Twenty Channels, Fifteen Instructions — And The Channels Are The Smaller Half
Yesterday we looked at a card of twenty YouTube channels for learning tech, and read it as a map of who owns which word — the niche principle, printed twenty times. That post is Nineteen Of These Channels Own One Word.
This card also has twenty channels on it. It is doing something else entirely, and the difference is worth a whole article.
Count what comes with it. Five groups, four channels in each — and three lines of instruction per group. Fifteen instructions against twenty names. And the opening line sets the brief for all of them: your AI learning plan should produce something you can use.
Now read the fifteen instructions and notice what is missing from every single one of them. Not one says watch. They say choose, recreate, write down, record, check, explain, test, define, build, add, review. The channels are the raw material. The instructions are the course.
The same three steps, five times
Put the five trios side by side and they have one shape.
| STEP 1 · DO ONE THING | STEP 2 · TEST IT AGAINST SOMETHING YOU DID NOT WRITE | STEP 3 · RECORD WHAT YOU STILL CANNOT DO |
|---|---|---|
| One course, one paper, one concept, one workflow, one evaluation | A fresh example, the original source, a tiny test, an incomplete input | The part you cannot explain, the assumption, the failure case, the error |
| Narrow enough to finish this week | Something that can disagree with you | A sentence, written down, that you keep |
| Without it you collect tabs | Without it you are grading your own homework | Without it you rewatch instead of advancing |
| Easy, and everybody does it | Harder, and some people do it | The one almost nobody does — and the only one that tells you what is next |
The five groups, as given
The group names, the channels and the “best for” lines are the card’s own; the three instructions are from the post beneath it.
| Group | The four channels | What they are best for | The three instructions |
|---|---|---|---|
| 1 · Build a foundation | codebasics · Krish Naik · freeCodeCamp.org · DeepLearning.AI | Practical AI roadmaps, ML and GenAI, long-form courses, certificates | Pick one course and put time on the calendar · recreate an example before changing the inputs · write down the part you still cannot explain |
| 2 · Understand the research | The AI Epiphany · Andrej Karpathy · Yannic Kilcher · Two Minute Papers | Deep learning and research, LLMs and neural networks, paper deep dives, summaries | Choose one paper connected to your project · record its question, evidence and limitations · check the original paper before repeating a claim |
| 3 · Strengthen the math and ML | Sebastian Raschka · sentdex · 3Blue1Brown · StatQuest | ML and LLM fundamentals, Python and ML, mathematical intuition, statistics | Use these when a concept keeps blocking progress · explain it in plain language, then test a tiny example · keep a notebook of assumptions and failure cases |
| 4 · Build an AI project | Google for Developers · Hugging Face · AssemblyAI · Nicholas Renotte | Developer tools, transformers and open-source, voice AI, hands-on builds | Define the input and the expected output first · build one small workflow you can demonstrate · test it with a new example beyond the tutorial |
| 5 · Learn what production requires | CampusX · Full Stack Deep Learning · Weights & Biases · NVIDIA Developer | Deployment, production AI and LLMOps, evaluations and observability, infrastructure | Add a simple evaluation before expanding · check what happens when an input is incomplete · record errors, response time and the help it needs |
WATCH THE ORIGINAL — 20 SECONDS
All twenty channels with their “best for” lines, and the five colour-coded groups along the bottom.
▶ 20 YouTube channels to learn AI — ordered by what you are ready for — BANI Academy on Facebook
Step two is a checker, and that is not an accident
Look at what every middle instruction actually asks for: recreate an example before changing the inputs, check the original paper before repeating a claim, test a tiny example, test it with a new example beyond the tutorial, check what happens when an input is incomplete.
Each one introduces something that can tell you that you are wrong, and crucially, something you did not produce yourself. A new example you have not seen. The paper rather than the summary of it. An input that is broken on purpose.
That is the same structure we wrote about when taking apart an agent architecture: the thing that produces the work cannot also be the thing that decides the work is good. Following a tutorial and feeling that it made sense is marking your own homework — you understood the explanation, which is not the same as being able to do it. The full version of that argument is in The Thing That Makes It Can Never Be The Thing That Grades It.
And the strongest single line on the whole card belongs here: check the original paper before repeating a claim. It is written for group two, but it applies to anybody who has ever quoted a number from a video. The summary is somebody else’s reading. If you are going to say it out loud, open the source.
Which group are you actually in?
| Where you are | The group | What to do first | What “done” looks like | The trap |
|---|---|---|---|---|
| You have never trained anything | 1 · Foundation | Recreate one worked example exactly, then change one input | A finished tutorial you can run again from scratch | Starting with the maths group because it feels more serious |
| A concept keeps blocking you | 3 · Maths and ML | Explain it in plain words, then test the smallest possible case | A one-paragraph explanation you wrote, and a tiny working example | Watching four explainers of the same concept |
| You want to ship something | 4 · Build | Define input and expected output before opening anything | One small workflow you can demonstrate to another person | Building before you decided what the output is |
| You keep quoting things you read | 2 · Research | Pick one paper tied to your project; record its limitations | Notes with question, evidence and limitations — in that order | Repeating a claim from a summary |
| Something works but nobody else uses it | 5 · Production | Add one evaluation before you add one feature | A number you can watch, and a record of what breaks | Scaling a thing you have never measured |
Four honest caveats
You need one group, not five. The card says it in its own subtitle — ordered by what you are ready for — and the post says choose the group that matches the work in front of you. Twenty channels is a reference. Four is a month. Trying all five groups at once is the most common way to finish none of them, which is the argument we made for narrowing in Pick One Per Stage And Learn It Properly.
Group five is mostly a team problem. Evaluations, observability, infrastructure and deployment matter enormously — once something you built is being used by people who are not you. Studying them before you have that is learning to read a dashboard for a system that does not exist yet. We made the same point about a skills chart in Ten AI Skills, But They Are Not Ten Separate Things.
Channels change and handles move. Any list of twenty names is a snapshot. Before committing a month, check the upload date on the most recent video — a recommendation stays in circulation long after a channel goes quiet, and nothing on a card tells you that.
“Something you can use” is doing a lot of work. The post’s closing section is unusually specific for a channel list: a working prototype, a clear explanation, or a useful evaluation all count. That is a generous definition and also a strict one — it rules out a folder of notes, a certificate, and four hundred saved videos. If what you have at the end of a month is none of those three, the plan did not run.
Questions people ask
How long should one group take? Long enough to produce the one artefact and no longer. If a month passes with nothing demonstrable, the problem is step one being too big, not the channel being wrong.
Can I skip the foundation group? If you already ship software, often yes — go straight to the build group and come back to the maths when something blocks you, which is literally what group three’s first instruction says.
What counts as “a new example beyond the tutorial”? Any input the video did not use. A different file, a longer one, one in your own language, one with a mistake in it. Those four cover most of what breaks.
What if I am learning this to make videos about it? Then the artefact is doubly valuable — it is the thing you learned and the thing nobody else has footage of. A real project with real numbers is the one angle a back catalogue cannot copy, which we argued in Nineteen Of These Channels Own One Word.
Keep the next lesson tied to the next missing thing
That phrase is the post’s own, and it is the best sentence in it. A learning plan is not a playlist in order; it is a loop where what you could not do last week chooses what you study this week.
Which is why the third instruction in every group is the one that matters. Write down the part you cannot explain. Record the assumption that failed. Note the error and what it needed. Those notes are not admin — they are the syllabus, written by the only person who knows where your gaps actually are.
Twenty channels give you five starting points. Fifteen instructions give you a method. Only one of those two halves will still be useful when the channel names have changed.
Learning a lot does not make anyone good. Doing a lot does.
If you want a deadline rather than five new subscriptions, we run a free challenge — there is no barrier to entry, only a commitment fee that is refunded when you finish. It exists because a deadline produces the artefact, and a playlist produces a playlist.
Which group matches the work in front of you right now? Tell us in the comments under the original reel.
More on building a channel for a global audience at mmoyoutube.com.
