Teach us four things.
On video, in your own way.
Everyone teaches the transformer on a whiteboard and shows us their best session with AI agents. Then you pick any two of three more topics. This is how we learn how clearly you think, how deep your foundations go, and how well you can teach something new.
Have you applied yet? Fill in the application form first, then come back here.
Don't know some of these topics yet? That is completely fine. It is completely fine if you have never studied some of these topics before. We want to see how fast you can learn something new, grasp it, and teach it from first principles.
What we are really judging
The reading material we share is only a reference. Read other blogs, watch other videos, go to the papers, run the code. What we care about is what you build on top of it:
- The story. Where you start, how you build up, and why each idea follows from the last.
- Your examples and metaphors. The comparisons that make a hard idea click for a beginner.
- First principles. Why things are the way they are, not just what they are. Show the mathematics where it matters.
- Joy. Be engaged and curious. We want to feel that you love this.
- Proactiveness. If you run your own experiment and show us the results, that is brilliant. “I learned this, tried it, and here is what I found” is exactly what we look for.
Please do not read from slides. If you put a finished presentation on screen and read it out, we will not move forward. We want to watch you think and draw, as if you were teaching a room of people who have never seen this before.
Build your own way in
The best lessons start from the problem that made the idea necessary. A few ways you could begin:
Start with the history. How did we model sequences before? Walk through RNNs and LSTMs, why they were slow and forgot long-range context, how attention first appeared in sequence-to-sequence translation, and why “Attention Is All You Need” threw the recurrence away entirely.
Start with why this is hard. Why is reinforcement learning with a single reward at the end so difficult? Why did earlier attempts at reasoning with RL struggle (sparse rewards, credit assignment, reward hacking)? Once that is clear, the aha moment becomes genuinely surprising.
Start by drawing the GPU. Sketch its parts: streaming multiprocessors, tensor cores, registers and shared memory, L2 cache and HBM. Show where the data lives and how it moves. Then derive the two limits, compute and bandwidth, from that picture.
Start with the engine. A language model is powerful, but on its own it can only predict text. Then build up everything around it: the loop, the tools, the context, memory and permissions, until you have an agent that can actually do work.
These are only examples. Choose your own starting point, stitch ideas together, and go as deep as you can.
How to record a whiteboard lesson
Your camera must be on, in every video. We need to see your face while you teach. If you record your screen, keep a webcam bubble visible the whole time. If you film a board or paper, set up the camera so both you and the board are in the frame. Videos where we cannot see you will not be reviewed.
You do not need a studio or a real whiteboard. Any of these is perfectly fine:
Paper or a board + your phone
Write on a whiteboard, paper or a notebook, and film it with your phone on a stand or a stack of books. Frame it so we see both you and the board.
A digital board
Draw on Miro, Excalidraw or tldraw, and record with Loom, OBS, or a Zoom or Google Meet recording with your webcam on.
A tablet
Use an iPad or Android tablet with any notes app. Record through Zoom, Meet or Loom so your camera shows alongside the drawing.
Tips
- Start from a blank page. Build the drawing as you talk. A pre-drawn diagram you only point at is close to reading slides.
- Sound matters most. Find a quiet room and keep the microphone close. Your phone's mic is fine.
- Write big and in landscape. If we cannot read it, we cannot follow you.
- Camera on, face visible, in every video. Good light on your face makes a big difference: face a window or a lamp.
- Do a 30-second test first to check sound, light and framing.
- One take is fine. No editing needed. Small mistakes, and how you recover from them, are part of teaching.
- Length: roughly 10 to 25 minutes per lesson is a good range. Depth matters more than length.
The questions
Questions 1 and 5 are required for everyone. Then choose any two of questions 2, 3 and 4. So you send us four videos in total.
The Transformer, from scratch
Explain the transformer architecture from first principles, as if your audience has never heard of it. Draw every block yourself and show how the blocks connect: tokens, embeddings, positions, attention, the MLP, residual connections, normalisation, and how the next token is predicted. The more depth, the better.
What this tells us: Foundational understanding and depth.
Primer and reading list →What is a harness? Pi vs Hermes
Teach what an AI agent harness is and why it matters. Then explain what the open-source Pi harness does, what Hermes Agent does, and how the two differ. Start from zero and build up. Applications and examples are welcome.
What this tells us: Whether you can pick up new material quickly and teach it well.
Primer and reading list →DeepSeek's "aha moment"
Explain the DeepSeek-R1 "aha moment" in as much depth as you can. Go as far back as you like and build a story around it: what reinforcement learning is, what GRPO does, how rewards work, what actually happened during training, and what it means. Show the mathematics.
What this tells us: Storytelling and mathematical strength.
Primer and reading list →The GPU roofline plot
Explain the roofline model from first principles: the two limits of a GPU (peak compute in FLOP/s and memory bandwidth in bytes/s), what arithmetic intensity is, and why it decides how much of the GPU you actually use. Work through real examples.
What this tells us: First-principles thinking about hardware and numbers.
Primer and reading list →Your best session with AI agents
Screen-record and walk us through the best session you have had with AI agents (Claude Code, Codex, Cursor or anything else). Show the actual session: what you set out to do, how you split the work, how many agents or sessions you ran at once, what you built and what results you got. Keep your camera on so we can see you as you talk.
What this tells us: Whether you understand that one person with agents can now do 20 times more work.
Primer and reading list →How to submit
- Create one Google Drive folder with your four videos:
Q1_Transformer,Q5_Agentsand the two you chose, named from this list:Q1_TransformerQ2_HarnessQ3_AhaMomentQ4_RooflineQ5_Agents - Set the folder's sharing to “Anyone with the link” → Viewer. Check it by opening the link in a private or incognito window. If we cannot open it, we cannot review it.
- Email rajatdandekar@vizuara.com with:
Subject:
Vizuara Task Submission | Your Full Name | Role(s)Body: your full name, the email you applied with, phone, college, the role(s) you applied for, and the Google Drive folder link.
- Send it by Friday, 20 November 2026, 11:59 pm IST. Send only one email. If you need to fix something, reply in the same thread.
Questions about the task? Email rajatdandekar@vizuara.com. We read every message.