A reference map

Paths of AI development specialization

These are different paths, not steps. Each is a legitimate home for a career — the ordering just organizes where each one sits on the spectrum from making AI to using it. The topics under each path are starting points: specific enough to search, study, and go deep on.

Making AI Using AI

01 02 03 04 5a 5b 5c 06

Foundational AI research

Inventing new architectures, training methods, and theoretical understanding. The artifact is knowledge — papers and ideas, sometimes reference implementations.

Key topics: transformer internals · attention mechanisms · scaling laws · optimization theory · mechanistic interpretability · sparse autoencoders · activation patching · experimental design · reading & reproducing papers
Examples: The Transformer paper · Chinchilla scaling laws · Anthropic's circuits work · RL from human feedback

Frontier model development

Actually training the large models. Blends research with enormous engineering effort. The artifact is a model.

Key topics: pretraining pipelines · tokenization · data mixing & curation · SFT · RLHF · DPO · Constitutional AI · reward modeling · synthetic data generation · compute planning · safety red-teaming
Examples: Anthropic · OpenAI · Google DeepMind · Meta FAIR · xAI

ML systems & infrastructure

Building the systems that make training and inference possible. Deeply technical but largely separate from research questions about what to train.

Key topics: CUDA · Triton · FlashAttention · PagedAttention · quantization (GPTQ, AWQ, FP8) · speculative decoding · KV cache management · FSDP · tensor & pipeline parallelism · continuous batching · NCCL / InfiniBand
Examples: vLLM · SGLang · TGI · Megatron-LM · DeepSpeed

Model adaptation

Specializing existing foundation models for a domain. Less common than it was, still relevant where base models fall short or have to run somewhere specific.

Key topics: supervised fine-tuning · LoRA / QLoRA · DPO · PEFT methods · knowledge distillation · synthetic data generation · domain-specific eval design · on-device deployment (GGUF, MLX, CoreML)
Examples: medical / legal fine-tunes · on-device models · code-specialized models

AI infrastructure & tooling

Building the layer other AI developers build on top of. Your users are other engineers — so you codify patterns, abstractions, and primitives that others will depend on.

Key topics: agent loop patterns (ReAct, tool use) · streaming (SSE) · function calling semantics · context window management · prompt caching · vector DBs · hybrid retrieval & reranking · MCP (Model Context Protocol) · observability & tracing · eval harnesses
Examples: LangChain · Vercel AI SDK · Mastra · Braintrust · Langfuse · OpenRouter

AI-native products

The product is the AI interaction. Strip out the model and there's nothing left. You live and die by prompt quality, context engineering, latency, and the shape of the agent loop.

Key topics: prompt engineering (few-shot, CoT, XML tags) · context engineering · token budgeting · RAG (chunking, embeddings, reranking) · agent design (planning, reflection, memory) · eval-driven development · structured outputs · latency optimization · model routing & cascades · hallucination mitigation
Examples: Cursor · opencode · t3chat · Perplexity · Claude.ai · Character.AI

AI features in traditional products

Bolting AI onto software that would exist without it. The core value prop isn't AI — it's payroll, analytics, project management, whatever — and AI is one feature among many.

Key topics: prompt engineering basics · guardrails & moderation · simple RAG · cost & latency budgeting · A/B testing AI features · fallback & error handling · AI disclosure UX · user feedback loops
Examples: Notion AI · Slack AI · Salesforce Einstein · in-app support bots · "ask your data" in BI dashboards

AI-augmented software engineering

Using AI tools to build regular software where the AI isn't part of the final product — it's a coworker, like an IDE. The artifact is normal software; AI is invisible to the end user.

Key topics: Claude Code / Cursor / Copilot workflows · context files (CLAUDE.md, AGENTS.md) · agentic coding patterns · TDD with AI · reviewing AI-generated code · delegation judgment · sub-agents & slash commands
Examples: shipping any normal app with Claude Code · Cursor · Copilot as a coworker

Overlays

Cross-cutting specializations

These don't sit on the spectrum — they cut across it. A safety researcher might work inside a frontier lab or an AI-native product team. An evals engineer might sit at a tooling company or inside a traditional SaaS team.

// 01

Safety & alignment

RLHF · Constitutional AI · sparse autoencoders · activation patching · circuit analysis · scalable oversight · deceptive alignment · capability evaluations · responsible scaling policies.

// 02

Evals & reliability

benchmark design · LLM-as-judge · pairwise comparison · rubric-based grading · statistical significance · regression testing · golden datasets · observability tooling.

// 03

AI product & UX

streaming UX · citation & provenance display · uncertainty communication · error recovery patterns · multi-turn conversation design · progressive disclosure · trust signals · input affordances for AI.

// 04

Policy & governance

AI policy research · EU AI Act · NIST AI RMF · responsible scaling policies · frontier safety commitments · third-party evaluations · standards development · government engagement.