Ilya Sutskever: “If you learn all of these, you’ll know 90% of what matters”

Original Article ↗ Hacker News Discussion ↗

A purported 30-item AI reading list attributed to Ilya Sutskever and reportedly shared with John Carmack is circulating, prompting scrutiny over its authenticity and whether it really represents “90% of what matters” in modern machine learning. Commenters note that while the papers and courses linked are foundational—covering CNNs, RNNs, LSTMs, transformers, and Kolmogorov complexity—the list omits areas like reinforcement learning, diffusion models, and many practical LLM techniques, and assumes a heavy math and CS background. Much of the debate centers on how long it would take to truly learn this material, who the list is realistically for, and whether such theory-heavy study without a concrete project goal is the best path into contemporary AI work.

Authenticity and Possible Product Tie‑in

  • Multiple commenters question whether the list is genuinely from the cited researcher or just “someone’s bookmarks.”
  • Some evidence is cited (interviews, social posts, an ex-employee’s onboarding notes), but others note there is no direct, explicit confirmation that this exact list is authentic.
  • A few see it as subtle marketing for a VC‑backed browser; others push back, saying the content looks legitimate and useful.
  • Overall status: plausible but unverified; several people explicitly flag that it may not be the actual list.

Scope and Content of the List

  • Despite being described as ~30 “papers,” it includes a full CNN course, a ~500‑page Kolmogorov complexity book, and other long texts.
  • Core topics include classic deep learning, RNNs/LSTMs, attention/transformers, and some foundational theory (e.g., Kolmogorov complexity).
  • One commenter notes a specific chapter range in the Kolmogorov book as especially recommended.

Relevance, Coverage, and Obsolescence

  • Some argue the list omits major modern areas: reinforcement learning, diffusion models, graph neural networks, low‑bit networks, and LLM‑era engineering (in‑context learning, RAG, tools, multimodality).
  • The claim that this covers “90% of what matters today” is viewed as bold and “very opinionated.”
  • Several note the list is years old and may be dated relative to current LLM practice.
  • Others suggest the missing topics might reasonably be the remaining “10%.”

Effort Required and Intended Audience

  • Commenters stress that “reading” ≠ “learning”: true understanding would take substantial time.
  • Estimates range from a year of full‑time effort to several years part‑time, depending on prior math/CS background.
  • For someone with no relevant background, some suggest 5+ years of full‑time study to reach the level assumed.
  • Many emphasize that the list was reportedly tailored for a highly experienced engineer, not for beginners.

Learning Strategies, Constraints, and Tools

  • Discussion on prerequisites: calculus, linear algebra, statistics, algorithms, and some learning theory are recommended.
  • Strategies: follow bibliographies forward and backward, look for survey papers, build a “narrative arc” across seminal works.
  • Several stress having a concrete project or goal; otherwise the material is likely to be forgotten and used only for “armchair” commentary.
  • Some suggest using modern language models as tutors: read, then ask questions when stuck.

Time, Life Responsibilities, and Trade‑offs

  • A long subthread debates the feasibility of “locking yourself in a hotel for a week” to study.
  • One side argues that most professionals could carve out such time if they truly prioritize it; others counter that family, health, and financial constraints make this unrealistic for many.
  • There is reflection on prioritization, ambition, and the tension between career advancement and other life responsibilities.

Practical Handling and Community Resources

  • Commenters share backups and alternative mirrors (simple HTML lists, wget one‑liners), and note the original site sometimes behaves oddly.
  • One person printed the combined PDFs as a spiral‑bound volume (~360 pages, including one extra paper).
  • Another created a public reference manager group and reports basic stats: items span from the early 1990s to 2020, and include a mix of papers, preprints, a course, a dissertation, a book, and blog posts.
  • Some ask how to filter out AI/LLM content from the site entirely; suggestions include browser extensions and ML‑based classifiers, with the aside that simple regex filters might cover most cases.