Generative AI Roadmap for Beginners
A learning roadmap for generative AI that starts with what the technology actually is, moves through prompting and retrieval, and ends at evaluation and deployment — with the prerequisites stated up front.
Who this is for
Someone who can write Python and wants to work on generative AI seriously. If you do not write code yet, the zero to AI engineer roadmap is the better starting point.
Before you start
You need Python, git, and enough linear algebra to recognise what a vector is. You do not need to derive gradients, and you do not need any prior machine learning.
Phase 1: What the technology is
Understand tokens, embeddings, attention and next-token prediction well enough to debug output and reason about cost. This is where most self-taught people skip ahead, and it shows later.
- Read what is an LLM first.
Phase 2: Working with models
Prompting that goes beyond tricks, structured output, function calling, streaming, and the API details that bite: rate limits, token accounting, retries, and cost.
The most valuable habit here is testing. Write a small evaluation set before you start tuning prompts, or you cannot tell whether a change helped.
Phase 3: Retrieval
Chunking, embeddings, vector search, hybrid retrieval, reranking, and how to tell whether a bad answer is a retrieval problem or a generation problem. That distinction saves more time than anything else in this phase.
Phase 4: Evaluation
The skill that separates people who can demo generative AI from people who can ship it. Build test sets, score outputs, and wire evaluation into your change process so regressions get caught.
Phase 5: Deployment
Serving, caching, cost budgets, monitoring and safe rollout of model changes. This is LLMOps, and it is a discipline of its own.
Phase 6: Going deeper
Pick a direction: agentic systems, fine-tuning and model adaptation, or AI system design. These diverge here, and the right choice depends on the problems you want to work on rather than on which is trendier.
How long this takes
Working steadily, six to nine months gets you from first principles to a solid professional level. The variance is wide, and the difference is almost entirely whether you build things. Reading about agents is not the same as having debugged one.
Go deeper
- Pick a track with the AI engineering course.
- See the AI engineer roadmap for the placement-oriented version.
Related reading
AI Photo: How Artificial Intelligence Creates and Edits Images
What an AI photo is, how text-to-image generation works, and why the LLMs, RAG and agents behind it are where the real engineering starts.
How to Become an AI Engineer: 90-Day Roadmap
A week-by-week 90-day roadmap from programming basics to your first deployed AI project — Python, ML fundamentals, LLMs, RAG, and a portfolio you can defend.
Build Your First AI Agent in an Afternoon
Build your first AI agent in an afternoon — a working agent needs tools, a loop, and clear boundaries, and here is the simplest possible version to start.