Why Video Courses Don't Stick for DSA and Generative AI

By Pooja Goenka · 2026-09-09

Forty hours of DSA videos, three notebooks of notes, and then a two pointer question in a screening round goes badly. The usual explanation is that you didn't practice enough. Often true. It also lets the format off the hook, and the format deserves some of the blame.

I want to be specific about why, because "video bad, interactive good" is a marketing line and I would rather give you something you can check.

The gap between following and doing

A good video explanation is easy to follow. That is what makes it good, and it is also the trap. Psychologists call it the fluency illusion: when material is presented smoothly, we read our own comprehension off how smooth it felt and overrate how much we have actually learned. Rozenblit and Keil found the same effect outside the classroom with what they named the illusion of explanatory depth. People are confident they understand how a zipper works right up until they are asked to explain it step by step, and then the confidence falls apart.

A recorded lesson is engineered for smoothness. The instructor already knows where the hard part is, has rehearsed the sentence that gets past it, and has cut the twenty minutes where they were confused. You get the clean path. You never get the experience of being lost and finding your way out, and that experience is the part that transfers.

Why video is a bad container for an algorithm

An algorithm is a state machine. It has a current state, a rule for the next step, and a stopping condition. A video is a timeline with a position and a play rate. Those two things do not map onto each other.

You feel it the moment you get stuck. You want to go back one operation, and what you get is a scrub bar that jumps three seconds and drops you mid sentence. You want to hold the frame where the left pointer moved and stare at it for a minute, and the narration keeps going. You want to change the input array to see whether your theory about the loop invariant survives, and you can't, because it is a video of an array.

Cognitive load researchers have a name for the underlying problem too: the transient information effect. Narration and animation vanish as they play, so your working memory has to hold every earlier step while processing the current one. A static diagram you can look at twice costs you nothing. This is one of the better replicated results in instructional design, and it points the same way every time. For anything with several interacting parts, people learn more from material that stays on screen than from material that plays.

You can search a page, you can't search a video

Two months later you are revising for interviews and you half remember something about why the slow pointer starts at the head. In a written, visual lesson that is a fifteen second lookup. In a video course it is a rewatch, or a scrub hunt through a forty minute recording where the answer sits somewhere near minute twenty six.

Video is also sized wrong, and not by accident. Runtimes get chosen for retention curves and ad breaks. Some ideas need ninety seconds. Some need an hour of sitting quietly with one diagram. Very few of them need twenty two minutes.

What this looks like in DSA

Take memoization. The standard line is that you store intermediate results so you don't recompute them. You can repeat that in an interview and get nothing for it, because it describes the fix rather than the problem.

Now draw the call tree for a naive fib(6) and expand it yourself. fib(4) appears twice. fib(3) appears three times. fib(2) appears five times. fib(1) appears eight. You count the duplicates on your own screen, and memoization stops being a trick you memorised for the round. You worked it out before anyone told you.

A video shows you that tree for eight seconds on the way to the code. The tree was the lesson.

The same goes for almost every DSA topic that people find slippery. Why quicksort's partition step is correct. Why a heap is an array. Why BFS gives you the shortest path on an unweighted graph and quietly stops doing that the moment you add weights. Every one of those is a thing you understand by watching state change under your own hand, one step at a time, on inputs you chose.

What this looks like in generative AI

The GenAI version is worse, because the vocabulary hides so much.

"An embedding places semantically similar text close together in vector space." Everyone can say that sentence. Very few people who have only watched it explained can tell you what changes when you swap the model, or why "bank" sitting near "river" lands somewhere different from "bank" sitting near "loan", or what "close" is measuring in the first place.

Attention is the same story. A heatmap over a sentence on a slide is a picture of the answer. What you need is to add a token and watch the weights redistribute, because that is the moment it stops being a diagram and becomes something you can reason about.

And the agent loop. Model, tool call, result, model again, until a stop condition fires. Four boxes on a slide, thirty seconds of narration, apparently obvious. Then you step through it with a real tool that returns a real failure, and you finally see why anyone bothered building LangGraph.

Where video is still the better format

I am not going to pretend video is useless. Our team learns from it constantly.

Video wins when the thing you are learning is a workflow rather than a concept. Watching an experienced engineer set up a project, move around their editor and drive a debugger teaches you a dozen small things nobody ever writes down. It wins on the messy parts too. A live debugging session where the first three theories are wrong is genuinely valuable, and it is exactly the footage a polished course cuts.

And some days you want a human voice and no decisions. That is a real reason. Starting badly beats not starting.

A study loop that works for both

Whatever course you end up using, the shape that works is roughly this:

1. Watch the thing move before you read about it, so you have a picture of the state changing.
2. Predict the next step before you reveal it. This is where the learning happens and it is the step everybody skips.
3. Write the code yourself from a blank file, rather than editing someone else's.
4. Get it checked immediately, while you still remember what you were thinking.
5. Come back in a week and rebuild it from nothing.

Steps two and five are retrieval practice, which is the best supported finding in this whole area. Roediger and Karpicke's work on the testing effect showed that people who spend study time trying to recall material beat people who spend the same time rereading it, and the gap grows the further you get from the study session. Rereading feels better while you do it. Recalling works better afterwards.

How the LogicWiz courses are built around this

This is the reasoning behind both of our tracks, and I will say plainly that we built them this way because the alternative had not worked for us.

The DSA track has 99 curated problems. Concepts arrive as visual simulations and short stories instead of a lecture, then you write the solution in the in browser playground and it is auto judged against the test cases right away. There is XP and a leaderboard, which is a small thing that does measurably get people back the next day. When you are stuck, the AI chat gives you a nudge rather than the answer. That was a deliberate choice and an irritating one to build.

The Generative AI course has 43 lessons across 10 modules, running from Python basics through retrieval augmented generation, LangGraph orchestration, multi agent systems, and what is actually happening inside a transformer. Every lesson pairs written material with an in browser lab, and there is an AI tutor you can interrupt in the middle of a lesson. The first four lessons need no account and no card, which is on purpose. An hour is enough to work out whether this format suits you.

Both tracks are free while we are pre launch.

None of that makes us better than a good video course at everything. What it does well is the one thing video does worst: letting you sit inside a mechanism and poke at it until it makes sense.

Common questions

Is visual learning actually better, or is this the learning styles myth?

Two different claims, worth keeping apart. The learning styles idea, that you are a "visual learner" and your friend is an "auditory learner" and you each need different material, has not survived testing. That is not the argument here. The argument is about the subject. Algorithms and model internals are things with states and flows, so material that shows state changing fits them better than material that describes it, and that holds for everyone, including people who are certain they learn best by listening.

Can I learn DSA without watching any videos?

Yes, and most people who get good at it end up spending the bulk of their time in a problem editor rather than a video player. Videos are useful for a first pass at an unfamiliar topic. The reps are what carry you through a screening round.

How long does it take to become interview ready?

That depends on where you start and how consistent you are, and anyone giving you a number in weeks is guessing. A more honest measure is problems solved from scratch with no hints. When you can read an unseen medium problem and know within two minutes which shape it is, you are close.

Do I need Python before starting the generative AI course?

No. It begins at Python basics and builds up to agentic systems across the 43 lessons. If you already write Python, skip ahead. Nothing later is locked behind lessons you passed over.

Is it genuinely free?

The first four GenAI lessons need no signup and no card. Past that you make an account, and both tracks are free while we are pre launch. Current terms are on the pricing page.

What should I do if I have already bought a video course?

Finish it, but change how you use it. Watch a segment once at normal speed, close it, and rebuild the idea in a code editor or on paper without looking. If you can't, go back to the exact minute you needed. You will get far more out of the same course, and you will find out quickly which parts you only thought you understood.

Start the DSA track or the Generative AI course. The first four GenAI lessons need no signup, so you can tell within an hour whether this way of learning suits you.