https://ojhaharsh.github.io/30u30/
A GitHub repository called 30u30 carries an ambitious subtitle: "30 Days. 30 Papers. 90% of AI." With 1,420 stars, it turns Ilya Sutskever's reading list for John Carmack into a 30-day course: one paper per day, each with paper notes, from-scratch Python implementations, five exercises, and visualizations. Looking into a few lessons — Day 13's Attention unit includes a ~30,000-character implementation.py building a Transformer from scratch, plus a minimal training script, cheatsheet, and exercise workbook — the content is seriously made. The repo was created January 19, with the final commit on February 19 adding RLHF solutions for the last lesson (Bradley-Terry preference model, synthetic teacher evaluator, committed all the way through). Then it stopped, untouched for eight months.
The backstory of the list
In mid-2024, a "list of ~30 papers Ilya Sutskever gave to John Carmack" circulated on Twitter and LinkedIn, accompanied by the famous quote: if you really learn these, you'll know 90% of what matters today. Carmack mentioned the exchange in a 2023 interview, so the recommendation actually happened earlier — some estimate 2021 or 2022. An archived version of the original list (the Twitter thread and an Arc folder) contains 27 papers — the contemporary LessWrong compilation was even titled "The 27 papers."
Where 27 and 30 differ
Comparing the lists: the first 28 days of 30u30 map one-to-one to the original list — from Karpathy's RNN blog post, colah's LSTM visual explainer, AlexNet, the two ResNet papers, to Attention Is All You Need, Neural Turing Machines, Scaling Laws, CS231n, and on to Kolmogorov complexity material and Shane Legg's *Machine Super Intelligence*. Day 29 (PPO) and Day 30 (RLHF) are not in Ilya's original list — the repo author added them. It's not an unreasonable addition (RLHF genuinely belongs among "what matters today," and the final commits seriously implement it), but the "Master Ilya's 30 Papers" framing quietly attributes two personal picks to Ilya.
It's not really "30 papers"
Another unstated point: the "30 papers" are a mixed bundle. Counting the composition: 21 papers, plus 4 blog tutorials (Karpathy, Olah, Sasha Rush's Annotated Transformer, Scott Aaronson's complexity dynamics), one open course (CS231n), and two academic works (Legg's doctoral dissertation and the Kolmogorov complexity textbook). Aaronson's "The First Law of Complexodynamics" and the "coffee automaton" piece are, strictly speaking, long blog posts rather than papers. Ilya's original list was like this by design — he wasn't aiming for a literature review, but the shortest path to genuine understanding, and blogs and books belong on that path.
A time capsule
The oldest items are Hinton's 1993 MDL paper and the Kolmogorov complexity material; the newest is 2020's Scaling Laws. Nothing after the Transformer — no diffusion models, no agents. The "today" in "you'll know 90% of what matters today" refers to today as of four or five years ago. This is a foundational map from around 2021: still valid as a base layer, but what lies beyond it is up to you to supplement.
Two complementary learning paths
Comparing learning paths: repos like 30u30 take the depth route — "papers + from-scratch implementation," one month, thirty papers, internalizing the theoretical lineage. Another route is the engineering skills chain, such as a 58k-star open course followed for months — from command line and network protocols up to machine learning fundamentals, one lesson per day. The former gives you the "why," the latter the "how." Foundational papers and foundational engineering are each other's missing half.
An ending frozen at 100%
The 30u30 README ends with "All units complete, bonus papers coming soon!!!" Eight months later, the bonus papers never came. 1,420 stars, 30 units, from-scratch implementations that all run — stopping at the exact moment of 100% completion is, in its own way, a dignified ending.
Sources
- GitHub Ojhaharsh/30u30 (created 2026-01-19, last commit 2026-02-19, 1,420 stars, CC BY-NC-ND 4.0, 30 directories under
papers/, Day 13 implementation spot-checked) - Original list source: email received by Carmack (mentioned in a 2023 interview; circulated on Twitter in May 2024, archived via @keshavchan's post and an Arc.net folder)
- The 27-paper list corroborated by three sources: LessWrong's "The 27 papers" post and GitHub dzyim/ilya-sutskever-recommended-reading
- PPO and RLHF confirmed absent from the original list
- Composition statistics (21 papers / 4 blogs / 1 course / 2 works / 2 additions) verified day by day against the 30u30 directory