The Captions Are Garbled Too. This Is A Picture Of A Watchlist.

“30 Videos To Master Agentic AI Before 2027.” A flowchart: a YouTube logo at the top, five branches, six video tiles down each branch. The caption tells you the order matters — “Start at the top left. Finish the column before you jump” — and ends with “Save this watchlist — it’s the whole path in one image.”

It is a lot of work to make, and the five column headings are genuinely well chosen. But before you save it, look closely at the fifth column, second tile from the bottom. The caption under it reads:

“Rornt Burrlé Agents, Ãoehtr abılq Εncyreenb — Developers”

That is not a video title. That is what happens when an image model is asked to draw words.

It is not just the thumbnails

Generated thumbnails would be a small thing. A thumbnail is decoration; you would still have the title and the channel, and you could go and search for them.

But the captions are generated too, and once you are looking for it you cannot stop seeing it. In the Production column, a channel is credited as “Andreıy Eopmerry”. In the MCP column, one reads “RAG & MCP Fundamentals — PurdueCoeCoemy”. Another in the same column manages two attributions in one line: “Intro to MCP Servers — Cousteau for Engineers — NetworkChuck”. The last tile in the Build column trails off into “CrewAaIlt AI Agets —ıt Cheab”.

Meanwhile the thumbnails behind them carry their own layer of the same thing: “RAKT ÄOOHO FUNDAMENTALS”, “L Bulding More Effectve efRocsaess fnicrntus”, “Apofunny Shs”.

So this is not a designed graphic with generated pictures in it. The whole card is an AI-generated image of a watchlist — layout, tiles, captions and all. Some captions came out as clean English. Several did not. Both outcomes are the same process.

Which is why the clean ones are not evidence either

This is the part that matters, and it is easy to miss. If the broken captions were the only problem you could simply skip them and use the twenty-five that read properly.

We tried to verify the entries. The one we could pin down completely is in the first column, fourth tile, and it reads perfectly cleanly:

“AI Agents Fundamentals: A 2-Hour Course — Theo Huang”

The video that exists is “AI Agents Fundamentals in 21 Minutes”, by Tina Huang. Not two hours — twenty-one minutes. Not Theo — Tina. She does have longer agent videos, including one at forty-four minutes, but nothing matching the title and duration on the card.

Nothing about that tile looks wrong. The letters are crisp, the name is a plausible name, the format is right. It is wrong in exactly the way a text-shaped guess is wrong, and it passes the only test a reader can apply by eye.

What we checked, and what we found

We did not attempt all thirty. Here are the ones we looked at closely, including the ones that hold up.

Five captions with visible corruption, one with two channel names, one with none, one that does not parse, and one that reads beautifully and is verifiably wrong. That last category is the one that makes the card unusable rather than merely untidy, because it means legibility carries no information about accuracy.

Diagram: an image model draws text as texture

None of this is a failure of effort. It is what the tool does. An image model is trained to produce images that look like images of a thing; letters are shapes that appear in images, and it renders them the way it renders brickwork — convincingly at a glance, arbitrary up close. It has no concept that a caption must resolve to something that exists.

The five column headings are the actual curriculum

Here is the good news, and it is better news than it sounds. The thing the card is actually arguing survives the image entirely. The caption spells it out:

Foundations — how agents actually work. Build AI Agents — make one yourself. Multi-Agent Systems — get them working together. MCP & Tool Connectivity — plug them into real tools. Production, Evals & Security — ship it without breaking things. Start at the top left. Finish the column before you jump.

That is a sound order, and the ordering claim is the card’s real contribution — not the thirty tiles. It puts building one yourself second, before multi-agent, which is right: most people who skip it end up orchestrating something they could not have written. It puts evals and security last but inside the path, rather than as an optional appendix, which is where most self-taught routes leave them. And “finish the column before you jump” is the one instruction that actually changes behaviour, because the failure mode in this subject is sampling five introductions and building nothing.

You do not need anybody’s thirty links to follow it. Five rungs, and at each one you search for yourself.

Rung four is worth a note, since it is the one most people reach first and understand least. We have written up how an MCP server actually reaches your assistant and when it loads, and what happened when a card got the endpoint and the scopes mixed up — which is exactly the mistake rung four exists to stop you making.

Four honest caveats

1. We did not check all thirty entries. We examined fourteen closely and verified one completely. Several of the clean-looking ones — Microsoft Developer, Jeff Su, IBM Technology, NetworkChuck, Anthropic — are consistent with real channels and real subjects, and may well be exactly right.
2. “Unverified” is not “invented”. Two entries we searched for did not turn up, which may mean our search was poor, the video is recent, or the title is slightly off. We are reporting what we could and could not confirm.
3. Letter-soup is a tool failure, not a character failure. The likely story is simple and blameless: somebody generated a good-looking graphic, most of the captions came out readable, and it shipped. The point of this article is what the tool does to a reference document, not what the author intended.
4. The curriculum is not in question. Our argument is entirely with the thirty tiles. The five rungs and the instruction to finish a column before jumping are sound, and we would keep both.

The one-line version

The thumbnails are generated, the captions are generated, and the entry we could fully verify reads cleanly and names the wrong person and the wrong duration. You cannot sort the real tiles from the invented ones by looking, because legibility was never the test.

So take the five column headings, which cost nothing and survive everything, and search for the videos yourself. And if you make cards like this: generate the mood, typeset the facts. The moment a graphic contains something a reader is supposed to go and find, it needs a font, not a diffusion model.