TL;DR
- Access to information is not the same as learning. A learner still needs to know what to learn, in what order, what to skip and where to spend more time.
- When you learn with ChatGPT, you control the questions. A beginner does not know what they do not know, so the important question often never gets asked.
- Reading a good explanation feels like understanding. A slightly different scenario is what shows whether the mental model holds.
- Pravaha sequences the path for each learner and then turns the tables: an AI coach asks the questions, and the path responds to what it finds.
I build Pravaha, an AI coach for engineering teams. When I show it to people, one question comes up more than any other:
“AI chatbots like ChatGPT can answer almost any question. Why do I need Pravaha for learning? I can just use ChatGPT directly.”
Fair question. In fact, that is pretty much where I started as well.
Answers were never the scarce part of learning
A while ago, I ran a 12-week bootcamp for eight freshers we had hired. Whenever someone had a doubt, the answer was usually somewhere on the internet. Documentation, Stack Overflow, blog posts, YouTube videos. Knowledge was never really scarce.
ChatGPT made that knowledge dramatically easier to reach. Instead of searching through ten links, you ask a question and get a reasonably good explanation immediately.
But the fundamental problem with learning did not disappear.
A learner still needs to know what to learn, in what order, what can be skipped, what deserves more time, and how one concept connects to another. That is surprisingly difficult when the subject is new to you.
How many times have you started reading an article about one topic, clicked a link to understand something mentioned inside it, opened another tab, searched something else, and 20 minutes later found yourself in a Reddit thread that has very little to do with what you set out to learn?
That is the obvious problem. There is another, much more important one.
Whoever asks the questions decides what gets learned
When you use ChatGPT for learning, you control the questions.
That sounds like a strength. For a beginner, it is also a weakness, because you do not know what you do not know.
Say I am learning vector databases. I may happily ask about embeddings, cosine similarity and Qdrant. I may not even realise that my understanding of tokens, context windows, chunking, or how an LLM actually consumes the retrieved text, is weak.
The two exchanges below cover the same learner and the same topic. Watch who is asking the questions, and whether chunking ever comes up.
The learner asks GAP STAYS HIDDEN
A clear, correct explanation of embeddings as vectors that place similar meanings close together.
Another clear, correct explanation, with a worked example.
Working code.
Why this falls short. Every answer is right, and the learner leaves feeling they understand retrieval. Chunking and context windows never come up, because the learner did not know they were questions.
ChatGPT will very patiently answer every question I ask. It will not necessarily tell me about the important question I never thought of asking.
A coach asks GAP SURFACES
Your search returns the right document, but the answer still gets a number wrong from the middle of a pricing table in it. What would you check first?
The right document did come back. What did the model actually receive from it?
Why this is good. The learner never asked about chunking. A scenario made the gap visible in two turns, and now the path knows where to send them next.
Feeling you understand is not the same as understanding
This is where structured learning still matters. Good learning is not a collection of explanations. It needs sequencing, deliberate practice, feedback and assessment.
What should I learn first? What depends on it? Where should I spend more time? What can I safely skip because I already understand it?
And most importantly: do I actually understand the concept, or do I only feel like I understand it after reading a good explanation?
That last one is the dangerous one. You read an explanation from ChatGPT and think, “Yes, makes complete sense.” Then someone gives you a slightly different scenario, and the mental model falls apart. The learner in the first exchange above would have said they understood retrieval. The pricing table is what showed they did not.
Pravaha turns the questions around
That is the problem we are trying to solve with Pravaha.
Pravaha builds a structured learning path for every learner, based on what they already know and what they are trying to achieve. I have written separately about why a learning path is only personal if it knows what to skip.
Within that path, learning is multimodal. Videos, interactive content, curated articles and simulations, depending on what works best for that particular concept.
The part I find more interesting happens after that. Instead of asking “Did you finish the lesson?”, an AI coach starts asking questions. Not just recall questions. Scenarios. Trade-offs. Debugging situations. “What would you do if…?” The pricing table question earlier is one of these.
The learner is no longer the only one deciding what gets asked. The system probes the learner.
That changes the experience quite a bit. The goal is not to check whether you remember the definition of a concept. It is to find out whether you have built the right mental model. If that model is weak, the learning path responds accordingly.
When answers are abundant, structure matters more
This is why I do not think ChatGPT and structured learning are really competing with each other. ChatGPT is an incredibly powerful way to access knowledge. But giving someone unlimited access to knowledge does not create a learning journey.
You still need structure. You still need sequencing. You still need practice. And you still need someone, or something, to occasionally turn the tables and ask you the questions.
AI has made answers abundant. That may make good learning structure more important, not less.