The craft of pushing AI forward, from reading papers to building world models.
It's 2026 when I'm writing this and AI is moving faster than anything we've seen before. It feels like every other week we have a new model, new architecture, new breakthroughs, or some incredible accomplishment. It's exciting but incredibly defeating since it feels like it would take a century to catch up and learn what's happening.
This is why I'm choosing to cover AI research. I want to discuss what AI research is, how to go about it, where to start, and more. Unlike machine learning where we are applying known techniques to business problems, AI research is about producing new knowledge. I hope to cover new architectures, training methods, and results from the most important papers of our century with a heavy focus on robotics.
I've made this list in order so topics build on top of each other, but AI research is so vast that you're better off hopping around different topics that you're interested in. As always, feel free to click on links to learn more and recursively build your understanding.
Researching is a skill. I'll walk you through how I read white papers, what's worked for me, and how to reproduce results from papers that you want to learn from. This section will give you a good understanding of the research process and you should even be able to extend papers, design your own experiments, and even try publishing on your own.
The three-pass method: skim, study, then rebuild it in your head.
If you can't rebuild it, you don't understand it yet.
Ablations, baselines, and seeds to prove your idea is the reason.
What a score really measures, and when a leaderboard lies.
This section covers modern AI architectures that fundamentally changed society and the reason why you're probably so keen on learning AI in the first place. These topics were directly responsible for creating ChatGPT and other modern LLMs. We'll cover the white papers that produced these findings and how to apply them.
The idea behind modern AI: every token attends to every other.
Chop an image into patches and treat them like words.
Models that teach themselves from unlabeled data.
Generates by learning to undo noise, one step at a time.
A straighter road from noise to data, in far fewer steps.
Reinforcement learning is learning by trial and error: an agent acts, the world responds, and reward is the only teacher. The split that matters most here is model-free versus model-based: whether the agent maps the world first or learns purely from outcomes. That question leads directly to imagination training, where a policy trains inside a learned simulator instead of the real world.
The core loop: an agent acts, the reward arrives, the policy updates.
React from experience, or learn a world you can plan inside.
The policy optimizers behind game agents and language models alike.
Let the policy practice inside the model's own dream.
The capstone, and the fastest-moving corner of the field. A world model is a network that predicts what happens next, given a state and an action. It's a physics engine learned from data instead of programmed. In one recent year, V-JEPA 2, Genie 3, Dreamer 4, and NVIDIA's Cosmos all landed, and robot policies like Pi0 started turning those predictions into real-world motion. This is where the previous three tiers were heading all along.
Networks that predict what happens next: learned physics engines.
Compress the world to its essentials, then predict in that space.
Learning how the world moves from a million hours of video.
Genie, Cosmos, Dreamer: one model, many worlds.
From predicted futures to motor commands, fifty times a second.