How to Learn AI/ML From Scratch (No PhD Required)

You do not need a PhD to work in AI. You do not even need a Masters. The most in-demand AI roles in 2026 (AI Engineer, AI Product Engineer) hire self-taught engineers all the time.

What you do need: real skills, real projects, and a clear plan. Here is exactly how to learn AI from scratch in 6 to 9 months.

What You Are Actually Learning#

Three distinct skill sets are often confused under "AI":

1. AI Engineering: building products with existing models (LLMs, vision models, etc.)
2. ML Engineering: building and deploying custom ML models
3. AI Research: creating new models and techniques

The first one (AI Engineering) is the most accessible and most in-demand. The second is harder and requires more math. The third requires a PhD.

This guide is for AI Engineering. If you want ML Engineering or research, you need additional math depth.

Prerequisites#

You need basic programming. Specifically:

1. Python at intermediate level (loops, classes, list comprehensions, async)
2. Comfort with the command line
3. Ability to use Git and GitHub
4. Familiarity with at least one web framework (Flask, FastAPI, or similar)
5. Basic SQL

If you do not have these, spend 1 to 2 months on them first. Free resources:

• Python: realpython.com
• Command line: missing.csail.mit.edu
• Git: learngitbranching.js.org
• FastAPI: their official docs
• SQL: sqlbolt.com

The 6-Month Roadmap#

Month 1: Foundations

Goal: understand how modern AI works at a conceptual level.

Resources:

1. Watch Andrej Karpathy's "Zero to Hero" YouTube series (free, ~25 hours)
2. Read "Designing Machine Learning Systems" by Chip Huyen
3. Skim the original Transformer paper ("Attention is All You Need")

Build:

• A simple language model from scratch in Python (Karpathy's nanoGPT tutorial)
• Do not worry about training it well, just understand the code

By end of month 1, you should be able to:

• Explain what a transformer does conceptually
• Read a paper abstract and roughly understand it
• Write Python that calls the OpenAI API

Month 2: Build with LLMs

Goal: get comfortable building features with existing models.

Resources:

1. OpenAI Cookbook on GitHub (free)
2. Anthropic's prompt engineering docs (free)
3. Read Eugene Yan's writings on LLM patterns

Build:

• A "talk to my notes" chatbot using embeddings and retrieval
• Use FAISS or pgvector for vector storage
• Build it as a real app with a UI (Streamlit or Next.js)

Specific tasks:

1. Get OpenAI API key
2. Build the chatbot in 2 weeks
3. Deploy it (Vercel free tier works)
4. Document the project in a README

By end of month 2, you should have shipped 1 real project.

Month 3: RAG Deep Dive

Goal: master retrieval-augmented generation, which is what 80% of AI engineer jobs work on.

Resources:

1. LangChain docs (free, but skim)
2. LlamaIndex docs (free)
3. Anthropic's blog on RAG
4. Pinecone's RAG learning resources

Build:

• Extend your chatbot from month 2
• Try different embedding models (OpenAI, Cohere, open source)
• Try different vector databases (FAISS, Pinecone, pgvector)
• Try different retrieval strategies (basic, hybrid, reranking)
• Compare results

Key concepts to learn:

1. Chunking strategies (size, overlap, semantic)
2. Embedding models (closed vs open source, dimensions)
3. Vector similarity (cosine, dot product, euclidean)
4. Hybrid search (semantic + keyword)
5. Reranking with cross-encoders
6. Query rewriting

By end of month 3, you should have a chatbot that performs significantly better than the month 2 version, and you should know why.

Month 4: Evaluation

Goal: learn to measure if AI systems actually work.

This is the most undervalued skill. Most AI engineers cannot evaluate their systems. If you can, you stand out.

Resources:

1. Eugene Yan's blog post on LLM evals
2. Anthropic's tool use and eval documentation
3. HumanEval, MMLU benchmarks (read about them, do not focus on the numbers)

Build:

• Create an evaluation set for your chatbot (50 to 100 questions with ideal answers)
• Build a script that runs your chatbot on the eval set
• Use an LLM-as-judge (GPT-4 or Claude) to score outputs
• Build a dashboard showing pass rate
• Iterate on your chatbot and watch the score change

This is the single most impressive thing you can put on a resume in 2026.

Month 5: Production AI

Goal: understand what it takes to run AI in production.

Topics:

1. Cost optimization (token counting, caching, model routing)
2. Latency optimization (streaming, batching, prompt compression)
3. Observability (LangSmith, Helicone, custom logging)
4. Safety (prompt injection, jailbreaks, output filtering)
5. Reliability (retries, fallbacks, circuit breakers)

Build:

• Take your chatbot and add all of the above
• Set up a logging dashboard
• Add a fallback to a cheaper model when GPT-4 is too slow
• Add basic prompt injection defenses

Resources:

• LangChain's production guide
• Eugene Yan's "patterns for building LLM systems"
• The book "Designing Data-Intensive Applications" (not AI-specific but invaluable)

Month 6: Build a Portfolio

Goal: have 3 to 5 strong projects that get you interviews.

By month 6, you should have:

1. The talk-to-my-notes chatbot (refined)
2. One agent project (an LLM that uses tools)
3. One eval framework you built
4. One open source contribution to an AI library
5. One blog post explaining a technical concept

Spend month 6 polishing these. Write READMEs. Add demos. Get them on GitHub with good documentation.

Then start applying.

Resources Worth Your Money#

Free resources are abundant. Most are great. A few paid resources are worth it.

Definitely free

1. Andrej Karpathy's YouTube channel
2. Sebastian Raschka's blog and books
3. Anthropic's documentation
4. OpenAI cookbook
5. Eugene Yan's writings
6. Hugging Face course
7. Fast.ai

Worth paying for

1. Coursera Deep Learning Specialization ($50/month, 3 months) - $150 for a credential
2. DeepLearning.AI short courses ($50 each) - specific topics
3. Latent Space podcast and community ($300/year) - network value
4. ChatGPT Plus or Claude Pro ($20/month) - daily AI partner

Total cost over 6 months: less than $1k if you are frugal.

Skip

1. Most Udemy courses with "AI Master" in the title
2. LinkedIn Learning AI courses (shallow)
3. Cheap bootcamps under $2k
4. Generic "Prompt Engineering Certifications"

Building Real Projects (Not Tutorial Clones)#

The biggest mistake new AI learners make: they follow tutorials and never build anything original.

Tutorial output: a "ChatGPT clone" that does what 1M tutorials show.

Real project: a tool that solves a problem in your life or work.

Examples of real projects:

1. A "rabbit hole" research agent that explores a topic deeply
2. A code review bot for your specific codebase
3. A personalized news summarizer for your interests
4. A meeting notes processor that extracts action items
5. A study assistant for a specific subject you are learning
6. A LinkedIn DM assistant that drafts replies

The litmus test: does this project solve a real problem? If yes, build it. If no, pick a different one.

How to Network While Learning#

Visibility helps you get hired.

1. Tweet about what you are building (X/Twitter is the AI engineer's home)
2. Write blog posts (Substack or your own site)
3. Open source your code on GitHub
4. Join AI Discord communities (Anthropic, Eleuther AI)
5. Attend AI meetups in your city
6. Comment thoughtfully on AI papers and blogs

You do not need 10k followers. You need 3 to 5 well-connected people who know you build things.

Common Pitfalls to Avoid#

Pitfall 1: Too much theory, no building

Reading papers and watching lectures is comfortable. Building is hard. Force yourself to build every week.

Pitfall 2: Building tutorial clones

You learn nothing from cloning. Start with tutorials, then deviate immediately.

Pitfall 3: Chasing every new framework

LangChain, LlamaIndex, AutoGen, CrewAI, DSPy. New frameworks weekly. Pick one and go deep before moving to another.

Pitfall 4: Ignoring math entirely

You do not need a PhD. But you should understand:

• What an embedding is (a vector)
• Why cosine similarity works
• Why temperature affects output
• What perplexity measures
• What attention does conceptually

Skipping these makes you a shallow AI engineer.

Pitfall 5: Avoiding the messy parts

Production AI is messy. Caching, retries, fallbacks, monitoring. Bootcamps skip this. Real jobs are 50% this.

Landing Your First AI Job#

After 6 months of focused learning and building, you are ready to apply.

Optimize your resume

Your resume should emphasize:

1. The 3 to 5 projects you built (with links)
2. Specific tools and APIs you used
3. Quantified outcomes ("Built RAG system over X dataset, achieved Y precision")
4. Open source contributions

Run it through the free ATS checker against AI Engineer JDs. Iterate.

Apply to 30 to 50 roles

Mix of:

• 10 AI-first startups (highest reply rate)
• 10 mid-size tech companies (Square, Datadog, Snowflake)
• 10 enterprise AI teams (Capital One, JPMorgan)
• 5 to 10 top AI labs (long shots)

Use the free AI cover letter generator for each.

Network in parallel

For each company you apply to, find a senior engineer on LinkedIn. Send a thoughtful message asking about their team. Half will reply.

Prepare for AI-specific interviews

Use the free mock interview tool to practice:

• "How does RAG work?"
• "Walk me through how you would build X"
• "What are the tradeoffs between temperature 0 and 0.7?"
• "How do you evaluate an LLM output?"

These are standard AI engineer interview questions.

Salaries to Expect#

For self-taught AI Engineers in 2026, at first job:

• US base: $130k to $200k
• Total comp at startups: $150k to $250k
• Total comp at FAANG: $200k to $350k (harder to get into)
• UK base: 70k to 110k GBP
• India base: 18 to 35 LPA

Self-taught engineers can earn the same as bootcamp grads. The market does not care how you learned, only what you can do.

What to Do This Week#

If you are starting from zero:

1. Set up a Python dev environment
2. Sign up for OpenAI API access (free credits for new accounts)
3. Watch the first 3 hours of Karpathy's "Zero to Hero"
4. Sketch out what your first project will be

If you already know Python:

1. Skip to month 2 of the roadmap
2. Start building your first chatbot today
3. Open a GitHub for your AI work

The market for AI engineers is the best market in tech right now. The gap between people who think about learning AI and people who actually do is enormous. Be on the right side of that gap.