LoRA, low-rank adaptation, is arguably the most popular parameter-efficient fine-tuning method for LLMs. But how does it actually work? Check out the video to learn LoRA and friends (LoRA+, QLoRA, VeRA, and DoRA)! youtu.be/U80tjcThl9Q

What? A 37-minute video? Ain't nobody got time for that! I hear you. Check out the 10-second version of JEPA.

I didn't know what JEPA is, and at this point I am too afraid to ask ...😬 so I made a video tracing the last 30+ years of self-supervised learning, covering ideas from contrastive learning, distillation, masked modeling, JEPA, and world models. youtu.be/gVEr2cnDE_8

The 60-Year Hunt for AI's Most Important Function I was trying to understand how SwiGLU works, but I couldn’t find an explanation that clicked for me. So I made this video to explain it from first principles. Check it out: youtu.be/JRaPNrpsQ9s

**Modern Transformer - Complete Guide** Interested in learning the recent advances in transformers? After 14 videos, I've finally completed this series! 🥳🥳🥳 Check out the course here: www.youtube.com/playlist?lis...

The Most Underrated Layer Inside Every AI Model Virtually every AI model has normalization layers. BUT, what makes them so essential? 🤔 New video on learning the role of normalization in stabilizing training and alternatives like DyT and Derf. youtu.be/JHl_gwVoh-k

How is DeepSeek V4 so INSANELY cheap? 🤔 Compared to a GQA baseline, it's new *compressed attention* mechanism (CSA and HCA) slashes the KV cache memory cost by 98% 🤯 at a 1M-token context! Here’s how: youtu.be/q8holiIirgo

How do we make attention actually capture context? Exclusive Self Attention (XSA) is an interesting variant that improves attention with minimal cost in speed & memory. Check out the video here: youtu.be/2eZKT4H9_iQ

**Modern Transformer architecture explained** I compiled a list of videos on the Transformer architecture into a short "YouTube course". www.youtube.com/playlist?lis... Hopefully, this would be helpful for beginners in the community. Happy learning! 😎

Finally got some time to read the DeepSeek Engram paper! Idea: Replace repeated reconstruction with direct lookup of common knowledge. It’s so intuitive that it feels strange this wasn’t part of the design from the start. Video summary here: youtu.be/87Q8nf1XHKA

Introducing *Edit-by-Track*, a framework that enables precise video motion editing via 3D point tracks. Our method supports a wide range of motion editing applications, including object removal, shape deformation, dynamic view synthesis, and many more! Video explainer: youtu.be/aAj6PIgx20o

Wondering how DeepSeek v3.2 rivals SOTA models (e.g., GPT5/Gemini 3 pro) while being ~30x cheaper? 🤔 Let's learn how the base model works! We'll focus on attention, the need for KV caching, and key ideas for improving attention (MQA/GQA/MLA/DSA). youtu.be/Y-o545eYjXM

Sharing the slides for a talk on faculty job search Hope it's helpful to people exploring and preparing for the process. Feedback is welcome! www.dropbox.com/scl/fi/p7xdt...

Here's a simple fix: Instead of just presenting information, tell a story. A story about how you encounter a problem, solve it, introduce a new conflict, and resolve it, and so on. Keep creating curiosity loops to engage/re-hook your audience.

How to organize your talk? I used to present like this, thinking that I was being "academic", "organized", and "professional". BUT, from the audience's viewpoints, this sucks. 😱 Look how far they need to hold a long-term context to just make sense of what you're saying!

In an era of billion-parameter models everywhere, it's incredibly refreshing to see how a fundamental question can be formulated and solved with simple, beautiful math. - How should we orient a solar panel ☀️🔋? - Zero AI! If you enjoy math, you'll love this! Video: www.youtube.com/watch?v=ZKzL...

*Slides without slide titles* When I first tried presenting WITHOUT slide titles, everything flowed so much better! (totally validated ... by me)! Give it a shot! Once you try it, you’ll never want to go back.

*Empty initial slides* What’s a better starting point than that default slide layout? A completely blank slide. It helps you explore the design space and focus on delivering a clear, compelling story.

*Bullet points* The second thing the layout prompts you to do? ("Click to add text"). Start a bullet list. Among so many creative forms of presenting your ideas, it nudges you toward the most boring one: a list. 🔢

*Slide title* The first thing this layout does is to ask you to add a slide title. Seems reasonable, right? visuals, this encourages you to 1) lead your presentation with text instead of visuals and 2) cram in many titles in a talk, making it harder to maintain a narrative flow.

Why is the "Title and Content" slide layout BAD? Most people prepare their presentation from this default layout. I used it for years without questioning it. BUT, this essentially guides you toward developing poor presentation. Why? 🤔

Why More Researchers Should be Content Creators Just trying something new! I recorded one of my recent talks, sharing what I learned from starting as a small content creator. youtu.be/0W_7tJtGcMI We all benefit when there are more content creators!

Policy gradient methods rock! These are the core techniques for making your transformer "chat" and "reason", a robot that manipulates objects, and a drone that maneuvers in a complex environment. BUT, how do we learn all the developments in the past 30+ years?

Awesome! 🤩 So glad to hear the authors enjoyed the video, totally made my day!

We had a blast at CVPR2025! There was so much to learn! I am particularly excited to meet many new friends and reconnect with old ones. I feel energized. Already looking forward to the next one!

Kullback–Leibler (KL) divergence is a cornerstone of machine learning. We use it everywhere, from training classifiers and distilling knowledge from models, to learning generative models and aligning LLMs. BUT, what does it mean, and how do we (actually) compute it? Video: youtu.be/tXE23653JrU

How LLMs Learn to Reason with Reinforcement Learning Full video: www.youtube.com/watch?v=mg-i...

RL is so back! Reinforcement learning is a key driver in aligning LLMs and enhancing their reasoning capabilities. BUT, it’s a tricky topic to wrap your head around (at least for myself 😵‍💫). So, I put up a video breaking down the basics in a way that clicked for me. I hope it helps you, too!

🥺

IVE leverages VLMs to • extract semantic scene graphs, • imagine novel scenes, • predict their physical plausibility, and • generate executable sequences. IVE is a memory-guided agentic exploration framework that operates fully automatically, enabling more diverse and meaningful exploration.