§ 1.2 AI Taxonomy & Industry Landscape

1. Overview

Before studying LLMs in depth, you need a mental map: where they sit in the broader AI family, which research communities created them, who is building them commercially today, and which curriculum parts you need based on the career you are targeting. This page answers all four questions. Read it once now; the job table in §5 will make more sense after you have worked through Parts IV–X.

The key insight from the diagram: LLMs are deep learning applied to text at scale. Reinforcement Learning (RL) overlaps both DL and classical ML; it reappears in post-training (RLHF, GRPO) where it is used to align LLMs to human preferences. Symbolic AI remains relevant as the verifier in neuro-symbolic systems and in formal-methods tooling.

2. AI vs ML vs DL vs LLM

These terms are not synonymous: each inner set is a strict subset of the outer one. The distinctions matter in practice because they determine what kind of data, compute, and expertise each approach requires.

Where does Reinforcement Learning (RL) fit?

RL is a learning paradigm, not a model family. An RL agent learns by trial-and-error with delayed reward signals rather than from labelled examples. It overlaps classical ML (tabular Q-learning) and DL (deep Q-network, neural policy gradient). In the LLM context, RL appears in post-training: RLHF and GRPO use policy-gradient algorithms to push the base model toward preferred outputs, making RL an essential ingredient of ChatGPT-class alignment. Parts XVII and IX cover this in depth.

Why is "AGI" slippery?

Artificial General Intelligence has no agreed mathematical definition. Common informal definitions differ on the test (Turing test? all human cognitive tasks? ARC-AGI benchmark?), the subject (human-level? expert-level?), and the timescale (stateless test? lifelong learning?). Frontier labs use the term differently: OpenAI defines AGI as "highly autonomous systems that outperform humans at most economically valuable work"; Anthropic focuses on "transformative AI"; Google DeepMind uses AGI as an internal milestone framework. For this curriculum, treat AGI as a direction of travel, not a crisp target.

3. Sub-fields & Transformer Adoption

Since 2020, the transformer architecture has colonized every major AI sub-field. The diagram below shows the adoption sequence; each node gives the canonical first transformer-dominant paper in that domain.

4. Industry Landscape (2026)

The LLM industry has consolidated around a handful of frontier labs — organizations that can afford the compute to train at the frontier (>$50M per run) and have the talent density to push architecture, post-training, and systems simultaneously.

4.1 United States & Europe

4.2 China

Chinese labs have closed the frontier gap rapidly. DeepSeek-V3 (2024) matched GPT-4-class quality at a reported training cost of <$6M; DeepSeek-R1 matched o1-class reasoning. The ecosystem is structurally different from the US: domestic compute (Ascend, Cambricon) + heavy quantization pressure from US export controls drives aggressive efficiency research (MLA, FP8 training, mixture-of-experts routing innovations).

4.3 Open-Source Ecosystem

The open-source layer wraps around every frontier lab:

• Hugging Face — model hub, transformers library, datasets, Spaces; de-facto model distribution layer
• vLLM — PagedAttention-based inference server; de-facto production serving engine (Part X)
• SGLang — structured generation with fast prefix-caching; competitor to vLLM at scale
• LangChain / LlamaIndex — RAG pipelines and agent orchestration (Parts XVIII, XIX)
• Triton — Python DSL for GPU kernels; Inductor backend in torch.compile (Part XIV)
• PyTorch / JAX — training frameworks; PyTorch dominates research, JAX at Google
• Megatron-LM — NVIDIA's reference for 3D-parallel pre-training (Part XI)
• DeepSpeed — Microsoft's ZeRO optimizer and pipeline-parallel framework (Part XI)

4.4 Hardware Ecosystem

NVIDIA holds ~80% of AI training compute through the CUDA moat: NCCL collectives, cuBLAS/cuDNN kernel libraries, FlashAttention CUDA extensions, Megatron-LM, DeepSpeed, and all of torch.cuda are optimized for NVIDIA hardware first.

5. Job Map & Curriculum Priority

Different roles hire on very different subsets of this curriculum. The diagram below shows the primary career paths; the table maps each role to the curriculum parts that interviewers will actually quiz you on.

6. Minimal Demo — Job-Role Skill Radar

Enter a role number (1–4) on stdin and the program prints a priority bar for all 22 curriculum parts. Try switching roles to see how different the study plans are: a Kernel Engineer (3) needs almost no RAG knowledge, while an Applied AI Engineer (4) barely needs CUDA depth.

7. Production & Source Pointers

Where to find the canonical implementations of the concepts introduced on this page:

• PyTorch nn.Transformer — torch/nn/modules/transformer.py. The canonical reference for multi-head attention + encoder/decoder.
• Hugging Face transformers — src/transformers/models/— one sub-directory per architecture (llama, gpt2, mistral, gemma, qwen2, …).
• vLLM engine — vllm/engine/llm_engine.py— LLMEngine class; read to understand continuous-batching + PagedAttention glue.
• Megatron-LM — megatron/core/tensor_parallel/— column / row parallel linear layers as described in Narayanan et al. 2021.
• FlashAttention — flash_attn/flash_attn_interface.py; CUDA kernel in csrc/flash_attn/.
• nanoGPT — model.py (github.com/karpathy/nanoGPT) — 300-line GPT-2 you can read in one sitting.
• DeepSeek-V3 — inference/model.py (github.com/deepseek-ai/DeepSeek-V3) — MLA attention + MoE routing in ~800 lines.

8. References

Papers

• Vaswani et al. 2017 — Attention is All You Need. arXiv:1706.03762
• Dosovitskiy et al. 2020 — An Image is Worth 16×16 Words (ViT). arXiv:2010.11929
• Radford et al. 2021 — Learning Transferable Visual Models From Natural Language Supervision (CLIP). arXiv:2103.00020
• Gulati et al. 2020 — Conformer: Convolution-augmented Transformer for Speech Recognition. arXiv:2005.08100
• Chen et al. 2021 — Decision Transformer: Reinforcement Learning via Sequence Modeling. arXiv:2106.01345
• Jumper et al. 2021 — Highly accurate protein structure prediction with AlphaFold. Nature 596.
• Brown et al. 2020 — Language Models are Few-Shot Learners (GPT-3). arXiv:2005.14165
• Ouyang et al. 2022 — Training language models to follow instructions with human feedback (InstructGPT). arXiv:2203.02155
• DeepSeek-AI 2024 — DeepSeek-V3 Technical Report. arXiv:2412.19437
• DeepSeek-AI 2025 — DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL. arXiv:2501.12948
• Shazeer et al. 2017 — Outrageously Large Neural Networks: The Sparsely-Gated MoE Layer. arXiv:1701.06538

Lectures

• Stanford CS221 (Intro to AI) — Percy Liang — covers AI taxonomy and sub-field survey
• Stanford CS224n (NLP with Deep Learning) — unit 1: from classical NLP to transformers
• Stanford CS336 (Building LLMs from Scratch) — lecture 1: full LLM landscape overview
• MIT 6.S191 (Intro to Deep Learning) — lecture 1: DL ecosystem and industry landscape
• Berkeley CS294-196 (LLM Agents) — lecture 1: LLM landscape and agent paradigms
• Karpathy "State of GPT" (Microsoft Build 2023) — 90-minute pipeline walkthrough
• Oxford Deep Learning for NLP (Hilary term) — tutorial 1: overview of NLP sub-fields

Textbooks

• Russell & Norvig — Artificial Intelligence: A Modern Approach (4th ed.) — Part I: overview; Appendix: AI sub-fields
• Goodfellow, Bengio, Courville — Deep Learning (MIT Press 2016) — Chapter 1: Introduction and landscape
• Jurafsky & Martin — Speech and Language Processing (3rd ed. draft) — Chapter 1: NLP overview

Blog Posts & Reports

• AI Index 2024 Annual Report (Stanford HAI) — compute trends, lab landscape, benchmark progress
• Anthropic "Core Views on AI Safety" (anthropic.com) — mission and AGI-risk framing
• OpenAI Charter (openai.com/charter) — AGI definition used by OpenAI
• Epoch AI "Machine Learning Hardware" tracker — GPU capability and cost over time
• Semianalysis "The AI Chip Landscape" — deep dives on NVIDIA / AMD / custom silicon economics

Code & Repos

• github.com/huggingface/transformers — all major LLM architectures in one repo
• github.com/vllm-project/vllm — PagedAttention inference server
• github.com/NVIDIA/Megatron-LM — 3D-parallel pre-training reference
• github.com/karpathy/nanoGPT — 300-line GPT-2 reference implementation
• github.com/deepseek-ai/DeepSeek-V3 — MLA + MoE architecture in production

9. Interview Prep

These questions appear in ML breadth / system-design screens across frontier labs: