Learning AI with AI

How I used different LLMs to build a learning path for AI engineering

A futuristic isometric illustration of a flow from a large pile of AI resources, through a series of data-processing steps, leading to a neatly organized course curriculum.

(Note: If you're here to learn AI, see the compiled AI course. The article below describes why and how I assembled it.)

I’ve long wanted to train myself into an AI engineer with a strong machine learning background. The timing has finally worked out, so I’m diving in! Over two years, I’ve amassed a collection of about 200 AI-related resources—blog posts, research papers, tutorials, GitHub repos—none of which I had found time to properly explore. It was time to turn this hoard of links into a coherent learning path.

And that’s where the real fun began: ChatGPT plus some custom scripting to wrangle it all into shape.

The workflow I used

                  [ 200+ ML/AI-related links ]
                               |
                               v
   +-----------------------------------------------------+
   |    Script & GPT summaries (Node.js/TS, ChatGPT)     |
   +---------------------------+-------------------------+
                               |
                               v
   +-----------------------------------------------------+
   |      Curriculum outline (ChatGPT 4, Llama-3.3)      |
   +---------------------------+-------------------------+
                               |
                               v
   +-----------------------------------------------------+
   |       Final course layout (GSheets + article)       |
   +---------------------------+-------------------------+
                               |
                               v
                        Start Learning!

1. Building my master list

I’d been bookmarking articles, saving links in random places, and occasionally copying them into a text file. This hodgepodge of resources ballooned to around 200 URLs. I’ve published the raw link list here. They cover everything from fundamentals of linear algebra to advanced tips on prompting LLMs.

2. Flattening the chaos into a spreadsheet

I wanted to run a script over all my sources. For this, I had to flatten my 200-URL chaos. I fed the entire link list into ChatGPT o1, asking it to convert the text file into a neat spreadsheet. If you’re curious, the resulting spreadsheet is here.

3. Summarizing each resource

Next, I wanted at-a-glance summaries of all these articles. I wrote a small Node.js/TypeScript tool to loop through all the URLs, fetch the content, and then call an LLM to generate short summaries.

The script did its thing—see the results in this new sheet.

4. Using AI to make a curriculum

Armed with 200 summarized resources, I now wanted a “course outline” or “learning path”. So I asked ChatGPT 4o (because o1 didn't accept an Excel file) to find patterns in the spreadsheet and propose a curriculum. I gave it a previous outline I’d built in an audio chat (ChatGPT doesn't support sharing audio chats, so no link), and built a new, more complete course outline.

Note: As always with LLMs, I didn't trust the output blindly. After each pass, I’d do a quick manual review, tweak the prompts, or shuffle the structure. The result was a pretty good draft of what to learn first, second, third, and so on.

5. Categorizing everything

Then came a categorization step. I wanted each resource tagged with one of my pre-defined topics, like “ML intro”, “Deep nets”, “LLMs”, “MLOps”, etc. I made my script more generic so it could prompt GPT for these labels automatically.

Note: For this step, I intended to use DeepSeek v3 but accidentally ended up with meta-llama/Llama-3.3-70B-Instruct-Turbo-Free. The result was surprisingly OK, so I didn’t bother re-running it. You can see the categorization results here.

6. Finalizing the learning list

Finally, I combined the outline from ChatGPT with the link-level info from Llama-3.3 to produce my final list. I sorted it in Google Sheets by the numbers the LLM gave me, and made the final layout in Google Sheets.

I've also asked o1 to give me a formula to convert each row to a piece of Markdown, which (after some manual editing) resulted in this AI course page.

The LLMs’ combined guesses suggested 7,409 minutes of study time. That’s about 20 very full workdays (six hours a day) or roughly four weeks. In practical terms, I’ll probably double that, because “reading time” can expand if you want to do sample projects, plus, some of the estimates were off, like for item #1, which is a Coursera course of ~80 hours, not 3. I might skip such long materials, though.

Where I go from here

The next step: actual study. My final learning list is locked and loaded. The plan is to work through each topic, probably in fits and starts, but at least I know where I’m headed.

I’ll keep posting because learning in public motivates me. I hope that sharing my progress will help others, too.

If you’re also trying to assemble your own “Course on AI”—or on anything else, really—I’d highly recommend using AI. Make sure to drop a link about it on Bluesky: https://bsky.app/profile/veszelovszki.com

Happy learning!