Math principles + runnable code + discussion in one integrated package, GitHub 76.9k★／12.2k fork, adopted by 500+ universities in over 70 countries, physical book published by Cambridge University Press. The codebase has largely stabilized since the 2.0 release in December 2022—a classic textbook representative of "solid foundations but no longer growing."
One of the "three classics" of Chinese AI, but the original book's content has not been updated since 2016; its activity is mainly reflected in early contributions—a reference point of "classic but outdated," which has stagnated after core maintainers left.
Full-pipeline Agent tutorial: paradigms, frameworks, custom frameworks, MCP/A2A, Agentic-RL, evaluation, 58.8k★／7.2k fork, covering the entire chain from ReAct to GRPO reinforcement learning. Deep dive in DataWhale feature.
30.1k★／3k fork, supporting 50+ models, with dedicated sections for AMD/Ascend/Apple M-series hardware—targeted coverage of the domestic compute ecosystem that is almost never seen in international tutorials.
29k★／2.7k fork, hand-writing Transformers and training practice rather than calling APIs; has been incorporated into the CCF × DataWhale × GitLink AI accessibility curriculum with free compute provided.
The 2026 Spring version has fully pivoted toward Agents: Context Engineering, Flash Attention, Self-Correction, Test-Time Scaling; completely free on Bilibili/YouTube, rebuilt from scratch every year.
Offered for three consecutive terms (Spring 2024/2025/2026), hand-writing the full pipeline from data cleaning to deployment, echoing Karpathy's nanochat philosophy; YouTube playlist publicly available.
Advanced to V6 in the 2025–2026 academic year; anyone can audit for free via Zoom without registration or university affiliation; lectures uploaded to YouTube 1–2 weeks later, often featuring top researchers like Hinton and Vaswani.
Complete 2024 lectures publicly available on YouTube; lecture notes updated annually, completely free and open—the standard path for systematically learning NLP + deep learning fundamentals.
Current semester lectures are restricted to on-campus Canvas access; only past recordings are public—a common misconception is that it is "fully open"; you need to combine past recordings + assignments for self-study.
2026 version completed on-campus; online version opens weekly livestreams, fully open-sourced after the course ends; an intensive "one-week bootcamp" design, suitable for quickly building a comprehensive understanding of deep learning.
The Agents Course (early 2025) and MCP Course (launched 2025.5.22) successively became the hottest courses of the year, closely following protocol trends; free learning + certificate upon completion. Deep dive in Hugging Face feature.
In October 2025, Andrew Ng personally taught the new course "Agentic AI" (four major patterns: Reflection/Tool Use/Planning/Multi-agent), accompanied by multiple short courses; certificates available upon completion.
Completely redesigned in November 2024 with a modular interactive design; March–April 2025 partnered with Kaggle for "GenAI Intensive" livestream courses; some courses offer certificates.
About 20 courses; the claim of "launching in March 2026" is questionable (first-hand evidence shows the infrastructure existed at least as early as August 2025, more likely a brand relaunch); the newly launched Claude Certified Architect certification in March 2026 is official, at $99/Pearson VUE proctored. Deep dive in Big Tech training comparison feature.
YouTube subscribers surpassed 1 million in early 2026; nanochat (2025.10) "$100 to train the best ChatGPT" sparked extensive tech media coverage, hand-writing code from scratch, "only one complexity knob." Deep dive in Karpathy feature.
YouTube subscribers approximately 8.38 million, with about 70k growth in the last 30 days; animated visualizations of mathematical intuition, intensively completed the Transformer/Attention series in 2024.
Single episodes on Bilibili can reach hundreds of thousands to millions of views; "three-pass paper reading method," covering 67 classic papers from AlexNet to Sora.
YouTube subscribers approximately 1.61 million; hand-drawn illustrated statistics, "explaining it in plain English"; published a companion illustrated textbook in 2025.
Long-term active creator; showcased a self-developed robotics project at WAIC 2025; classic paper deep reads + CS231n Chinese walkthroughs, a dual track of "science communication + practice."
Released "How to Solve It With Code" and the Solveit platform in October 2025, shifting from "top-down deep learning teaching" to "teaching human-AI collaboration while preserving engineering judgment."
Skill Builder + AWS Certified AI Practitioner tiered certification, Pearson VUE proctored; the only one with an official "commitment-to-achievement" scale data closed loop (31M+ people trained).
Microsoft Learn + AI-900/AI-103 + AB-730; AI-102 will retire on 2026.6.30, with AI-103 (Agent direction) taking over.
Baidu Smart Cloud Academy (released September 2025) + "Generative AI Application Engineer," jointly certified by the Ministry of Industry and Information Technology Education and Examination Center; over 15,000 certified learners.
HCIA-AI / HCIP-AI / HCIE-AI; V4.0 adds Pangu Ultra MoE content; over a decade of ICT certification history gives it widely recognized market credibility.
Clouder free certification + ACA/ACP Large Model Engineer certification, 600–1200 RMB; ACP requires in-person proctoring, which is rare domestically.
AI University + Prompt/Fine-tuning/Agent Engineer certification; 4 million developers is an overall ecosystem figure; the first "Spark Camp" had only 115 people—an elite cultivation model.
About 72.6k★, official repository of API usage examples and best practices; a quick reference rather than a tutorial, not constituting a systematic course.
Strong at technical teardowns and in-depth industry reporting; suitable for "tracking developments" rather than "systematic learning"—setting expectations apart from systematic courses.
The company does not maintain an independent teaching repository; teaching content is highly dependent on third-party communities like DataWhale.
Chinese version of Andrew Ng's deeplearning.ai, Zhejiang University's "AI: Models and Algorithms," etc., aimed at a broader range of learners without a CS background.
With 36 resources laid out on a single map, what you see is not a collection of isolated course listings, but an educational supply structure that is rapidly reorganizing. The following four cases each represent the extreme of four paradigms: open-source community crowdsourcing (DataWhale), platform-type certified certificates (Hugging Face), personal-brand teaching (Karpathy), and how the companies building models teach themselves (a comparison of official training across 16 companies).
Founded in 2018, the GitHub organization datawhalechina has 215 repositories; its core mechanism is "open-source tutorials + team-based check-ins + community co-creation." The three currently hottest projects form a complete progression path: hello-agents (58.8k★, full Agent pipeline, 2026's hottest project, project lead Chen Sizhou, advising expert Zhu Xinzhong), self-llm (30.1k★, local deployment and fine-tuning, including AMD/Ascend/Apple M hardware sections), and happy-llm (29k★, hand-writing LLM principles, incorporated into the CCF×DataWhale×GitLink AI accessibility curriculum). The core mechanism is the "Whale Elite Tutor" system—large numbers of current students participate in content creation, ensuring rapid iteration while forming a self-sustaining talent funnel. The limitation is that fewer than one-tenth of the 215 repositories may be truly active, and newcomers can easily lose their way on the GitHub organization page. The closest overseas mirror is the Hugging Face course matrix.
On huggingface.co/learn, seven or eight "foundation-type" courses (NLP/LLM Course, Deep RL Course, Diffusion Models Course) are maintained long-term; what truly reveals the strategy is the newly added Agents Course in early 2025 (five units covering smolagents/LangGraph/LlamaIndex) and the MCP Course (launched 2025.5.22, only a few months after the MCP protocol itself was released)—arriving almost in sync with the protocol's industry adoption. Class Central lists over 80 HF online courses, distributed in partnership with DataCamp. Behind the "free learning, certificate upon completion" model is a clear acquisition logic: while earning their certificates, learners have already made dozens or hundreds of real calls in HF's model library, Spaces, and Transformers library—which also explains why Anthropic Academy only arrived in March 2026. The closest domestic mirror is DataWhale.
The starting point of his teaching reputation is the YouTube playlist "Neural Networks: Zero to Hero"—starting from backpropagation, hand-writing code up to GPT; his teaching philosophy is "don't give you a ready-made library, make you hand-write it first." In July 2024, he founded Eureka Labs; the core course LLM101n planned to guide students through building a language model from scratch, but as of mid-2026 it has long remained in early development without a full release; in 2026, Karpathy joined Anthropic's pretraining team, further lowering Eureka Labs' priority. The control group is nanochat, released in October 2025—the tagline "The best ChatGPT that $100 can buy," a single code repository covering the full pipeline from tokenization to deployment, running through on an 8×H100 node in about 4 hours for roughly $100, retaining only one complexity knob (transformer depth), igniting the community within weeks. The same person: the institutional path hasn't fully released in two years, while the personal path crystallized in weeks—this contrast is more persuasive than any theoretical analysis, and together with Stanford CS336 and DataWhale happy-llm, it confirms that "training from scratch" is becoming the new consensus paradigm for advanced learners.
The 16 companies present a clear three-tier structure: companies with cloud computing DNA (AWS, Microsoft, Baidu, Huawei, Alibaba Cloud, iFlytek) graft large model training onto existing certification exam systems, creating a complete "course + exam + certificate" closed loop—AWS is the only company with an official "commitment-to-achievement" closed-loop data (31M+ people trained); pure model companies (Anthropic, OpenAI) only tried certification for the first time in the past year, and the value is yet to be market-tested; and a group of companies (Meta, Zhipu, ByteDance, xAI, DeepSeek, Kimi) basically don't treat education as a product line. Counter-intuitive detail: the claim that "Anthropic Academy launched in March 2026" is itself questionable—first-hand evidence shows the infrastructure existed at least as early as August 2025, more likely a brand relaunch; Microsoft AI-102 and AWS ML Specialty are retiring old certifications in the same time window, making way for generative AI/Agent new certifications; when DeepSeek and Kimi have no official training, a large number of scam training classes masquerading as "MIIT certifications" appear on the market—the gap between hype and educational supply will always be filled by someone, not necessarily a good actor. Selection guide: for resume-worthy certificates, prioritize Huawei HCIA/HCIP-AI or AWS Certified AI Practitioner; for enterprise bulk purchasing, AWS's verifiable closed-loop data is the first choice.
The four collectively answer the same question: when model capabilities iterate every few months, how can "teaching AI" itself keep up? The answer so far is—no single model can keep up alone, but companies chase hotspot speed, individuals chase trust depth, communities chase systematic completeness, and cloud computing giants chase scaled certification; only when the four curves叠加 together do they barely match the iteration speed of the models themselves.
First published 2026-07-24