04 — YouTube Channels & Podcasts

Video and audio content for ML exists on a spectrum from “watched once and understood nothing because it was too shallow” to “paused 47 times and filled a notebook.” This document lists sources that trend toward the latter. The filtering criterion: does this content require you to think, or just consume?

One honest warning: video content is high-bandwidth but low-retention without active engagement. Watch with a notebook. Pause and reproduce code. Reading a paper is often more efficient than watching a 90-minute lecture on it.

YouTube Channels

🔴 Tier 1 — Technical Depth, Primary Learning Material

Andrej Karpathy

  • Channel: youtube.com/@AndrejKarpathy

  • Signal: ⭐⭐⭐⭐⭐

  • Why it’s essential: His “Neural Networks: Zero to Hero” playlist builds everything from the micrograd automatic differentiation engine through a GPT implementation, from first principles. This is not a tutorial — it’s a masterclass in how transformers actually work. Watch with the code open.

  • Key playlist: “Neural Networks: Zero to Hero” (~15 hours total)

  • Also watch: His Stanford CS231n lectures (2016 version on YouTube) and “Let’s build GPT: from scratch, in code, spelled out.”

  • Cadence: Irregular. He posts rarely but every video is an event.

  • Best for: Months 2-8 of the roadmap. Return to it repeatedly as your understanding deepens.

Yannic Kilcher

  • Channel: youtube.com/@YannicKilcher

  • Signal: ⭐⭐⭐⭐⭐

  • Why it’s essential: Paper walkthroughs at a level that assumes you can read math but don’t want to spend 8 hours decoding notation alone. His GPT-4 paper explanation, DALL-E 3, and attention paper walkthroughs are among the best secondary resources for understanding papers.

  • Format: Paper reads (30-90 min per paper), with whiteboard notation. Sometimes manic. Always technically honest.

  • Volume: Very active. Watch 1-2 per week aligned with your reading.

  • Best for: Months 4-13. Paper comprehension support tool.

Stanford University (CS231n, CS224n official uploads)

MIT 6.S191 Deep Learning

  • Channel: MIT OpenCourseWare / IntroToDeepLearning

  • URL: introtodeeplearning.com

  • Signal: ⭐⭐⭐⭐

  • Why it matters: Updated annually. The 2025 version covers diffusion models, LLMs, and RL with modern examples. Dense but accessible. 1-2 hours per lecture.

  • Best for: Month 3-5. Good as a second pass after foundations.

🟠 Tier 2 — Practical & Supplementary

Sentdex (Harrison Kinsley)

  • Channel: youtube.com/@sentdex

  • Signal: ⭐⭐⭐

  • Why it’s useful: Practical code-first tutorials. Good for getting something running quickly. Lower on theory. His PyTorch series is decent for beginners.

  • Best for: Months 1-3 as a code supplement. Not for depth.

Two Minute Papers (Károly Zsolnai-Fehér)

  • Channel: youtube.com/@TwoMinutePapers

  • Signal: ⭐⭐⭐

  • Why it’s useful: 2-5 minute paper summaries with visual demos. Good for staying current on what’s published, especially in vision/graphics.

  • Honest warning: Very surface-level. “Two minutes” means you get the claim, not the mechanism. Use as a discovery tool, then read the actual paper.

  • Best for: All 13 months as a 10-min/week paper discovery digest.

Hugging Face (official channel)

  • Channel: youtube.com/@HuggingFace

  • Signal: ⭐⭐⭐⭐

  • Why it’s useful: Practical tutorials for the entire HF ecosystem: transformers, diffusers, peft, trl. The people who built the libraries explain how to use them.

  • Best for: Months 5-13. Direct tool usage tutorials.

fast.ai (Jeremy Howard)

  • Channel: youtube.com/@howardjeremyp

  • Signal: ⭐⭐⭐⭐

  • Why it’s useful: The video companion to fast.ai courses. Top-down learning philosophy. Real models, real datasets, real code.

  • Best for: Months 2-5. Especially effective for practitioners who learn by building.

Umar Jamil

  • Channel: youtube.com/@umarjamilai

  • Signal: ⭐⭐⭐⭐

  • Why it’s useful: Builds ML architectures (Transformers, LLaMA, Stable Diffusion) from scratch with very clean code. Less famous than Karpathy but excellent implementation quality.

  • Best for: Months 5-10. Implementation companion for key architectures.

Aleksa Gordić — The AI Epiphany

  • Channel: youtube.com/@TheAIEpiphany

  • Signal: ⭐⭐⭐⭐

  • Why it’s useful: Paper implementations in PyTorch, deep dives into transformer variants, graph neural networks. Technical rigor.

  • Best for: Months 6-13. Architecture deep dives.

🟡 Tier 3 — Entertainment / High-Level Awareness

Lex Fridman

  • Channel: youtube.com/@lexfridman

  • Signal: ⭐⭐ (for technical learning) / ⭐⭐⭐⭐ (for big-picture exposure)

  • Honest assessment: Long-form interviews (3-6 hours) with prominent ML researchers and technologists. You will hear Yann LeCun, Geoffrey Hinton, Ilya Sutskever, Sam Altman. The problem: 4-hour interviews are not efficient learning mechanisms. You will feel informed while learning little reproducible knowledge. The occasional interview with deep technical content (his Karpathy interviews, his Wolfram interview) is worth watching.

  • Rule: Watch for inspiration and big-picture, never as a substitute for learning the actual math or code. Budget: max 1 interview per month.

AI Explained

  • Channel: youtube.com/@aiexplained-official

  • Signal: ⭐⭐⭐

  • Honest assessment: Better-than-average explanations of recent AI developments for a non-specialist audience. Good if you need to explain concepts to colleagues or stay aware of the applied landscape.

What Most People Get Wrong About ML YouTube

  1. Confusing watching with learning. You can watch 10 Yannic Kilcher paper reads and retain nothing if you don’t pause, implement, and verify. The correct workflow: watch 10 minutes → pause → open paper → find the equation being discussed → reproduce the logic in a notebook.

  2. Algorithm capture. YouTube’s algorithm will funnel you into shallow “AI is amazing” content. Use subscriptions-only, not the homepage. The homepage is not optimized for your learning.

  3. Tutorial completion theater. Finishing a 30-part tutorial series and feeling like you know the material. The test: can you implement it from scratch without the video? If not, you watched, you didn’t learn.

Podcasts

🔴 Tier 1 — Technical Substance

Machine Learning Street Talk (MLST)

  • Hosts: Tim Scarfe, Keith Duggar, Yannic Kilcher (guest contributor)

  • Where: Spotify, Apple Podcasts, YouTube

  • Frequency: Every 1-2 weeks

  • Signal: ⭐⭐⭐⭐⭐

  • Why Tier 1: The most technically demanding ML podcast. Interviews go deep into the actual mechanism of papers, models, and architectures. Not comfortable listening — you need to pay attention. Guests include leading researchers (Max Tegmark, Gary Marcus, Yoshua Bengio). Covers alignment, ML theory, LLMs, and philosophy of intelligence.

  • Best for: Months 6-13. Requires mathematical foundations to appreciate.

Gradient Dissent (Weights & Biases)

  • Host: Lukas Biewald (W&B CEO) and guests

  • Where: wandb.ai/podcast, Spotify

  • Signal: ⭐⭐⭐⭐

  • Why Tier 1: Focused specifically on ML experimentation, model training, production systems, and the tools practitioners actually use. Less hype than Lex Fridman, more implementation-focused. Good for understanding how frontier labs run experiments.

  • Best for: Months 7-13. Production mindset.

TWIML (This Week in Machine Learning)

  • Host: Sam Charrington

  • Where: twimlai.com, Spotify

  • Frequency: Weekly

  • Signal: ⭐⭐⭐⭐

  • Why Tier 1: Longest-running technical ML podcast (~600+ episodes). Sam Charrington interviews practitioners and researchers. Good signal-to-noise, topic diversity (not just LLMs — covers computer vision, RL, ML systems, healthcare ML).

  • Best for: All 13 months as background research audio.

🟠 Tier 2 — Good, Not Essential

Latent Space Podcast

  • Hosts: swyx + Alessio Fanelli

  • Signal: ⭐⭐⭐⭐

  • Why listen: AI engineering track. Paired with the newsletter. Covers production AI systems, agent frameworks, LLM infrastructure. Very applied.

  • Best for: Months 8-13.

Lex Fridman Podcast

  • Signal: ⭐⭐ (for learning) / ⭐⭐⭐⭐ (for inspiration)

  • Honest assessment: Same note as the YouTube channel. Use with extreme discipline. The Karpathy episodes, Hinton episode, and Bengio episodes have genuine technical content. The Elon Musk episodes do not.

The Retort AI

  • Signal: ⭐⭐⭐

  • Focus: Covers alignment, policy, and the technical/policy intersection. Good for the months you’re working through alignment material.