Ilya Sutskever's 30 Papers: The Reading List That Covers 90% of What Matters

TL;DR
Ilya Sutskever's rumored 30-paper ML reading list, now on 30papers.com. What the papers cover, the compression thesis behind them, and what HN thinks.
Last updated: July 7, 2026
Back in 2022, Ilya Sutskever reportedly gave John Carmack a list of roughly 30 research papers with the advice: "If you really learn all of these, you'll know 90% of what matters today." The list was never officially published. Someone on Twitter compiled a speculative version in 2024. Now a first-year CS student at Trinity College Dublin has built 30papers.com to make that list more accessible with plain-language explanations.
The project hit the Hacker News front page with 271 points and a discussion that's equal parts appreciation, skepticism about the list's authenticity, and complaints about the website's animations.
What's on the List#
The 30 papers span the foundations of modern deep learning:
Neural network fundamentals:
- CS231n (Stanford's visual recognition course)
- ImageNet Classification with Deep CNNs (AlexNet, 2012)
- Deep Residual Learning (ResNet)
- The Unreasonable Effectiveness of RNNs (Karpathy's famous blog post)
- Understanding LSTM Networks (Colah's explanation)
Attention and transformers:
- Neural Machine Translation by Jointly Learning to Align and Translate
- Attention Is All You Need
- The Annotated Transformer (Harvard's implementation guide)
Scaling and training:
- Scaling Laws for Neural Language Models
- GPipe: Pipeline Parallelism for Training
Theory papers:
- Kolmogorov Complexity and Algorithmic Randomness
- A Tutorial Introduction to the Minimum Description Length Principle
- Quantifying the Rise and Fall of Complexity in Closed Systems (The Coffee Automaton)
The theoretical papers are what make this list distinctive. They're not standard deep learning curriculum - they're information theory and complexity theory papers that connect to Ilya's thesis that learning is compression.
The Compression Thesis#
Several HN commenters picked up on why the Kolmogorov complexity papers are included. As one explained: "Ilya argues that the reason why neural networks generalize - why they work at all - is because they are effectively finding a simple description of their training data, converging down onto the limit of the Kolmogorov complexity."
Another linked this to Solomonoff induction, which "combines Kolmogorov complexity with Bayes rule to provide a general framework for inductive inference, and naturally formalizes Occam's razor."
This is the reading list's hidden curriculum: the papers don't just teach you how to build neural networks, they explain why they work. Good models compress their training data; bad models memorize it.
What HN is Saying#
The Hacker News discussion raised several concerns.
Is this actually Ilya's list? Multiple commenters questioned the provenance. One noted: "Someone posts on X, 'These are Ilya's 30 papers', gives no source, doesn't say where he got it from, and isn't connected to either Ilya or Carmack. Then someone vibe codes a barely usable website based on that, and it lands on the HN front page?"
The author acknowledged this on the site: "rumoured list of papers that Ilya Sutskever gave to John Carmack." Whether it's the exact list or a reasonable reconstruction, the papers themselves are canonical - these are the foundational works in deep learning.
The UX is rough. The site features heavy animations with scrolling effects. Multiple commenters reported headaches and dizziness: "I scoffed at your comment and went to the website. After scrolling a bit, I find myself having a mild headache and slight dizziness."
The technical issues are more severe. LaTeX formulas render incorrectly with flattened subscripts and superscripts. Images and tables don't render at all. One commenter posted the direct paper links as a service to others.
The format question. A core debate: what value does the site add? One commenter asked: "Is it just rehosting the list, plus a reformatted copy of the papers? I was hoping you'd have at least annotated them with what you'd learned?"
The author, a first-year CS student, explained his motivation: "When I was getting into reading research papers I ended up burning a ton of my Claude usage asking questions other people have probably already asked." The site hosts papers with inline plain-language explanations of difficult terms - essentially baking in the questions you'd ask Claude.
Reading order matters. Several people noted the list isn't ordered for learning: "The paper introducing the attention mechanism probably ought to precede 'Attention Is All You Need.'" This is a reasonable critique - the list was given to Carmack, who already had significant ML background.
The Meta-Commentary#
The discussion produced a useful perspective on curated reading lists in the LLM era:
"Compiled resources for nerds are catnip. Hit that bookmark/upvote button to never get to it :)"
There's truth here. The list has circulated for years, spawned multiple GitHub compilations, and even a Manning book. Most people who bookmark it won't read the papers. But that's always been true of reading lists.
What's different now: you can actually process these papers efficiently. Tools like Claude, NotebookLM, and various PDF-to-audio services make it practical to work through dense research. One commenter even shared their own tool for generating teacher-style audio explanations of papers.
The Actual Links#
For those who just want the papers without the animations:
- CS231n
- AlexNet paper
- ResNet
- Karpathy's RNN post
- Understanding LSTMs
- Attention Is All You Need
- The Annotated Transformer
- Neural Turing Machines
- Scaling Laws
- Kolmogorov Complexity book
The full list is available on several GitHub repositories, including this curated version with summaries and study roadmaps.
Why This Matters#
The list's real value isn't as a reading assignment - it's a map of what one of the field's most influential researchers considered foundational. The theory papers alongside the architecture papers. The explanatory blog posts alongside the formal research. The Stanford course that taught a generation of ML engineers.
If you're learning ML in 2026, you have better resources than this list. But if you want to understand how the people who built modern AI thought about these problems, this is the reading.
FAQ#
Is the 30 papers list really from Ilya Sutskever?#
Its provenance is unconfirmed. The story is that Ilya Sutskever gave John Carmack a list of roughly 30 papers in 2022, but the list was never officially published, and the version that circulates was compiled speculatively on Twitter in 2024. Even 30papers.com describes it as a rumoured list. The papers themselves are canonical works in deep learning either way.
What papers are on Ilya Sutskever's reading list?#
It spans neural network fundamentals (CS231n, AlexNet, ResNet, Karpathy's RNN post, Colah's LSTM explainer), attention and transformers (the Bahdanau attention paper, Attention Is All You Need, The Annotated Transformer), scaling and training (Scaling Laws, GPipe), and theory papers on Kolmogorov complexity and the Minimum Description Length principle.
Why does the list include Kolmogorov complexity papers?#
They connect to Ilya's thesis that learning is compression: neural networks generalize because they find a simple description of their training data, converging toward its Kolmogorov complexity. Those theory papers explain why deep learning works, not just how to build it.
In what order should I read the 30 papers?#
The list is not ordered for learning. Commenters pointed out, for example, that the paper introducing the attention mechanism should come before Attention Is All You Need, since the list was originally given to someone who already had significant ML background.
Continue Reading#
- Detecting LLM Text with Classical ML: TF-IDF Still Works
- LLM Architectures Got Complicated Fast
- Transformers.js: Run AI Models Directly in the Browser
Sources#
Get the next deep dive like this in your inbox
One email a week on News and the rest of the AI dev stack. Free.
Read next
Terence Tao Digests the Jacobian Conjecture Counterexample: How Claude Fable 5 Broke an 87-Year-Old Math Problem
Claude Fable 5 found a counterexample to the 87-year-old Jacobian conjecture, and Terence Tao worked through the math. How it works and what HN is saying.
9 min readGhost Font: Text That Humans Can Read But AI Cannot
Ghost Font hides text in video using motion, so humans can read it but frame-by-frame AI models cannot. How it works and how HN commenters broke it.
6 min readLLM Architectures Got Complicated Fast
LLM architectures got complicated: MoE routing, mixed attention variants and integrated vision encoders replaced the simple transformer. What changed and why.
6 min readNew here? Start with
Technical content at the intersection of AI and development. Building with AI agents, Claude Code, and modern dev tools - then showing you exactly how it works.







