YouTube Channels¶

The best video pedagogy in ML systems, curated. Watch on 1.5x speed. Take notes.

The rule¶

One lecture per week. Not two. Not zero. Lectures are dessert; code is dinner. If you watched a lecture and didn’t write code afterward, you didn’t learn anything, you consumed content.

Note-taking format (in your lab notebook, one page per lecture):

Lecture: [Title] — [Speaker] — [Channel]
Date watched:
One-sentence summary:
Three things I learned:
1.
2.
3.
One thing I'll try in code this week:
One question I still have:

Tier 1 (mandatory)¶

GPU MODE¶

URL: https://www.youtube.com/@GPUMODE
Companion: https://github.com/gpu-mode/lectures (code + slides)
Why: The core lecture series for this entire roadmap. See 01_gpu_mode.md for the full lecture-by-lecture roadmap alignment.
Pace: one lecture per week, phase-aligned.

Umar Jamil¶

URL: https://www.youtube.com/@umarjamilai
Why: Best paper-by-paper implementation walkthroughs on YouTube. He builds Llama, DeepSeek-V3, RoPE, GQA, RLHF, DPO, distributed training from scratch on camera. Every video is 2–4 hours; treat as full guided lab sessions.
Playlist priority for this roadmap:
1. “Coding LLaMA 2 from scratch in PyTorch”
2. “Coding a Transformer from scratch on PyTorch, with full explanation, training and inference”
3. “Coding DeepSeek-V3 from scratch”
4. “Distributed Training with PyTorch: complete tutorial with cloud infrastructure and code”
5. “Direct Preference Optimization (DPO) explained: Bradley-Terry model, log probabilities, math”
6. “Mistral / Mixture of Experts, explained”
7. “Flash Attention explained” — companion to your Triton FA2 work
8. “LoRA & QLoRA” — for Phase 5–6

Stanford CS336 recordings¶

URL: https://www.youtube.com/playlist?list=PLoROMvodv4rOY23Y0BoGoBGgQ1zmU_MT_ (Spring 2025 playlist — verify current-year link at https://stanford-cs336.github.io/)
Why: The single best-fit university course for this whole trajectory. Lecture 6 (Kernels/Triton), Lecture 10 (Inference), Lectures 7–8 (Parallelism) are your material.
Pace: 1–2 lectures per week during Phase 1–4.

Tier 2 (highly recommended)¶

Andrej Karpathy¶

URL: https://www.youtube.com/@AndrejKarpathy
Why: Zero-to-Hero series is the foundation for Phase 1. Micrograd → Makemore → nanoGPT → GPT-2 rebuild → tokenizer. His “Let’s build the GPT tokenizer” and “Let’s reproduce GPT-2” are must-watch.
Also: “[1hr talk] Intro to Large Language Models” as a mental-model refresher.

Sasha Rush¶

URL: https://www.youtube.com/@srush_nlp (variable; his content also lives on GitHub as tutorials)
Why: GPU Puzzles, LLM Training Puzzles, Tensor Puzzles — interactive companions to his notebooks. Great in Phase 1–2.

3Blue1Brown¶

URL: https://www.youtube.com/@3blue1brown
Why: His “Neural Networks” series (specifically the attention/transformer chapters) are the best visual intuition-builders. Watch Phase 1 for intuition anchors; don’t rely on them for depth.

Yannic Kilcher¶

URL: https://www.youtube.com/@YannicKilcher
Why: Paper walkthroughs. Uneven quality but useful for a first-pass understanding of a new paper before you sit down with it yourself. Not a substitute for reading.

Tier 3 (targeted, watch when relevant)¶

NVIDIA (GTC + On-Demand)¶

GTC playlist: https://www.nvidia.com/en-us/on-demand/
YouTube: https://www.youtube.com/@NVIDIADeveloper
What to watch:
- Every GTC keynote from Jensen (know what’s shipping)
- CUTLASS deep-dive sessions (Vijay Thakkar, Cris Cecka)
- TensorRT-LLM deep-dives
- Hopper/Blackwell architecture deep-dives
Filter: most GTC content is product marketing. The named technical deep-dives are gold.

AMD ROCm (awareness)¶

YouTube: https://www.youtube.com/@AMDDevCentral (search ROCm/MI300X)
Why: Phase 2 awareness of the alternative stack.

PyTorch¶

URL: https://www.youtube.com/@PyTorch
What to watch: PyTorch Conference recordings each year — several inference/serving/compile deep-dives per conference.

MLSys conference (published talks)¶

URL: https://mlsys.org/ + their YouTube (search “MLSys 20XX”)
What to watch: Best-paper talks. FlashInfer, Mooncake, etc.

USENIX (OSDI / NSDI / ATC recordings)¶

URL: https://www.usenix.org/publications/multimedia
What to watch: Orca (OSDI ‘22), DistServe (OSDI ‘24), and other systems papers you’re reading. Getting a 20-min talk before reading the paper is often the best on-ramp.

MIT HAN Lab (Song Han)¶

URL: https://www.youtube.com/@MITHANLab
Why: The AWQ / SmoothQuant / TinyChat authors’ lab. Occasional deep-dive on efficient AI.

Latent Space (Swyx + Alessio)¶

URL: https://www.youtube.com/@LatentSpaceTV
Podcast form; watch/listen when they study inference-eng people specifically: Tri Dao, vLLM/SGLang crews, DeepSeek/Kimi people. Skip the general-AI-industry episodes.

AI Engineer Summit (talks)¶

URL: https://www.youtube.com/@aidotengineer
Filter aggressively: most is product-layer; the infra track sometimes has serving deep-dives.

Awareness-tier (rare-glance)¶

TensorPilot — https://www.youtube.com/@tensorpilot — short-form GPU/systems content; verify current activity before subscribing.
Efficient ML (Song Han’s course, MIT 6.5940) — lectures released publicly some years; check MIT OCW.
CMU 10-708 / 10-414 / Sasha Rush’s CMU/Cornell lectures — for adjacent depth.
Latent Space Papers Podcast, Machine Learning Street Talk, Cognitive Revolution — background listening; low code density.

What to unsubscribe from¶

Any “AI news” channel with weekly “Everything that happened in AI this week” videos. Zero technical content.
Any channel whose thumbnails feature large arrows and shocked faces.
Any channel selling a course whose curriculum overlaps with this roadmap. You already have the syllabus.

Watchlist queue (managed like your paper queue)¶

Maintain a ~/notes/youtube-queue.md file:

## To watch this week
- [ ] GPU MODE Lec X — topic (Phase Y aligned)

## To watch this month
- [ ] Umar Jamil DeepSeek-V3 (Phase 4)
- [ ] CS336 Lec 10 Inference (Phase 4)

## Watched — rated 5/5
- [x] Simon Boehm equivalent talks etc

Rating your notes forces you to be honest about what actually added value.