AI / ML / DL / GenAI map
- AI is the broadest umbrella: any technique that mimics intelligent behavior
- ML is the subset that learns from data rather than following explicit rules
- Deep Learning is the subset of ML using multi-layer neural networks; Generative AI is the subset of DL focused on creating new content
AI, ML, deep learning, and generative AI are nested subsets, not synonyms. Each level adds a more specific constraint: AI includes any intelligent behavior, ML requires learning from examples, deep learning requires neural network depth, and generative AI requires creating new outputs rather than classifying existing ones.
- Clarifies which tool fits which problem - not every task needs a transformer
- Explains the capability progression: rules -> learning -> representation -> generation
- Connects to the history timeline and the practical "which algorithm to use" decision tree
- Place a given technique into the AI / ML / DL / GenAI hierarchy without ambiguity.
- Distinguish discriminative from generative models with a one-sentence rule.
- Identify which paradigm dominates each contemporary application area.
- Refute the common misconception that "GenAI" is a separate field from machine learning.
TL;DR
AI is the field. ML is the data-driven subset of AI. DL is the neural-network subset of ML. Foundation models (and most of what people mean by “generative AI”) are a deployment pattern within deep learning, not a separate discipline.
Intuition
The four words AI, ML, DL, and GenAI are often used interchangeably in the press and treated as four parallel fields. They are not parallel; they nest.
Reading from outside in:
- Artificial Intelligence is the broad goal of building machines that perform tasks normally requiring human intelligence. It includes everything from chess engines and SAT solvers to robotics and ML. Russell and Norvig [1] give the canonical textbook definition.
- Machine Learning is the subset of AI where the system learns from data rather than from hand-written rules. SVMs, decision trees, logistic regression, and neural networks are all ML. Symbolic planners and SAT solvers are AI but not ML.
- Deep Learning is the subset of ML that uses many-layered neural networks [2][4]. CNNs, RNNs, transformers all live here. Random forests do not.
- Foundation Models are very large neural networks (transformers, mostly) trained once on web-scale data and adapted to many downstream tasks [3]. Most LLMs and modern image generators are foundation models. Generative AI is the use case of producing new artefacts (text, images, audio); the underlying model is almost always a foundation model.
Paradigms compared
| Symbolic AI / GOFAI | Statistical ML | Deep Learning | Foundation Models | |
|---|---|---|---|---|
| Representation | Hand-crafted rules and logical predicates. | Hand-crafted features fed into a parametric classifier. | Learned hierarchical features inside a neural network. | Learned features at scale, transferred to many tasks. |
| Learning signal | None - rules are written by experts. | Labelled examples, typically thousands to millions. | Labelled examples, typically millions. | Self-supervised on unlabelled data, then optionally fine-tuned or RLHF-tuned. |
| Strength | Interpretable, verifiable, sample-efficient. | Strong on tabular data with limited samples; calibrated probabilities. | Dominates perceptual tasks (vision, speech) with sufficient data. | One model, many tasks; in-context learning; emergent capabilities. |
| Weakness | Brittle outside its rule set; expensive to scale. | Plateau without good features; struggles on raw signals (pixels, waveforms). | Data and compute hungry; opaque; no guarantees. | Hallucinations, alignment, cost, prompt injection. |
| Canonical example | Deep Blue chess engine (1997); MYCIN expert system (1976). | sklearn pipelines; XGBoost on Kaggle (2014-). | AlexNet (2012); ResNet (2015); seq2seq (2014). | GPT-4 (2023); CLIP (2021); Stable Diffusion (2022). |
Generative versus discriminative
A clean cross-cut on the same models:
| Family | Models the joint p(x, y) or marginal p(x)? | Models the conditional p(y | x)? | Examples |
|---|---|---|---|
| Discriminative | No | Yes | Logistic regression, SVMs, ResNet image classifiers, BERT for sentiment. |
| Generative | Yes | Sometimes (by Bayes rule) | Naive Bayes, GMMs, VAEs, GANs, diffusion models, autoregressive LLMs. |
A discriminative model only answers “what is y given x?”. A generative model can also draw new samples of x (or of x and y jointly), which is why text-to-image and chatbots feel different from spam classifiers even though all four are deep learning.
“Generative AI is a separate field” is a category error. Generative AI is a use case (producing new artefacts), implemented today with foundation models, which are a kind of deep learning, which is a kind of machine learning, which is a kind of AI. The hierarchy nests; it does not branch.
A second common confusion: not every generative model is an LLM. Diffusion models for images, VAEs for compression, and autoregressive flows for tabular data are all generative; only autoregressive transformers over text are LLMs.
Where each paradigm wins today
| Application | Dominant paradigm | Why |
|---|---|---|
| Tabular data with < 1M rows (credit, fraud, churn) | Statistical ML (XGBoost, logistic regression) | Hand-crafted features + tree boosting still beats deep nets at this scale. |
| Image classification, object detection, segmentation | Deep Learning (CNNs, ViTs) | Pixels are too high-dimensional for hand-crafted features to compete. |
| Speech recognition, machine translation | Deep Learning / Foundation Models | Sequence-to-sequence transformers dominate end to end. |
| Open-ended question answering, code completion, chat | Foundation Models (LLMs) | Trained on text-of-the-internet; in-context learning generalises. |
| Robotic control, planning under uncertainty | Hybrid: classical control + RL + LLM tool use | Pure deep RL is still data-hungry; symbolic planners are still useful for verification. |
| Theorem proving, SAT, constraint solving | Symbolic AI (with ML guidance) | Soundness matters; LLMs are used to suggest tactics, not to verify proofs. |
Pop quiz
- 1. Which of the following is part of AI but NOT machine learning?
- 2. Which is a generative model?
- 3. A team trains GPT-2 from scratch on Wikipedia for sentence completion. Which categories does this system belong to?
- 4. You have 5,000 labelled rows of customer features and need to predict churn. Which paradigm is most appropriate?
Interview Q&A
In one sentence, what is the difference between machine learning and AI?screening
AI is the goal of building machines that perform tasks normally requiring intelligence; ML is the subset of AI in which the machine learns its rules from data rather than having them written by a human. Every ML system is an AI system; not every AI system is an ML system.
A recruiter says they are hiring a 'GenAI engineer, not an ML engineer'. What does that mean in practice?phone
In most companies it is shorthand for “engineer working with foundation models, primarily LLMs, primarily via APIs and fine-tuning rather than training from scratch”. The work is still machine learning: prompt engineering, retrieval-augmented generation, evaluation against benchmarks, fine-tuning open-weights models, alignment via RLHF or DPO. The role differs from a traditional ML engineer in that the model is usually pre-trained and the engineer adapts it, rather than designs and trains it.
It does not mean the underlying field is different. A GenAI engineer who cannot reason about loss functions, evaluation, and overfitting is the same kind of underqualified as an ML engineer who cannot.
Why are random forests often better than deep nets on tabular data?onsite
Three reasons.
- Inductive bias. Trees split on individual features, which matches the structure of tabular data (each column is its own concept). Convolutions, attention, and pooling assume spatial or sequential structure that tabular data lacks.
- Sample efficiency. Trees handle thousands of rows well; deep nets typically need an order of magnitude more data to outperform a tuned XGBoost on the same task.
- Heterogeneous features. Trees handle mixed numerical and categorical features without rescaling. Deep nets need careful preprocessing (standardisation, embedding) for the same input.
The current research consensus (e.g. Grinsztajn et al. 2022, “Why do tree-based models still outperform deep learning on tabular data?”) is that the gap on small-to-medium tabular benchmarks is real and not just a tuning artifact. Deep nets close the gap when you have millions of rows or when you can warm-start from a pretrained tabular foundation model.
Why is the term 'foundation model' a useful category beyond just 'large neural net'?research
Bommasani et al. [3] introduced the term to capture two emergent properties that distinguish very large self-supervised models from regular deep nets:
- Homogenization. A single model serves as the substrate for hundreds of downstream tasks. The economic and infrastructural cost of training the base model gets amortised across applications, but the failure modes also propagate. A bug or bias in the foundation model shows up in every downstream system.
- Emergent capabilities. Some abilities (in-context learning, chain-of-thought reasoning, code generation) appear at scale and are absent at smaller sizes. This is qualitatively different from regular deep nets, where scaling typically improves the same skill the smaller model already had.
The category is useful because it forces you to think about the model as infrastructure, not as an artefact: who controls the base, how does it get updated, how do downstream consumers verify behaviour, who is liable when it is wrong. These are governance questions that did not arise when every team trained their own ResNet from scratch.
References
- Russell, S. and Norvig, P. (2020). Artificial Intelligence: A Modern Approach (4th ed.). Pearson. ISBN 978-0-13-461099-3
- LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learning. Nature 521, 436-444. doi:10.1038/nature14539
- Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., et al. (2021). On the Opportunities and Risks of Foundation Models. arXiv:2108.07258
- Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep Learning. MIT Press. https://www.deeplearningbook.org