Community Reconstructs List of 30 AI Papers from OpenAI's Former Chief Scientist Ilya Sutskever to John Carmack
Former OpenAI Chief Scientist Ilya Sutskever provided legendary programmer John Carmack with a list of about 30 core AI papers, stating that mastering these materials would cover 90% of the key knowledge in the AI field at that time.
This list originated from Carmack's request for reading suggestions when he shifted to AI research, to which Sutskever responded with about 40 papers and made the aforementioned comment; the original complete list is reportedly untraceable, and the community has reconstructed a version based on various recollections and public discussions, including topics such as RNN, LSTM, AlexNet, ResNet, attention mechanisms, and scaling laws.
Several open-source projects have transformed the listed papers into runnable code implementations. For example, the 30u30 project, with the slogan "30 papers in 30 days covering 90% of AI," offers implementations from scratch of multi-head attention, residual connections, character-level RNNs, variational autoencoders, etc., along with notes, visualizations, and exercises; some implementations use only NumPy to emphasize underlying principles.
The covered content includes the sequence-to-sequence and attention mechanisms from "Attention Is All You Need," the x + F(x) residual structure of ResNet that enables training of hundreds of layers, the mathematical description of the relationship between computing power and performance in Scaling Laws, the gating mechanisms of LSTM, and theories such as MDL and compression as intelligence; the project emphasizes that while frameworks evolve rapidly, the underlying mathematics and architecture remain stable over the long term.
Carmack confirmed in an interview that he had obtained and studied the list, and community projects have accumulated thousands of lines of code with participation from users across multiple countries; related implementations have been made public on platforms like GitHub, allowing for direct execution and extension.
This phenomenon represents a knowledge-structured-driven supply of educational resources, where developers transform an orally recommended list into executable code and systematic courses, shifting focus from tool chasing to foundational paper reproduction; the beneficiaries are engineers looking to systematically grasp the underlying principles of AI, while the pressure is on those who rely solely on the latest frameworks and prompt-based superficial learning paths.
Source: Public Information
ABAB AI Insight
Ilya Sutskever provided a reading list to John Carmack between 2019-2020, who confirmed in an interview that studying it led to a systematic understanding of AI; the community subsequently reconstructed this "mysterious list" and open-sourced implementations, showing a transformation from private recommendations by top researchers to publicly executable courses, covering core evolutions from RNN/LSTM to Transformers and scaling laws.
The resources are primarily directed towards zero-code reproduction and accompanying notes, rather than commercial courses or framework encapsulations; the motivation is to validate the claim that "mastering 30 papers can cover 90% of the key content" and to lower the barrier from theory to practice through runnable implementations, allowing learners to directly engage with principles of residuals, attention, gating, and other long-term stable architectures.
Similar paths can be seen in classic public resources like CS231n course notes and The Annotated Transformer, as well as multiple reproductions of Karpathy's blog and Olah's LSTM visualizations; we are currently in a phase where foundational AI textbooks are expanding from papers and blogs to systematic open-source implementations, with the community attempting to combat the knowledge fragmentation caused by rapid tool iteration.
Essentially, this represents a technological substitution and restructuring of the industrial chain: as the iteration cycles of upper-level frameworks and models shorten to months, the underlying mathematics and architectural principles become long-standing knowledge barriers; by transforming core papers into executable code, knowledge acquisition shifts from "reading abstracts" to "hands-on reproduction," thereby changing engineers' paths and dependency structures in mastering AI capabilities.
ABAB News · Cognitive Laws
Underlying architectures are more durable than upper-level frameworks
Runnable implementations solidify understanding better than paper abstracts
90% of key knowledge is concentrated in a few foundational structures