Ilya Sutskever's 30-Paper Reading List Is Making the Rounds Again

Most AI conversations right now orbit a single question: which model is smarter? But lately something quieter has been circulating in developer circles. Not a product. Not a benchmark. A reading list. Specifically, a set of 30 papers that OpenAI co-founder Ilya Sutskever — one of the people closest to the frontier of deep learning — is said to have handed off years ago. No flashy launch, just a list of PDFs. So why is it back?

“Read these 30 and you’ll understand 90% of what matters”

The story goes like this. Someone asked Ilya what they should read to really understand AI. He reportedly passed over a list of papers and said: read all of these, and you’ll know 90% of what matters in AI today.

That one line has stuck around for a reason. One of the deepest thinkers in deep learning didn’t say “keep up with the trends.” He pointed at 30 foundational papers — many of them old.

Worth flagging: this isn’t an official document Ilya published or curated for the public. It reads more like a private recommendation to an acquaintance that leaked online. So there’s genuine back-and-forth over whether it’s exactly 30 papers, or whether it’s truly his list at all. But the papers on it are undisputed classics, which is why it keeps getting shared regardless of the provenance debate.

What’s actually on the list

Skim it and you’ll hit name after name that anyone who has studied deep learning will recognize. A few highlights.

Attention Is All You Need — the paper that introduced the Transformer, the skeleton underneath nearly every chatbot you use today. The “T” in GPT is that Transformer.

AlexNet — the 2012 paper that crushed an image-recognition competition and burned “deep learning actually works” into the field’s collective memory. It’s widely treated as the starting gun for the current AI boom.

ResNet — deep networks used to get worse the deeper you stacked them. This paper fixed that with a deceptively simple idea: a shortcut, the skip connection.

There’s more in there — recurrent networks, information theory, computational complexity, and other classics. The telling part: this is not a collection of recent papers. A lot of them are years old, some more than a decade. It’s a list of foundations, not fashions.

Why point at the basics?

Here’s the thing worth sitting with. AI shifts every few months. New models, new techniques, every single week. So why would a frontier researcher tell you to read old papers?

Because the surface keeps changing, but the underlying principles mostly don’t. Even the slickest features in the newest model still stand on old pillars: the Transformer, gradient descent, backpropagation. Understand the pillars and each new release parses instantly — “ah, this is a variation on that idea.” Skip them and you’re stuck being surprised from scratch every time something ships.

That’s the same reason this list keeps getting summoned in community threads. When someone asks “what should I study right now,” it’s an answer with a long shelf life. Tools go stale. Principles don’t.

Reading it yourself, in an age when AI reads for you

There’s an irony here. You can hand a paper to an AI and get the gist in seconds. So why grind through 30 of them by hand?

Because reading a summary and wrestling with the original leave you with different things. A summary gives you the conclusion. The original shows you the reasoning that led there — why they chose this approach, what failed, where they got stuck. Real fluency usually accrues in that struggle, not in the takeaway.

And the more we hand judgment to AI, the more valuable it becomes to hold onto judgment of your own. To tell whether an AI’s answer is right or wrong, you need a standard inside your own head. The surest way to build that standard is to walk the fundamentals yourself.

The takeaway

The message behind Ilya’s 30 papers is almost embarrassingly simple: before you chase the new thing, get a grip on the things that don’t change. Few people will read all 30. But the list is still a useful mirror — a prompt to ask whether you’re chasing trends or building a foundation. In an era where AI reads more and more on your behalf, which 30 papers do you still want to read yourself?