AI vs Machine Learning vs Deep Learning

The nested-doll mental model

Visualize four concentric circles:

┌─────────────────────────────────────────┐
│ AI (the umbrella)                       │
│  ┌──────────────────────────────────┐   │
│  │ Machine Learning (learn from data)│   │
│  │  ┌────────────────────────────┐  │   │
│  │  │ Deep Learning (neural nets) │  │   │
│  │  │  ┌──────────────────────┐  │  │   │
│  │  │  │ Generative AI        │  │  │   │
│  │  │  │ (create new content) │  │  │   │
│  │  │  └──────────────────────┘  │  │   │
│  │  └────────────────────────────┘  │   │
│  └──────────────────────────────────┘   │
└─────────────────────────────────────────┘

Each inner layer is a subset of the outer one. Each inner layer also needs more data and compute, and gives up some interpretability.

Layer 1: AI (Artificial Intelligence)

The broadest term. AI is any system that performs tasks normally requiring human intelligence. That includes:

• Symbolic / rule-based AI — chess engines pre-2016, expert systems for medical diagnosis, classical planning systems, business rules engines
• Search and optimization — pathfinding, scheduling, constraint satisfaction
• Machine learning — see below
• Hybrid systems — most production AI in regulated industries

The "AI winter" of the 1990s was a crisis for symbolic AI specifically; ML pulled the field out of it.

Layer 2: Machine Learning

ML systems learn patterns from data instead of being programmed with explicit rules. Three sub-types you'll encounter:

• Supervised learning — labeled training data; the model learns input → output mapping. Classification, regression, ranking. Most production ML is supervised.
• Unsupervised learning — no labels; the model finds structure. Clustering, dimensionality reduction, anomaly detection.
• Reinforcement learning — learning by trial-and-error against a reward signal. Robotics, game playing (AlphaGo), recommendation systems, RLHF for LLMs.

Classical ML (XGBoost, random forests, logistic regression, support vector machines) requires you to engineer features — decide what columns to feed the model. This is human effort but also human knowledge baked in.

Layer 3: Deep Learning

Deep learning is ML using multi-layer neural networks that learn their own feature representations from raw data. The "deep" refers to the number of layers — modern frontier models have hundreds to thousands.

What changed in the 2010s: a combination of bigger datasets (ImageNet, web text), specialized hardware (GPUs, later TPUs), and architectural breakthroughs (CNNs for images, transformers for sequences) made deep learning dramatically outperform classical ML on perceptual tasks.

Today's headline architectures:

• CNNs (Convolutional Neural Networks) — image classification, computer vision
• RNNs / LSTMs — older sequence modeling, mostly superseded by transformers
• Transformers — language, modern vision, multimodal — see Transformer Architecture Explained
• Diffusion models — image generation (Stable Diffusion, Flux, DALL-E)
• GANs — older image generation, still used in narrow domains

Layer 4: Generative AI

Generative AI is deep learning models that produce new content (text, images, audio, video, code) rather than classifying or predicting from existing data.

The 2017 transformer paper → 2018 GPT-1 → 2020 GPT-3 → 2022 ChatGPT → 2026's frontier line is the most consequential decade in AI history. Foundation models — single large models trained once and used many times — are how the layer above (deep learning) became something most knowledge workers interact with daily.

See What is Generative AI for the full treatment.

Decision matrix: which to use?

The honest economic ranking

Roughly, by cost to train and deploy:

Most teams should use the API tier (foundation models via OpenAI / Anthropic / Google) for the deep learning capability, and reserve custom training for narrow cases where the economics warrant it.

The 2026 reality

The lines blur in practice:

• LLMs are deep-learning models, but used through APIs the deep-learning machinery is invisible — feels closer to rule-based at the call site.
• Modern recommendation systems combine classical ML, deep learning, and LLM features.
• RLHF (used to train ChatGPT, Claude) combines reinforcement learning + deep learning + supervised fine-tuning.

You'll rarely choose just one layer. Production AI in 2026 is hybrid. The mental model still helps — it tells you what knobs you're turning when you build.