Big Tech Is Giving Away the AI Training Most Engineers Are Skipping

You've probably scrolled past a post about "free AI certifications from Google" at least five times this year. Most of them are the same thirty-minute intro course with a shareable badge attached. What almost nobody is talking about is that Anthropic, NVIDIA, and OpenAI quietly built out real technical training catalogs, free, and they map onto exactly the skills that keep showing up in AI engineering job postings.

Anthropic Runs a Course Platform Most Engineers Never Open

Anthropic isn't only shipping model updates. It maintains a full course platform at anthropic.skilljar.com, and almost none of the engineers using Claude daily know it exists.

This matters because the courses are built by the same team that ships the product, not a training vendor reverse-engineering the API from documentation. Claude 101 covers how the model actually reasons. Claude Code 101 walks through building AI-powered applications. Claude Platform 101 goes into the API and integration layer directly. That's a different kind of training than a generic "intro to LLMs" course: it's closer to reading the commit history of the thing you're building on top of.

Picture a team shipping a customer support agent on top of Claude's API without ever going through Platform 101. They misjudge context window behavior, over-stuff the prompt with irrelevant history, and the agent starts hallucinating answers under load, right as a support ticket spike hits during a product launch. Nobody notices in staging because the test conversations are short. In production, with real multi-turn conversations, the failure shows up as wrong refund amounts and escalated complaints, and the postmortem traces back to a context management pattern the course covers in the first module.

Spend the hour on Platform 101 before you touch the API keys, not after something breaks in front of a customer.

NVIDIA's Catalog Is the Most Technically Dense of the Five

NVIDIA's free training, at nvidia.com/en-us/training/self-paced-courses, skips the fluff and goes straight to infrastructure-level concepts.

Three courses stand out: Generative AI Explained, a two-hour primer that doesn't require a machine learning background; Getting Started with Deep Learning, which covers foundational architecture from the company whose hardware runs most of the field; and Augmenting LLMs with RAG, a one-hour course that punches well above its runtime. RAG shows up constantly in job postings for AI engineering roles right now, and NVIDIA teaches the core mechanics, retrieval, chunking, embedding freshness, in an hour, with a certificate that means something to a recruiter scanning LinkedIn.

Think about the technical screen where a candidate is asked to design a retrieval pipeline for a support knowledge base. They can explain what RAG is conceptually, but they've never touched embedding drift or thought about what happens when the underlying documents change faster than the vector index refreshes. The interviewer asks one follow-up question about stale retrieval and the whole answer falls apart. That's not a knowledge problem you fix by reading more blog posts. It's a gap this specific course closes directly, because it's built around the failure modes, not just the definitions.

If you only have an hour this week for technical upskilling, this is the one to spend it on.

OpenAI Academy Is the First Platform Here That Assumes Competence

OpenAI Academy, at academy.openai.com, launched recently and structures its content by how deep you already are, instead of assuming everyone starts at zero.

AI Foundations is for someone brand new to using AI tools for work. Applied AI Foundations moves into building repeatable workflows instead of one-off prompts. Agents and Workflows, the advanced tier, is about directing AI agents through structured, multi-step tasks, and it doesn't waste time re-explaining what a prompt is. The tiering matters because most free training either targets total beginners or drowns everyone in jargon; this is one of the few that lets you skip straight to the tier that matches where you actually are.

An engineer already comfortable with prompting who burns a week working through AI Foundations out of habit, then rushes the Agents and Workflows material the night before building an autonomous triage agent, ends up shipping something that works fine on the happy path and falls over the first time it hits an ambiguous ticket, because the course section on handling multi-step failure states got skimmed instead of studied.

Skip the tier you've outgrown. Spend the saved time on the one that's actually going to be tested against a real deadline.

Google and Microsoft Are Optimizing for the Recruiter, Not the Engineer

Google's grow.google/ai and Microsoft Learn cover ground the other three platforms mostly ignore: getting past the resume filter in the first place.

Google offers AI Essentials, a five-hour course built as the cleanest on-ramp for someone starting from nothing, plus a course specifically about using AI tools in a job search. Microsoft Learn is free study material aimed at the AI-900 certification, alongside a Responsible AI principles course that's starting to appear as a listed requirement in job descriptions at larger companies. Neither is technically deep, and that's the point: these badges are built to be recognized instantly by an applicant tracking system or a non-technical recruiter doing a first pass, in a way that "completed a RAG tutorial on YouTube" never will be.

A strong candidate with a solid GitHub history and real project experience gets filtered out before a human ever reads the resume, because the job listing's screening keywords match "Microsoft Certified" and nothing on the resume matches that string. It's not a fair filter. It's also not going away, and it's free to satisfy.

Treat these two as resume infrastructure, not as a substitute for the technical depth NVIDIA and Anthropic provide.

What Job Postings Are Actually Asking For

A scan of more than 3,000 AI job listings surfaced a pattern worth paying attention to before picking which course to prioritize.

Python sat at the top of nearly every posting, which surprises no one. The bigger signal was buried further down: agentic AI, specifically LangChain, RAG pipelines, and multi-step agent orchestration, showed up constantly in postings for roles at the fresher end of the ladder, the ones candidates assume only need prompting skills. Most applicants for those roles hadn't touched the concept at all.

A team lead posts an opening for a mid-level AI engineer expecting basic RAG pipeline experience, and forty applications come in from people who can write a solid prompt but have never built a retrieval layer. The role stays open for three months, not because the skill is rare in absolute terms, but because almost nobody bothered to close the specific gap the job actually tests for.

That gap is exactly what NVIDIA's RAG course and OpenAI's Agents and Workflows tier are built to close, for free, this week.

Where to Actually Start

For someone already working at the AI Engineer level, the efficient path through this list isn't the beginner track, and it isn't all five platforms at once either.

NVIDIA's RAG Course

Start here. It's an hour, and it maps directly onto the skill gap employers are naming out loud in job postings right now.

Anthropic's Platform 101

Move here next if any part of your work touches the API layer directly. It prevents the exact context-management mistakes that show up as production incidents later.

OpenAI's Agents and Workflows Tier

Save this for last. It assumes you've already got the fundamentals of both retrieval and platform mechanics down, and it's wasted on someone who skips ahead without that base.

Stop treating "free AI course" as a synonym for beginner content. The technical tiers from NVIDIA, Anthropic, and OpenAI's advanced track are closing the exact gap that's showing up in the job market this year, and none of them cost more than a weekend. The next time a recruiter's technical screen asks you to reason about retrieval freshness or agent failure states, the difference won't be talent. It'll be whether you spent the hour.