dataofmen

30 Papers - 일리야 서츠케버 추천 AI 핵심 논문 목록

4 min read

30 Papers - 일리야 서츠케버 추천 AI 핵심 논문 목록

Hada URL: https://news.hada.io/topic?id=31224
Canonical URL: https://30papers.com/
Captured: 2026-07-10 06:35 KST
Fetch method: ultimate-fetcher --json via Jina for both Hada summary and canonical site.

Summary

30papers.com is a beginner-friendly learning site based on a rumored list of papers that Ilya Sutskever gave John Carmack. The site states that it currently has 27 items, not the full canonical 30. It gathers foundational papers, course notes, blog posts, and code-based explanations that trace the development of modern deep learning and large language models.

The value of the site is not merely that it lists papers. It lowers the entry barrier into the original papers by organizing them as a learning path across computer vision, sequence modeling, attention, transformers, external memory, graph neural networks, scaling laws, information theory, and complexity theory.

Hada summary

  • The project is based on the well-known story that Ilya Sutskever gave John Carmack a core AI reading list.
  • The list covers major development lines in modern AI: deep learning, computer vision, sequential modeling, attention, transformers, graph neural networks, scaling laws, information theory, and complexity theory.
  • It pairs papers with lecture notes, explanations, and code-oriented resources to lower the barrier to original papers.
  • The list is known as a 30-paper list, but the current website organizes 27 items.

Canonical site structure

The site includes entries such as:

  • CS231n: Convolutional Neural Networks for Visual Recognition — convolutional networks from first principles.
  • The Unreasonable Effectiveness of Recurrent Neural Networks — Karpathy’s practical character-level RNN exploration.
  • Understanding LSTM Networks — Christopher Olah’s visual explanation of LSTM gates.
  • ImageNet Classification with Deep Convolutional Neural Networks — AlexNet and the launch point of modern deep learning.
  • Deep Residual Learning for Image Recognition — ResNet and residual connections.
  • Multi-Scale Context Aggregation by Dilated Convolutions — wider receptive fields without losing resolution.
  • Recurrent Neural Network Regularization — correct dropout placement in LSTMs.
  • Neural Machine Translation by Jointly Learning to Align and Translate — attention for machine translation.
  • Attention Is All You Need — Transformer architecture.
  • The Annotated Transformer — executable explanation of the Transformer.
  • Neural Turing Machines — differentiable external memory.
  • Relational reasoning and graph-related models — relation networks and message passing.
  • Scaling Laws for Neural Language Models — power-law regularities in model scaling.
  • Minimum Description Length and Kolmogorov complexity — information-theoretic and complexity-theoretic foundations.

Why it matters

The site is useful because it frames modern AI as a sequence of conceptual moves rather than as a scattered list of famous papers. The path moves from perception architectures to sequence models, then to attention, memory, relations, scaling, and theory. This is closer to how a researcher builds mental models.

For beginners, the biggest obstacle in paper reading is not access to PDFs but lack of scaffolding. 30papers.com addresses this by combining original sources with plain-language summaries, contributor context, and implementation-oriented explanations.

Strategic interpretation

This is a research-learning infrastructure example. In the AI era, value often comes from turning raw expert artifacts into navigable learning paths. The same pattern applies to Henry’s wiki pipeline: a clipping is not enough; it becomes useful when it is decomposed into concepts, entities, relations, and human-readable notes.

The site also shows why paper reading remains important even when LLMs can summarize papers. Summaries are easy; building a durable map of architectural ideas and their dependencies is harder. That map is what lets a learner understand why modern LLMs look the way they do.

Key takeaways

  • Paper lists become more useful when organized as a conceptual path.
  • The barrier to original papers is reduced by summaries, diagrams, code, and prerequisite ordering.
  • Modern AI history can be read as a transition from CNN/RNN foundations to attention, memory, scaling, and theory.
  • The list is a learning scaffold, not a substitute for reading the original papers.
  • The current site says it has 27 items rather than a full canonical 30.

그래프 뷰