025 outperforms tuned tree ensembles on datasets up to around ten thousand rows.⁵ Neither result changes the practical point, which is that a cheaper, simpler, more explainable model is often better than a fashionable one.

The other half of the pair

Data Science, Formally#

Data science is the discipline of extracting knowledge and insight from data.⁶ The typical artifact is a presentation, a dashboard, an experiment readout, or a memo that changes what an organisation does.

Digital advertising gives a clean example. Every major ad platform runs a click-through-rate prediction model. It takes in signals about a user and a candidate advertisement, then outputs the probability that the user will click.⁷ That is machine learning. It runs continuously, and it is one of the most lucrative types of software ever built.

The same company may also run an analysis showing that firms in the travel sector are under-investing in paid advertising relative to their addressable demand. The analysis might conclude that assigning more salespeople to that vertical would pay for itself. That is data science. The output is a conclusion. The action is taken by humans reallocating a sales team.

Both types of work can exist inside one company. Both can be worth enormous sums. But they require different skills, different tools and different success metrics.

Where the lines blur

The Boundaries Really Are Fuzzy#

Nobody should pretend these definitions are enforced consistently. Usage varies by company, country and decade.

Job titles have splintered accordingly: data analyst, analytics engineer, data scientist, machine learning engineer, research scientist, MLOps engineer, and now AI engineer. The last title usually refers to someone who builds products on top of existing foundation models rather than training new models from scratch. That role barely existed before 2023 and is now one of the fastest-growing categories in the field.⁸

If you are hiring, describe the deliverable rather than relying on the title.

“We need a service that scores every incoming claim within 200 milliseconds” and “we need to know which of our five customer segments is actually profitable” are different job postings, even if both candidates have AI-related words on their résumés.

Inside the model

Neural Networks and Deep Learning#

Return to the bakery. One way to map the input attributes to a quantity is to feed them into a neural network, often called an artificial neural network to distinguish it from the biological kind.

Here is the whole idea without mathematics. Imagine a committee arranged in rows. The facts about tomorrow are handed to the first row. Each member of that row pays attention to some facts more than others. One might care mostly about the weather and slightly about the day of the week. Each member then passes a single number forward as its verdict.

The second row receives all the verdicts from the first row, weighs them according to its own preferences, and passes its own verdicts to the third row. After several rows, one final member outputs a number.

Nobody told any committee member what to care about. Every one of those preferences, known as weights, was discovered through the guess-check-adjust process described earlier. The circles in neural network diagrams are these committee members. They are called artificial neurons, or simply neurons.

Why use rows at all? Because layering lets the network build ideas in stages. Early rows can only combine raw facts. Later rows combine the combinations, which means they can represent things no single input captures by itself, such as “a cold wet Saturday in the second week of term.”

In image recognition this becomes visible: early layers detect edges, middle layers detect shapes such as eyes and wheels, later layers detect faces, cars or buildings. Depth is what makes that kind of abstraction possible, and depth is where the term “deep learning” comes from. Machine Learning, Explained draws that ladder in full.

A neural network is a very large mathematical function, fitted to data by repeated error correction. It can be powerful, but it is not magic.

Deep learning and neural networks are now used almost interchangeably. Historically, “neural network” was the older term, used from the 1940s through the 1990s and through several boom-and-bust cycles. “Deep learning,” referring to networks with many layers, became popular from roughly 2006 onward. It was a stronger brand, and the name stuck.

The brain analogy: handle with care#

Neural networks were inspired by neuroscience, but the resemblance is limited. Biological neurons communicate through spike timing, neurotransmitter chemistry, dendritic computation and neuromodulation. The human brain learns continuously from sparse experience, consumes around 20 watts, and does not run backpropagation in any demonstrated literal sense.

Calling a language model a “digital brain” is usually marketing, not science. It leads people to overestimate these systems in some ways and underestimate them in others.

The historical debt to neuroscience is still real. The 2024 Nobel Prize in Physics went to John Hopfield and Geoffrey Hinton, in the committee’s own words, for foundational discoveries and inventions that enable machine learning with artificial neural networks.⁹ The 2024 Nobel Prize in Chemistry was divided, with one half to David Baker for computational protein design and the other half jointly to Demis Hassabis and John Jumper for protein structure prediction, the work behind AlphaFold.¹⁰ Those awards marked an important shift. Neural networks had moved from being a computer-science curiosity to becoming general-purpose scientific infrastructure.

What changed after 2017

Transformers, Foundation Models and Generative AI#

The classical picture above, supervised learning, A to B, one model per task, described the field well for most of the 2010s. Then the architecture changed the economics.

Start with the phone keyboard#

The most useful mental model for a language model is predictive text on your phone. You type “I’ll be there in five” and it offers “minutes.” It is not thinking. It has simply seen that “minutes” often follows that phrase.

A large language model does the same basic thing: it predicts the next fragment of text. The difference is scale. It has been trained on a huge amount of text rather than your recent messages. It also considers far more context than a phone keyboard, including instructions given several paragraphs earlier.

That second difference is crucial. It is what the transformer architecture, introduced in 2017, made possible.¹¹ Its central mechanism is called attention.

Attention means that for every word the model processes, it works out which other words in the passage are relevant to it and gives them more weight. Older architectures processed text more rigidly and struggled with long-range connections. Attention made those connections easier to learn, and Machine Learning, Explained has the worked example.

The practical consequence was economic as much as technical. Attention allowed models to train efficiently on GPUs, the massively parallel hardware the industry already had. Training that would once have been impractical became possible. That efficiency unlocked scale. Scale then produced a qualitative shift: a system trained to predict the next word turned out to be able to summarise, translate, write code, answer questions and follow instructions.

Nobody explicitly programmed each of those abilities. They emerged from training at scale.

Where the training data came from#

The second ingredient was self-supervised learning, and it solved a major bottleneck.

Supervised learning needs labels, and labels usually need humans. To build a medical imaging model, radiologists must annotate thousands of scans. That is slow and expensive.

Self-supervised learning avoids that bottleneck with a simple trick: hide part of the data and make the model predict the hidden part. Take a sentence from a book, cover the last word, ask the model to guess it, then reveal the answer. No human labelled anything. The text itself provides both the question and the answer.

This means existing text, code, images, audio and video can become training material. The labelling ceiling disappeared, and training sets grew by orders of magnitude. What a model retains from all that reading, and what it does not, is the subject of What Pre-Trained Knowledge Can and Cannot Do.

The result: foundation models#

The output of that process is a foundation model.¹² When the primary modality is text, people often call it a large language model.

A foundation model is a very large pretrained network that can be adapted to many downstream tasks through prompting, fine-tuning or tool use. Instead of building one narrow model for each task, a company can start from a general model and adapt it.

This changed the economics. Under the older model, every business problem often required its own dataset, its own training run and its own deployment. Under the newer model, someone else may spend hundreds of millions of dollars building a general capability, and a business adapts it with instructions, retrieval, fine-tuning or integration into tools.

Generative AI is the popular label for systems whose output is content: prose, code, images, audio or video, rather than a number or a category.

It is useful to remember where it sits. Today’s generative AI sits inside deep learning. Deep learning sits inside neural network methods, which sit inside machine learning, which sits inside AI. Generative AI gets most of the attention right now, but it is still only one part of the field. The Rise of Artificial Intelligence covers what that attention has and has not been worth so far.

The three that cost money

Three Developments That Should Change How You Plan#

Of everything that has happened since 2023, three shifts have direct operational consequences: reasoning models, agents and retrieval. These are worth understanding properly.

1. Reasoning models, and paying for thinking time#

Until late 2024, many language models produced answers more or less immediately. That is fine for simple tasks but weak for problems requiring multi-step reasoning.

Reasoning models changed the pattern. They are trained to work through problems step by step before answering. In simple terms, they are given more room to think before they respond.

The training method matters. Rather than only being shown examples of good reasoning, these models are often trained on problems where the answer can be automatically verified, such as mathematics, competitive programming and formal logic. The model attempts a problem many times. The approaches that reach the correct answer are reinforced. Over many rounds, it learns reasoning strategies that nobody manually wrote into it.¹³

Beginning with OpenAI’s o-series in late 2024 and rapidly matched by DeepSeek’s R1, Anthropic’s extended-thinking models and Google’s Gemini reasoning variants, this established a second scaling axis: test-time compute.¹⁴

Historically, better performance usually required a bigger model and a more expensive training run. Now there is another lever: let the model spend more computation on a particular answer. On mathematics, competitive programming and formal reasoning benchmarks, the improvements have been steep. In July 2025 a model scored at the gold-medal threshold on the International Mathematical Olympiad problems, solving five of six under contest conditions, with the solutions graded by the competition’s own coordinators.¹⁵

The practical point is simple: quality has become a dial. Cheap and instant may be fine for routine classification. Expensive and deliberate may be worth it for contract analysis, financial reasoning or code review. AI budgeting is no longer just a single price per query. That said, the price of a given level of capability keeps falling hard. Holding capability fixed at GPT-3.5 level, the cost of serving an answer fell from twenty dollars per million tokens in late 2022 to seven cents by late 2024.¹⁶

2. Agents, and the gap between demo and deployment#

An agent is a model given the ability to act rather than merely respond. It can call external tools and APIs, search the web, write and run code, query a database, file a ticket, or operate software. What Is Agentic AI is the primer on the word itself, and Why Agentic Workflows Make AI More Useful is the one on what the extra machinery buys you.

Instead of saying, “Here is how you would reconcile those invoices,” an agent may try to reconcile them. It plans steps, executes them, notices failures and adapts.

This is where much enterprise attention has shifted. It is also where the gap between an impressive demo and a reliable production system is widest.

The reason is not mysterious. A single step that succeeds 95% of the time sounds excellent. But twenty such steps chained together succeed only about 36% of the time. Agentic reliability depends less on flashy model intelligence than on tight scoping, verification at each step, and well-designed moments where a human intervenes. Benchmarks that run the same task many times over show exactly this collapse, and measured agent capability is now often reported as the length of task a system can finish rather than as a single accuracy number.^{17 18}

Interactive // illustrative arithmetic, not a measurement

Pick how reliable one step is, and how many steps the agent has to chain together. The number underneath is what is left when they multiply.

Each step succeeds

80%90%95%99%99.9%

Steps in the chain

135102050

Reset

35.8%of runs finish every step without a misstep

• In plain numbersabout 36 runs in 100 get through all 20
• To hold the chain above 90%every step would have to clear 99.47%
• ReadingFinishing clean is now the exception, not the rule.

Independent steps is an assumption, and a strong one. Real steps correlate, a retry can rescue one, and a check step can catch a failure before it spreads. The multiplication is exact. The model of the world is not.

Standardised tool-calling interfaces have emerged to handle the plumbing. The most discussed is the Model Context Protocol (MCP), released by Anthropic in November 2024 and since adopted broadly across the industry.¹⁹

The point of MCP is straightforward. Rather than writing custom integration code for every combination of model and tool, you expose systems through a common interface so compatible models can use them. Think of it as a standard socket instead of a drawer full of proprietary adapters.

The practical point: agents fail on process design more often than on model quality. The organisations doing well with agents are the ones that simplify and map a workflow before automating it.

3. RAG, and the difference between memory and reference#

A foundation model knows only what was in its training data, and that training data has a cutoff date. It also does not automatically know your internal policies, contracts, customer records or operating manuals.

Retrieval-augmented generation (RAG) fixes this by changing the model’s job from recall to reading comprehension.

Here is the mechanism. When a question arrives, the system first searches your document repository for relevant passages. It then gives those passages to the model along with the question and an instruction to answer using the supplied material.

The model does not need to have memorised your refund policy. It needs to read the refund policy that was retrieved for it.

The search step relies on embeddings, which are a way of converting text into a list of numbers so that passages with similar meanings end up numerically close together. This allows search by meaning rather than by exact keyword. A query about “reimbursement for cancelled travel” can retrieve a policy document that says “refunds for aborted journeys.” Vector databases exist to store and search those numeric representations at scale, which is why they became a product category so quickly. If you want the mechanism in full that is Retrieval Augmented Generation, and What RAG Is Actually Used For covers where it earns its keep.

The practical point: RAG is usually the cheapest and most controllable route to making a general model useful on proprietary data. It also has a governance benefit. Answers can cite the source passages they used, which makes them easier to audit than answers pulled from a model’s internal memory.

The rest of the wall

Further Terminology You Will Encounter#

These are worth recognising rather than mastering. Skim them if you need to.

RLHF (reinforcement learning from human feedback) is the technique that helped turn raw text predictors into usable assistants. A model trained only to predict internet text may continue a rude question rudely. RLHF improves this by having humans compare pairs of candidate responses, training a second model to imitate those preferences, and then tuning the main model to score well against it. Related methods include DPO, which aims for a similar effect more directly, and constitutional or AI-feedback methods, where the model critiques its own outputs against a written set of principles rather than relying entirely on human raters.²⁰

Distillation means training a smaller model to imitate a larger one. The goal is to capture much of the large model’s capability at a fraction of the running cost.

Quantisation means storing a model’s internal numbers at lower precision. Roughly speaking, it rounds the numbers to shrink memory use and increase speed, usually with limited quality loss.

Mixture-of-experts is an architecture where the model contains many specialised subnetworks but activates only a few for each query. This allows a very large model to be cheaper to run because most of it stays idle on any given request.

Together, distillation, quantisation and mixture-of-experts explain much of the collapse in inference cost.²¹ Open-weight models released by Meta, Mistral, DeepSeek, Alibaba’s Qwen team and others have closed much of the gap to frontier proprietary systems, often within months.

Multimodality means models that can work across text, images, audio and video, instead of treating each format as a separate world.

Fine-tuning means continuing to train an existing model on your own examples so it adopts a particular style, format or narrow specialisation. This is different from RAG. Fine-tuning changes behaviour; RAG supplies knowledge. Teams often reach for fine-tuning when what they actually need is retrieval, and the measured comparison favours retrieval when the goal is new facts rather than new manners.²²

Context window means how much text a model can consider at once, including your instructions, retrieved documents and conversation history. It is a practical constraint on system design, although context windows have expanded dramatically.

Hallucination means a model producing fluent, confident, false output. This is not just a minor bug. It follows from how these systems work: they generate plausible continuations, and plausibility is not the same as truth. Hallucinations can be reduced with retrieval, grounding, verification and citation, but they cannot be treated as fully solved. Why AI Models Fall Behind the Real World is the piece on where the confident wrong answers come from.

Unsupervised learning means finding structure in unlabelled data, for example clustering customers into segments nobody defined in advance.

Reinforcement learning means learning by trial, error and reward rather than from labelled examples. It is the paradigm behind game-playing systems like AlphaGo and, in modern language models, an important part of instruction-following and reasoning.

Graphical models, planning and knowledge graphs are established AI toolkits that predate the deep learning era and remain in active use.

The part everyone forgets

The Tools That Do Not Learn Anything#

AI’s non-machine-learning tools have not disappeared. It is a mistake to assume every problem calls for a model.

Constraint solvers and mathematical optimisation quietly run global logistics, airline crew scheduling and electricity grids. These are problems where you need a provably optimal or provably feasible answer, not a good guess, and the underlying solvers got orders of magnitude faster over three decades without learning anything from data.²³

Knowledge graphs underpin search and pharmaceutical research. Classical planning and search remain central to robotics.

Increasingly, the strongest production systems are hybrids. A language model handles ambiguity, unstructured input and the human interface. A solver, database or rules engine handles anything that must be exactly correct. The olympiad-geometry system that reached medal standard in 2024 worked precisely this way: a neural network proposed constructions, and a symbolic engine checked them.

A model can interpret a customer’s messy email. A deterministic system should compute the refund amount.

The map

How It All Fits Together#

Picture a set of nested regions.

Artificial intelligence is the outermost set. It includes the entire collection of techniques for making computers behave intelligently: search, logic, optimisation, planning, knowledge representation, machine learning and more.

Machine learning is the largest and most economically important subset of AI, but it is not the whole field.

Neural networks are the most important family inside machine learning today, and deep learning means using them at depth. But plenty of valuable machine learning is not deep learning. Tree ensembles, linear and logistic regression, clustering, matrix factorisation and classical time-series methods all remain workhorses. Matrix factorisation, for example, was central to classic recommendation engines because it could infer taste from patterns in who liked what.

Generative AI, as practised today, sits inside deep learning. It is a subset of a subset, despite absorbing much of the public attention and investment.

Data science does not fit neatly inside any of these circles, which is why terminology arguments never fully resolve. Some people call data science a subset of AI. Others say AI is a subset of data science.

The more useful description is that data science is a cross-cutting discipline. It borrows from AI, machine learning and deep learning. It also adds statistics, experimental design, causal inference, data engineering and communication.

Causal inference is especially important because it helps distinguish “these two things move together” from “this one causes the other.” That is the difference between a correlation and a decision you can safely act on.

Why any of this matters

Why the Vocabulary Has Business Consequences#

Getting these words straight is not academic nitpicking. It changes what you build, who you hire, how you budget and how you judge success.

Enterprise surveys, including McKinsey’s annual State of AI work, consistently find that many organisations now report using AI somewhere, while a much smaller share report material impact on earnings. The Rise of Artificial Intelligence works through those numbers. The most-quoted version of the finding is a 2025 report from an MIT Media Lab research initiative, which examined enterprise generative-AI pilots and concluded that around 95% of them had produced no measurable profit-and-loss impact.²⁴

It is worth knowing what that sentence is. The report is an industry study built on roughly fifty structured interviews, a conference survey of about a hundred and fifty leaders, and an analysis of around three hundred publicly announced deployments. It is not a peer-reviewed experiment, and “no measurable impact yet” is a different finding from “failed.” It is still the number in the room, which is a good reason to know what it actually says. Its own explanation for the gap points at workflow integration, data readiness and unclear ownership rather than at model quality.

The pattern is familiar: many AI projects do not fail because the model is too weak. They fail because the organisation has not decided where the system fits, who owns it, what data it can trust, and what happens when it is wrong.

Three practical implications follow.

Know which deliverable you are commissioning. A project whose output is a running service needs engineering and operational discipline. It needs a latency budget, meaning a hard limit on acceptable response time. It needs version control so you can tell which model produced which decision. It needs an evaluation harness, meaning an automated test suite that catches quality regressions before customers do. It also needs someone accountable when it breaks. The trained model is usually the small box in a much larger diagram, which is the finding that gave the field the discipline now called MLOps, or LLMOps in its generative-AI variant.²⁵

A project whose output is an insight needs something different: statistical rigour, honesty about uncertainty, and a decision-maker who has agreed in advance to act on the finding if the evidence is strong.

Confusing the two produces prototypes that never ship and analyses nobody reads.

Data quality dominates model choice. The industry’s shift toward data-centric AI reflects hard-won experience: in high-stakes deployments, the compounding problems tend to start in the data rather than in the model.²⁶ For most business problems, a modest model on clean proprietary data beats a frontier model on messy data. Most companies do not need a custom foundation model. They need cleaner data, better workflow design, and someone accountable for deployment. Understanding Data is the primer on what that actually involves.

Not everything needs a neural network. If your problem is tabular prediction on a few hundred thousand rows, gradient boosting will likely be cheaper, faster, more accurate and easier to explain than deep learning.

If your problem must be auditable under regulation, interpretability and documentation may matter more than the last two points of accuracy. The EU AI Act’s prohibitions and AI-literacy duties have applied since February 2025 and its general-purpose model obligations since August 2025, while the high-risk regime, the part that actually drives documentation and traceability, was pushed back in July 2026 to December 2027 for standalone systems and August 2028 for AI embedded in regulated products.²⁷ The deadline moved. The design question did not.

“We used a language model” is not always a defensible answer to a regulator, a board, or a customer asking why they were declined.

Reference

Glossary#

Data science

Using data to answer a question so that a decision can be made. The deliverable is the decision, not a piece of software.

Deliverable

What a project actually hands over. For machine learning it is usually software that keeps running; for data science it is usually one answer.

Tabular data

Data in rows and columns, the shape most business data already has: transactions, orders, claims, sensor readings.

Gradient boosting

A method that builds many small decision trees in sequence, each one correcting the last. Still the strong default on tabular data.

Self-supervised learning

Learning from raw data by hiding part of it and predicting the hidden part, so nobody has to label examples by hand.

Foundation model

A large model trained broadly once, then adapted to many tasks instead of trained from scratch for each one.

Reasoning model

A model trained to work through a problem in steps before answering, spending more computation on harder questions.

Test-time compute

Computation spent while answering rather than while training. Buying accuracy with time and tokens at the moment you ask.

RLHF

Reinforcement learning from human feedback. People compare answers, and the model is tuned toward the ones they preferred.

DPO

Direct preference optimization. Reaches a similar result to RLHF by training on preferred-versus-rejected pairs, without a separate reward model.

Constitutional AI

Using a written set of principles, applied by the model to its own drafts, in place of some of the human rating.

Distillation

Training a smaller model to imitate a larger one, so most of the behaviour survives at a fraction of the running cost.

Quantisation

Storing a model's internal numbers at lower precision so it needs less memory and runs on cheaper hardware.

Mixture of experts

A model split into many sub-networks where only a few run for any given input, so it can be large without costing large.

Multimodality

Handling more than one kind of input or output in a single model: text with images, audio or video.

Solver

Software that finds an exact or provably good answer to a defined problem, such as a schedule or a route. It learns nothing.

Classical planning

Searching for a sequence of actions that reaches a goal from a formal description of the world. No training data involved.

MLOps

The engineering around a deployed model: pipelines, versioning, monitoring, retraining. Usually the larger half of the work.

LLMOps

The same discipline for systems built on language models, where prompts, retrieval, tools and evaluations also have to be versioned and watched.

Data-centric AI

Improving results by improving the data rather than the model, on the view that most remaining errors come from the examples.

The takeaway

The Short Version#

If someone says machine learning, they mean learning input-to-output mappings from examples, and the deliverable is running software.

If someone says data science, they mean extracting insight from data, and the deliverable is a decision.

If someone says a model learns, they mean it guesses, measures the error, adjusts its internal numbers, and repeats that many times.

If someone says neural network, they mean a large mathematical function built from layers of simple units and fitted to data. Deep learning is the modern name for using them at depth and scale. Neither is a digital brain; the relationship to biological brains is loose historical inspiration, not a close scientific match.

If someone says transformer, they mean the architecture that made training on internet-scale text efficient. Self-supervised learning is what removed the need for human labels at that scale. Together they produced foundation models.

If someone says generative AI, they mean the content-producing corner of that, which is a subset of a subset.

If someone says reasoning, agents or retrieval, pay attention, because those three have the clearest operational consequences: paying for more thinking time, letting a model act, and grounding answers in your own documents.

And if someone says AI, they may mean none of the above. The umbrella still contains valuable tools that involve no learning at all.

Everything else, including unsupervised learning, reinforcement learning, graphical models, planning, knowledge graphs and embeddings, is a specific instrument in the toolbox. You do not need to master each one. You do need to know what kind of problem you have, what artifact you are asking for, and who in your organisation is accountable for the result.

Sources

References#

Dates, figures and regulatory deadlines below are snapshots verified 2026-07-30 against the cited primary sources.

Footnotes#

1. Arthur L. Samuel, “Some Studies in Machine Learning Using the Game of Checkers,” IBM Journal of Research and Development 3(3), Jul 1959. ieeexplore.ieee.org. ↩

2. Tom M. Mitchell, Machine Learning, McGraw-Hill, 1997. cs.cmu.edu. ↩

3. Yann LeCun, Yoshua Bengio, Geoffrey Hinton, “Deep learning,” Nature 521, 2015. nature.com. ↩

4. Léo Grinsztajn, Edouard Oyallon, Gaël Varoquaux, “Why do tree-based models still outperform deep learning on typical tabular data?”, NeurIPS Datasets and Benchmarks, 2022. arxiv.org. ↩

5. Noah Hollmann et al., “Accurate predictions on small data with a tabular foundation model,” Nature 637, Jan 2025. nature.com. ↩

6. William S. Cleveland, “Data Science: An Action Plan for Expanding the Technical Areas of the Field of Statistics,” International Statistical Review 69(1), 2001. doi.org. On the job title, Thomas H. Davenport and D.J. Patil, “Data Scientist: The Sexiest Job of the 21st Century,” Harvard Business Review, Oct 2012. hbr.org. ↩

7. H. Brendan McMahan et al., “Ad Click Prediction: a View from the Trenches,” KDD 2013. research.google. ↩

8. Shawn Wang, “The Rise of the AI Engineer,” Latent Space, Jun 2023. latent.space. ↩

9. The Royal Swedish Academy of Sciences, “The Nobel Prize in Physics 2024,” 8 Oct 2024. nobelprize.org. ↩

10. The Royal Swedish Academy of Sciences, “The Nobel Prize in Chemistry 2024,” 9 Oct 2024. nobelprize.org. ↩

11. Ashish Vaswani et al., “Attention Is All You Need,” NeurIPS 2017. arxiv.org. ↩

12. Rishi Bommasani et al., “On the Opportunities and Risks of Foundation Models,” Stanford CRFM, Aug 2021. arxiv.org. ↩

13. DeepSeek-AI, “DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning,” Nature, Sep 2025. nature.com. ↩

14. Charlie Snell, Jaehoon Lee, Kelvin Xu, Aviral Kumar, “Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters,” 2024. arxiv.org. ↩

15. Google DeepMind, “Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad,” 21 Jul 2025. deepmind.google. ↩

16. Stanford HAI, Artificial Intelligence Index Report 2025, ch. 1. hai.stanford.edu. ↩

17. Shunyu Yao et al., “τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains,” 2024. arxiv.org. ↩

18. Thomas Kwa et al., “Measuring AI Ability to Complete Long Software Tasks,” METR, Mar 2025. arxiv.org. ↩

19. Anthropic, “Introducing the Model Context Protocol,” 25 Nov 2024. anthropic.com. ↩

20. Paul Christiano et al., “Deep reinforcement learning from human preferences,” 2017. arxiv.org. On DPO, Rafael Rafailov et al., “Direct Preference Optimization,” NeurIPS 2023. arxiv.org. On AI feedback, Yuntao Bai et al., “Constitutional AI: Harmlessness from AI Feedback,” 2022. arxiv.org. ↩

21. Geoffrey Hinton, Oriol Vinyals, Jeff Dean, “Distilling the Knowledge in a Neural Network,” 2015. arxiv.org. On quantisation, Tim Dettmers et al., “QLoRA: Efficient Finetuning of Quantized LLMs,” NeurIPS 2023. arxiv.org. On mixture-of-experts, Noam Shazeer et al., “Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer,” ICLR 2017. arxiv.org. On multimodality, Alec Radford et al., “Learning Transferable Visual Models From Natural Language Supervision,” ICML 2021. arxiv.org. ↩

22. Oded Ovadia, Menachem Brief, Moshik Mishaeli, Oren Elisha, “Fine-Tuning or Retrieval? Comparing Knowledge Injection in LLMs,” EMNLP 2024. arxiv.org. ↩

23. Robert E. Bixby, “A Brief History of Linear and Mixed-Integer Programming Computation,” Documenta Mathematica, ISMP extra volume, 2012. ems.press. On knowledge graphs, Google, “Introducing the Knowledge Graph: things, not strings,” 16 May 2012. blog.google. On the hybrid, Trieu H. Trinh et al., “Solving olympiad geometry without human demonstrations,” Nature 625, Jan 2024. nature.com. ↩

24. MIT Media Lab Project NANDA, The GenAI Divide: State of AI in Business 2025, Jul 2025. mit.edu. ↩

25. D. Sculley et al., “Hidden Technical Debt in Machine Learning Systems,” NeurIPS 2015. research.google. ↩

26. Nithya Sambasivan et al., “‘Everyone wants to do the model work, not the data work’: Data Cascades in High-Stakes AI,” CHI 2021. research.google. ↩

27. Regulation (EU) 2026/1744 of 8 Jul 2026 (Digital Omnibus on AI), amending Regulation (EU) 2024/1689. eur-lex.europa.eu. ↩