"Interactive Tools That Actually Help You Learn Transformers and Deep Learning"

Most people (including me) don’t learn AI by reading 50-page PDFs front to back.

We learn from:

• YouTube videos
• blog posts
• random tweets
• half-remembered formulas from a course

…and then we try to glue all of that together in our heads.

I’ve been building a platform to make this process less chaotic and more interactive: instead of passively consuming content, you explore concept maps, highlight research papers, and break down formulas step-by-step in a notebook. Breaking down and aggregating all of the information into one place.

In this post I’ll walk through three of the core workflows, using real screenshots from the app:

1. Learning from videos with concept maps
2. Discovering and navigating related research papers
3. Turning dense formulas into understandable notes ## 1. From “I watched a video” to “I understand the ideas”

Here’s a snapshot of the “watch + explore” experience:

![Video + concept map workspace]

On the left is a (great) YouTube lecture on transformers and neural networks. On its own, it’s easy to watch, nod along, and forget everything two days later.

On the right and below:

• Concept map — a visual graph of the major ideas (Fourier transform, CNNs, embeddings, etc.), so learners can see how topics connect.
• AI-generated explanations — clickable terms in the transcript that open definitions, examples, and follow-up questions.
• Question prompts — learners can ask targeted questions about what they’re seeing, not just generic “explain transformers” prompts.

The goal is to turn video watching from a passive activity into a structured learning session:

• Watch a segment
• Click the highlighted term
• See where it sits in the bigger picture
• Ask follow-up questions directly about that concept

2. Exploring papers without drowning in PDFs

Once you move past tutorials, you start touching research papers. That’s where a lot of learners bounce:

• The abstract is dense
• Terminology jumps levels quickly
• It’s hard to know which papers are foundational vs. niche

Here’s how the platform handles paper discovery:

![Survey paper with related results]

On the left: an open access survey paper

On the right: a ranked list of related papers, with:

• highlighted matching sections
• quick relevance scores
• one-click access to the full abstract / PDF

You can start from a high-level survey like:

“Review of deep learning: concepts, CNN architectures, challenges, applications, future directions”

…and immediately see:

• which works are extending ideas
• which ones are applying similar concepts in different domains
• which ones are more math/architecture-heavy vs. application-heavy

Instead of opening 20 tabs and skimming blindly, you get guided exposure to the “neighborhood” of that paper.

---

3. Turning attention formulas into something you can reason about

Even when you’ve found the right papers, the math can feel like this:

“I’ve seen this formula before, but I still don’t really get it.”

The notebook is built around that feeling.

Here’s a screenshot of a note that breaks down self-attention:

![Notebook with formula breakdown]

You can:

• paste or type formulas using LaTeX/KaTeX
• highlight specific pieces (like (QK^\top / \sqrt{d_k}))
• ask the AI to explain only that part in context
• structure your notes with headings, bullet points, and callouts
• save snippets from multiple papers, video keyword analysis, multiple follow up questions all into one place

For example, instead of just throwing the scaled dot-product attention formula at you, you will see a formula you can read and understand:

Try it out

If you’re learning AI / ML or working with LLMs and want to explore these workflows yourself:

I-O-A-I — broad AI learning (concept maps, simulations, research exploration)
👉 https://i-o-a-i.com

L-L-M — focused on transformers, attention, and large language model internals
👉 https://l-l-m.org

I’d love feedback from devs and researchers:

What part of learning AI has been hardest for you?

Which tools (if any) have actually helped?

What would make this kind of platform more useful for you?

Feel free to comment here or reach out — I’m actively iterating on this.

\text{Attention}(Q, K, V) = \text{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V