Seven AI YouTube channels, covering paper breakdowns and scratch-built code. Watch the ones that teach, skip the ones that perform.
Károly Zsolnai-Fehér · Research Made Visual
Károly Zsolnai-Fehér turns thirty-page papers into five-minute visual summaries, showing generative video and physics before they hit mainstream. Fastest pulse-check on the bleeding edge.
Watch Two Minute Papers ↗Yannic Kilcher · Line-by-Line Paper Breakdowns
Yannic Kilcher reads the paper so you do not have to, walking the math and architecture on a literal whiteboard. What separates him is the willingness to call a flawed methodology flawed, an exaggerated claim exaggerated, receipts always included.
Watch Yannic Kilcher ↗Jason Zhou · Applied Agentic Engineering
AI Jason skips the theory and goes straight to the workflow: RAG pipelines, multi-agent orchestration, production-ready tools built live, right on screen. It is the rare channel that treats "building an agent" as an engineering problem with real tradeoffs, not a demo chasing a wow moment, and the code proves it, business case after case, tutorial after tutorial.
Watch AI Jason ↗Harrison Kinsley · From-Scratch Python & ML
Harrison Kinsley has taught Python from first principles for over a decade, and his channel refuses every shortcut that would make content easier and understanding shallower. Reinforcement learning, self-driving simulations, neural networks: all built from scratch, no high-level library hiding the mechanics from you. This is not a channel for people in a hurry, it is for people who want to understand what is happening underneath the abstraction, one small piece at a time.
Watch Sentdex ↗Andrej Karpathy · Deep Learning From First Principles
Andrej Karpathy helped found OpenAI and ran AI at Tesla, and teaches like someone who actually built the frontier rather than read about it secondhand. His long-form "let's build from scratch" sessions construct GPT tokenizers and backpropagation engines live, on screen, nothing hidden, no slides standing in for actual code that runs. "Neural Networks: Zero to Hero" is the closest thing to a free graduate course in deep learning online, taught by someone with nothing left to prove and every reason left to simply teach it well, start to finish.
Watch Andrej Karpathy ↗Josh Starmer · Statistical Intuition, Built Up
Josh Starmer's premise is that intimidating notation is a teaching failure, not a prerequisite, and StatQuest proves the point relentlessly, video after video, algorithm after algorithm. Principal component analysis, probability, transformer architecture: rebuilt as step-by-step visual reasoning, scaffolded so each video assumes only the one directly before it. It will not tell you what shipped this week or which benchmark just got beaten by a newer model. What it gives you instead is the foundation underneath the acronyms, the kind that makes every other channel on this list easier to follow once you have it. Not the frontier — the bedrock underneath it.
Watch StatQuest ↗AI Explained · Sober Benchmark & Capability Analysis
Most AI channels repeat a company's press release with better editing and a friendlier tone. AI Explained tests the model instead: difficult logic and reasoning tasks designed to break the marketing claim, not confirm it for a thumbnail. That is the fold — it takes the industry's own benchmarks and runs them back through independent testing before vouching for anything, publicly, with failures included. The weekly roundups compress a chaotic week of releases into a dense, sober summary that treats your attention as finite and worth respecting, not a metric to maximize. A channel willing to say a benchmark does not mean what the announcement implies is doing something closer to journalism than most outlets that call themselves journalism.
Watch AI Explained ↗