In December 2021, the release of the Chinese beta version of \"Dive into Deep Learning\" (D2L) 2.0 marked a critical inflection point in open-source AI education. By expanding its code implementations to include PyTorch and TensorFlow alongside the original MXNet, the project signaled a recognition of the shifting developer landscape—a move that proved prescient given the subsequent consolidation of research workflows around PyTorch.
\nThe release of D2L 2.0 arrived at a moment when the deep learning framework wars were still actively contested, though clear leaders were emerging. The original iteration of the textbook was heavily associated with MXNet, a framework championed by Amazon Web Services. While technically robust, MXNet's market share was already waning relative to the explosive growth of Facebook's PyTorch and Google's TensorFlow. The 2.0 update, released initially as a beta for Chinese readers, directly addressed this friction by offering what the authors described as a textbook that is "executable and discussable" across all three major frameworks.
The core value proposition of the 2.0 release was its unified approach to code implementation. Unlike competitors such as the "Deep Learning" textbook by Goodfellow et al., which focused purely on theory, or Fast.ai, which emphasized a top-down, code-first approach using a specific library wrapper, D2L 2.0 attempted to balance mathematical rigor with implementation details. The authors provided parallel codebases for NumPy/MXNet, PyTorch, and TensorFlow, allowing students and researchers to toggle between frameworks without losing the theoretical thread.
This multi-framework support was not merely a feature update; it was a survival strategy for the curriculum. By late 2021, academic papers were increasingly implemented in PyTorch. Had D2L remained an MXNet-exclusive resource, it risked obsolescence despite its high-quality theoretical content. The beta release explicitly targeted the Chinese developer community first—a demographic that has historically been rapid adopters of new AI tooling—with the documentation noting the content was "oriented towards Chinese readers" with plans for continuous updates.
At the time of the beta release, the project reported utilization by "300 universities across 55 countries". This metric highlights the vacuum that existed for a comprehensive, open-source curriculum that could serve as a standard university text. The interactive format, likely leveraging Jupyter Notebooks, allowed for the immediate execution of complex models, lowering the barrier to entry for students lacking extensive software engineering backgrounds.
Viewing this 2021 release through the lens of the current AI landscape offers several insights. First, the decision to support PyTorch was validated by the industry's near-total shift toward that framework for research and Large Language Model (LLM) development in the years following. The D2L curriculum provided a critical bridge for students who learned the fundamentals just before the Generative AI boom of 2022-2023.
However, the "beta" status and the initial language restriction highlighted the difficulties of maintaining a living textbook. Keeping three separate framework implementations in parity is a significant engineering burden. In the years since, while TensorFlow remains relevant in production environments, the pedagogical momentum has swung decisively toward PyTorch. The D2L 2.0 release captures a specific moment in tech history where framework diversity was high enough to necessitate a "Rosetta Stone" approach to deep learning education.
Compared to the Coursera Deep Learning Specialization by Andrew Ng, which dominated the MOOC space, D2L offered a more granular, open-source alternative that did not require a subscription. Where Fast.ai abstracted away complexity to get students shipping models quickly, D2L 2.0 forced an engagement with the underlying tensor operations, a distinction that became crucial as engineers moved from fine-tuning models to architecting custom solutions in the LLM era.
\n\n