How to Learn AI From Scratch (Realistic Roadmap)

Search “learn AI” and you’ll find two kinds of lies: the six-week-bootcamp fantasy, and the gatekeeping myth that you need a math PhD to begin. The truth is a middle path that takes real months of consistent work — and is entirely walkable by a motivated beginner with a plan. This is that plan, with honest timeframes attached to every step.

A note on honesty before we start: we research these programs’ curricula, instructor track records, and learner outcomes — we haven’t sat every course end-to-end, and we won’t pretend otherwise. And where this roadmap names timeframes, they’re planning ranges for consistent part-time effort (call it 8–12 hours a week), not promises. Double the hours and things move faster; halve them and they don’t.

Before step one: decide what “learn AI” means for you

“AI” has become three different goals wearing one name, and picking yours changes the roadmap’s back half:

• AI literacy — using AI tools well at work. Weeks, not months. You don’t need this roadmap; you need a short course and practice (start with our best free AI courses and a weekend on prompting via the best prompt-engineering courses).
• Building with AI — ML fundamentals plus the ability to build applications with models, including LLMs. This roadmap, specializing toward applied AI at step 4.
• Data science/analytics — statistics, SQL, and analysis, with ML as one tool among several. This roadmap through step 3, then branching toward the data lane (our data analytics certifications guide maps that world).

The steps below serve the second and third goals — the ones that take months and change careers.

Step 1: Foundations — Python plus math intuition (months 0–3)

Two parallel tracks, and interactive beats passive for both.

Python. Every serious ML course assumes it, and learning Python while learning ML concepts doubles the difficulty of both. You need: variables, loops, functions, data structures, and a first pass at NumPy and pandas. You explicitly do not need to be a software engineer before starting ML — that’s the over-preparation trap, and it strands as many people as under-preparation does.

Interactive platforms are the right tool because they force you to type code from day one:

Our DataCamp vs Codecademy comparison breaks down which fits which learner, and our best Python courses for data guide covers the wider field including free options.

Math intuition. Here’s the honest version of the math question: you need working intuition for three things — linear algebra (what vectors and matrix operations do), probability (how uncertainty behaves), and derivatives (how change propagates, which is what “training” is). You do not need proofs. The applied path runs on understanding what the operations mean, and good ML courses refresh the mechanics as they go.

Free alternatives exist and are excellent — Khan Academy for structured math, and 3Blue1Brown’s Essence of Linear Algebra videos for intuition that sticks. Budget-zero learners can do this entire step free; Brilliant buys convenience and consistency, not access.

Timeframe honesty: 2–3 months part-time for both tracks together. Comfortable coders can compress to weeks; true zero-to-code beginners should let it take the full three months without shame.

Step 2: One core ML course, completed (months 3–6)

This is the roadmap’s keystone, and the instruction is deliberately narrow: one course, finished completely, every assignment done. Course-hopping is the most common failure mode in self-taught AI — five sampled introductions leave you with less than one completed course, because the compounding happens in weeks four through ten, exactly where samplers quit.

The default pick is the Machine Learning Specialization from DeepLearning.AI, hosted on Coursera — Andrew Ng’s modernized version of the course that onboarded a meaningful share of the people now working in the field. It’s beginner-friendly without being hollow: supervised learning, the workhorse algorithms, neural network fundamentals, and — most valuably — the practical judgment layer of why models fail and what to try next.

Practical notes: it’s billed the standard Coursera way — auditable free, paid for graded work and the certificate, and included in Coursera Plus. If your plan includes more Coursera programs after this one (step 4 likely does), read is Coursera Plus worth it before paying per-program — a multi-program year is exactly the case where the subscription wins. Alternatives by learning style — more visual, more hands-on, more academic — live in our best AI courses for beginners guide; pick one and apply the same rule: finish it.

Timeframe honesty: 2–3 months part-time. The advertised course hours undercount reality — assignments where you’re stuck are where the learning is, and being stuck takes time.

Step 3: Projects on data the course didn’t give you (months 5–8)

Course assignments are learning with a safety net: clean data, known answers, guided steps. Real skill starts when the net comes off. Immediately after (or during the tail of) your core course, build two or three small projects on data you chose — scraped, downloaded from a public source, or pulled from your own life or job.

What makes a good first project is smallness plus genuine friction: predict something about data you care about, classify something from a messy public dataset, analyze something from your workplace (nothing confidential). The goal is to hit the problems courses shield you from — data that needs cleaning, approaches that don’t work, ambiguity about what “good enough” means — and get through them anyway. That struggle is the skill. It’s also, not coincidentally, exactly what interviewers probe for: “tell me about a problem you hit and how you got past it” has no good answer that comes from coursework alone.

Two rules: keep each project under a month (scope creep kills more portfolios than laziness), and write down what failed as you go — you’ll want it for step 5.

Timeframe honesty: 2–3 months, overlapping the end of step 2. If a project is on month two, cut its scope in half and ship it.

Step 4: Specialize — LLMs and applied AI, or data (months 8–12)

With fundamentals and first projects behind you, generalist learning hits diminishing returns. Pick the lane that pulled you hardest during steps 2–3:

The LLM/applied-AI lane. Building applications on top of large language models — the retrieval systems, agents, and AI features that most “AI engineer” work now involves. The move here: DeepLearning.AI’s short courses on LLM application patterns (many free) for breadth, plus one deeper program on building with model APIs, and — most importantly — one substantial project: a working application that does something real, end to end.

The data lane. Deeper statistics, SQL fluency, and the analytics stack — the path where AI skills meet the enormous, stable market for data work. Structured options: a recognized certificate as spine — the Google Data Analytics Certificate is the famous on-ramp, with honest caveats we’ve written up separately — or a fuller career track:

Either lane, same rule as step 2: one substantial program plus one substantial project beats four introductions.

Timeframe honesty: 3–4 months part-time to real depth in one lane. Trying both lanes simultaneously reliably delivers neither.

Step 5: A portfolio that shows reasoning (month 10+, ongoing)

Now package the evidence. Three to five projects, public (GitHub is the default home), each with a write-up a non-expert could follow: the problem, why it’s non-trivial, your approach, what failed along the way, and what you’d do next. That failure section isn’t confessional decoration — it’s the single strongest hire-me signal a self-taught candidate has, because it’s the part that can’t be faked by following a tutorial.

Hard truths that save you months: a wall of certificates is not a portfolio — credentials get you glanced at, projects get you interviewed. The shared capstone from a popular course is not differentiation — thousands of identical ones exist. And a portfolio nobody can find is a diary — put the link on your LinkedIn and resume, and consider writing up one project as a post; explaining things publicly is both proof of skill and how opportunities find you.

Step 6: Maintenance mode (forever, honestly)

The field will move under your feet — that’s the deal you signed. Sustainable currency, part-time: follow a small number of high-signal sources rather than drowning in AI news; take a short course when something genuinely new arrives (DeepLearning.AI’s free short-course library is built for exactly this); and keep shipping a small project every couple of months. A few hours a week of maintenance beats quarterly panic-cramming, and beats it badly.

The realistic timeline, assembled

• Months 0–3: Python + math intuition (step 1)
• Months 3–6: core ML course, completed (step 2)
• Months 5–8: first independent projects (step 3)
• Months 8–12: specialization + substantial project (step 4)
• Months 10+: portfolio assembly, then maintenance (steps 5–6)

Total: a year, give or take, at consistent part-time hours — down toward six months for full-time-ish effort with prior coding, out toward eighteen for busy lives and true zero starts. Both ends of that range represent success. The only failure mode is the stall, and the antidote to stalling is the boring one: scheduled hours, one course at a time, projects small enough to finish.

Our honest take

Learning AI from scratch in 2026 is one of the highest-return projects a motivated person can take on — and it’s a project, with a project’s timeline, not a purchase with a checkout page. Some links here may earn us a commission, so note what we’re really telling you: most of this roadmap can be walked free or nearly free, the paid tools buy structure and convenience rather than access, and nothing you can buy substitutes for the projects only you can build.

Start with step 1 this week — a Python exercise and twenty minutes of math intuition — and let the streak do the rest. When you’re choosing your core course, our beginner AI course guide is the next read, and everything we publish on this path lives at the AI & Data Skills hub.