Five AI Roles, Five Free Resources — Read Your Row, Skip Four

Open the prerequisites page Google publishes for its Machine Learning Crash Course and read what it actually asks of you: algebra and logarithmic equations, tensor rank, matrix multiplication, mean/median/standard deviation, plus enough Python to run exercises against NumPy, pandas and Keras. The course behind that gate is genuinely good — Google rebuilt it in November 2024 into a free 15-hour self-study run with more than 130 exercise questions, adding large language models, AutoML and expanded coverage of data and responsible AI.

It is also the link somebody pastes into every "how do I start with AI" thread I see in Brazilian dev groups. And for four of the five roles a Brazilian dev or EM is realistically hired into right now, it is the wrong first click.

That is the flaw in free-resource lists, and it isn't the resources. Six good links in a flat row hand you the hardest decision — which one is yours — so you bookmark all six and open none. Below is the same catalogue, mapped to roles, with a skip line on every entry. Read your row and ignore the other four.

My rule of thumb: choose by the output the role gets judged on, not by how advanced you feel. Feeling behind pushes you toward the most technical link on the page, which is exactly how people spend three weekends on gradient descent for a job that will never ask them about it.

Role 1 — You translate AI for people who don't build it

EM, PM, designer, ops. You get judged on decisions and on the vocabulary you use to defend them — in English, in a meeting, with a hiring manager on the other side.

• Elements of AI — built by the University of Helsinki with MinnaLearn, no coding and no math required, and it ships as two courses in sequence: Introduction to AI, then Building AI. Over 2 million people have signed up and its graduates come from more than 170 countries, which is the part that matters for an EU-targeted hunt — a Dutch or German hiring manager has plausibly heard of it. Best for: the manager who has to hold the conversation, not write the code. Skip if: you already ship anything that calls a model — chapter one will bore you into quitting.
• A Brazilian footnote worth knowing: the Portuguese edition lives on elementsofai.pt — that's Portugal, not Brazil. Since the vocabulary you'll actually need in an interview is English anyway, do it in English and count the reading as practice for the screen.

Role 2 — You drive coding agents all day

You already live inside an agent, and you've hit the ceiling where it does 80% of the task and then quietly wrecks the last 20%.

• Anthropic Academy — three tracks (Build with Claude, Claude for work, Claude for personal) and certificates on completion. Inside the catalog there's a real ladder rather than a pile: Claude Code 101 → Claude Code in Action → Introduction to agent skills → Introduction to subagents. Best for: the dev who uses Claude Code daily and still can't explain why some runs go sideways.
• Claude Code in Action specifically — the registration button says FREE, and the curriculum is 11 sections that read like a senior engineer's checklist: Steer the Work, Steering Long Sessions, Configure Claude, A CLAUDE.md That Follows, Verification Skills, Permission Modes, Hooks, Automate Repeat Work, Routines and Headless, GitHub Actions and Code Review, Plugins — plus a quiz. Best for: closing the gap between "I use the tool" and "I let the tool run unsupervised." Skip if: you got pointed at the AI Fluency for educators / students / nonprofits editions — same framework, wrong audience for you.

Role 3 — You ship features that call a model

There's a ticket on your board this month with an API key in it. You get judged on whether the thing works in production and what it costs to run.

• OpenAI Cookbook — 82 entries as I write this (13 articles plus 69 examples), spanning agents, evals, RAG, fine-tuning, vision, voice and code modernization. The detail that turns it from a bookmark into a tool: every entry has a Markdown twin at /cookbook/<path>.md, and there's a single-file Markdown export of the lot. You don't sit and read that — you hand it to your own agent as reference so its first draft is written against a real recipe instead of its memory. Best for: someone with a ticket, not a curiosity. Skip: reading it front to back. It's a reference, and treating a reference like a course is how a Saturday disappears.
• Same rung on the Claude side: Building with the Claude API and Introduction to Model Context Protocol in the Anthropic Academy catalog, with Model Context Protocol: Advanced Topics after them.

Role 4 — You review ML work you will never do yourself

Staff engineer, EM, tech lead. Nobody expects you to train the model. Everybody expects you to smell it when an estimate or a metric is nonsense.

• Google's Machine Learning Crash Course — now it's the right row instead of the wrong one. 15 hours, 130+ exercise questions, and since the rewrite the modules run from linear and logistic regression through numerical/categorical data and overfitting into neural networks, embeddings and large language models, then production systems, AutoML and fairness. Best for: the EM who needs to push back on a data scientist's plan without bluffing. Skip: the Keras programming exercises if you'll never own a training job — read the visualizations and the data modules, because that's where the arguments in your standup actually come from.

Role 5 — You're deliberately switching tracks

Not "getting better with AI" — changing what your title says.

• Microsoft AI For Beginners — 12 weeks, 24 lessons, MIT license, authored by Dmitry Soshnikov, with notebooks in both PyTorch and TensorFlow/Keras. It teaches the field rather than a product: symbolic AI, neural networks, computer vision, NLP, genetic algorithms and reinforcement learning, plus a section on AI ethics. It's translated into 50+ languages and the repo carries a real pt-BR folder, not only pt-PT. Best for: the person who wants the discipline, not the tool. Skip it if: you can't name the thing you'll have built by lesson 24 — a 12-week curriculum with no target is the most expensive free resource on this page.

Choose in 60 seconds

PICK ONE — answer in order, stop at your first YES

1. Do you have to explain AI to people who don't build it?
   → Elements of AI (Helsinki + MinnaLearn). No code, no math.
2. Do you run a coding agent daily and keep hitting a ceiling?
   → Anthropic Academy: Claude Code 101 → Claude Code in Action
     → Introduction to agent skills → Introduction to subagents.
3. Is there a ticket on your board this month that calls a model API?
   → OpenAI Cookbook (+ Building with the Claude API for Claude).
4. Do you review, estimate or challenge ML work you don't do yourself?
   → Google ML Crash Course. Read the modules, skip the Keras exercises.
5. Are you changing tracks into ML on purpose?
   → Microsoft AI For Beginners. 24 lessons — and write your target
     project down before lesson 1.

No YES anywhere? You don't need a course this week. You need a task.

The failure mode that costs the most

It isn't picking the wrong course. It's picking three. Two of these five resources are references (the Cookbook, and Google's crash course once you drop the exercises) and three are sequential courses — and being 20% into a sequential course, three times over, is worth precisely nothing to your skills or your CV. One row, one pass. Then the next role, if you actually move into one.

Today: run the five questions above out loud and stop at your first yes. Open that single link, find the first exercise or recipe that touches something already on your board this week, and do only that. Close the other four tabs before you start — that part is the whole technique.

Sources

• Google — Machine Learning Crash Course: prerequisites and prework
• Google Developers Blog — Machine Learning Crash Course gets new updates
• Elements of AI — University of Helsinki and MinnaLearn
• Anthropic Academy — learning tracks
• Anthropic Academy — course catalog
• Anthropic Academy — Claude Code in Action
• OpenAI Cookbook — recipes and examples
• Microsoft AI For Beginners — 12-week, 24-lesson curriculum