The Missing Elements

§01The email that vanished

Sometime around 2020, John Carmack, the programmer behind Doom and Quake who was by then turning his attention from virtual reality to artificial general intelligence, asked Ilya Sutskever what he should read. Sutskever, then OpenAI’s chief scientist, handed him roughly 40 research papers and a promise that Carmack repeated to Dallas Innovates in February 2023: “If you really learn all of these, you’ll know 90% of what matters today.” Carmack said he plowed through them and the field began sorting itself out in his head.

Then the list disappeared. Carmack later said the email was lost to Meta’s two-year auto-delete policy, and he asked publicly on X for Sutskever to post a canonical version. Sutskever never did. What circulates now, on punkx.org, LessWrong, Turing Post, a dozen GitHub mirrors and the PaperTrace page that prompted this dispatch, descends from an arc.net folder shared in May 2024 by someone Turing Post describes as an OpenAI researcher. It holds 27 items. Nobody has credibly reconstructed the other dozen or so.

So the honest framing comes first. This is a list attributed to Sutskever, unconfirmed by him, and shorter than the one Carmack remembers. PaperTrace’s “top 30” version drops The Annotated Transformer and adds four items: Multi-Token Prediction (2024), Zephyr (2023), and RAG and Dense Passage Retrieval (both 2020). At least two of those postdate any plausible handoff to Carmack. They are community modernizations, not his picks.

The 27 are still worth reading as a document, for what they cluster around and, more usefully, for what they skip.

§02The thesis

The list is not a catalog of architectures. It is a theory of compression plus the handful of operators that theory needs, and it stops in 2020, the year Sutskever now says the age of scaling began. Lay it out like a periodic table and the most informative cells are the empty ones: value functions, continual learning, alignment. Those are the subjects his November 2025 conversation with Dwarkesh Patel puts at the center of what Safe Superintelligence Inc. is trying to solve.

§03What the list is actually about

Read the 27 for architectures and you get a tidy curriculum. AlexNet and Stanford’s CS231n cover convolution. ResNet and its identity-mappings follow-up cover depth. Bahdanau attention, Pointer Networks and the Transformer cover routing. LSTMs, Neural Turing Machines and relational recurrent networks cover memory. Relation networks and neural message passing cover reasoning over entities. GPipe, Deep Speech 2 and the Kaplan scaling laws cover scale.

The odd items are the tell. Why would a deep-learning syllabus include Scott Aaronson’s blog post on why the interestingness of a stirred cup of coffee rises and then falls while its entropy only climbs, plus the Coffee Automaton paper that formalizes it? Why a tutorial on Minimum Description Length, a Kolmogorov complexity textbook assigned from page 434 onward, Shane Legg’s thesis on universal intelligence, and Geoffrey Hinton and Drew van Camp’s 1993 paper on keeping networks simple by minimizing the description length of their weights?

Because they are one argument. Sort the 27 into functional columns (our classification, not Sutskever’s) and 7 of 27 land in a single bucket: compression, MDL and regularization. Hinton and van Camp’s bits-back idea runs straight into the 2016 Variational Lossy Autoencoder, which Sutskever co-authored. The RNN regularization paper he wrote with Wojciech Zaremba and Oriol Vinyals is a generalization fix. The Kaplan scaling laws belong here in spirit too: a language model’s loss, measured in nats, is literally a code length, so a scaling law describes how compression improves with resources.

Sutskever has made this argument in public. At the Simons Institute in Berkeley in August 2023, in a talk titled An Observation on Generalization, he argued that compressors and predictors correspond one to one, treated conditional Kolmogorov complexity as the ideal unsupervised learner, and described networks trained with stochastic gradient descent as small, approximate versions of that ideal. The theory bloc of the reading list is that talk in bibliography form.

§04Why the periodic-table metaphor earns its keep

In 1869 Dmitri Mendeleev published a table of the elements with gaps in it. He did not paper over them. He predicted what would fill them, eka-aluminium and eka-silicon among them, and gallium (1875) and germanium (1886) arrived roughly where he said they would. The value of his table was its blanks.

Arrange the Sutskever list the same way, with functional groups across the top (represent, route, remember, relate, compress, scale, evaluate, adapt) and levels of abstraction down the side. The six left-hand columns fill quickly. Represent holds five items. Route and Remember hold four each. Relate and Scale hold three each. Compress holds seven.

Then the right edge. Evaluate, the machinery of judging an action before its outcome arrives, holds exactly 1 of 27: Legg’s thesis, which defines intelligence over reward-seeking agents. There is no DQN, no AlphaGo, no PPO. Adapt, learning continually after deployment, holds nothing. Self-play, debate and alignment have no cell at all.

This is analyst synthesis and should be labeled that way. Sutskever has never called the list a periodic table or described any set of irreducible primitives. His stated aesthetic cuts the other way. In the Dwarkesh interview he describes research taste as a search for beauty, simplicity, elegance and correct inspiration from the brain, and says human learning points to a single underlying machine-learning principle he has opinions on but won’t discuss. Attention arguably absorbs half the left side of the table anyway: self-attention can emulate convolution, performs message passing on a fully connected graph, and works as content-addressable memory. Several of these elements may be isotopes of one.

As a predictive instrument, though, the table does what Mendeleev’s did. The blanks show where the next discovery is supposed to land.

§05What he said about the blanks

The November 25, 2025 episode, roughly 96 minutes long and still his most recent long-form interview, reads like an annotation of the missing columns.

On Evaluate, he described a man whose brain damage wiped out his emotional processing. The man still passed tests but became nearly unable to make decisions. Asked for the machine-learning analogue of emotion, Sutskever reached for the value function, which lets an agent know partway through a task that a path is bad instead of waiting for the final grade, and suggested that human value functions are modulated by emotions in ways evolution hardcoded. Then he made an MDL-shaped argument: emotions are simple, and simple things stay useful across a very broad range of situations. The compression column, extended into the empty one.

On Adapt, he argued that a human being is not an AGI in the pre-training sense. A person lacks most knowledge and relies on continual learning instead. His picture of a successful SSI system was a brilliant, eager teenager who learns each job after deployment: a process, not a finished artifact.

On the meta-question he was blunter. Value functions, he said, will make reinforcement learning more efficient, but anything you can do with one you can do without it, only more slowly. The fundamental problem is that models generalize far worse than people. Generalization is the oldest theme on a list that runs from 1993 to 2020, and it is still the first one.

§06The correction the premise needs

The question that sent us to the list assumed SSI is chasing recursive self-improvement and trying to give language models emotions. The transcript supports neither.

On recursive self-improvement, Sutskever went out of his way to separate SSI from it. He described self-improving AI as the single idea the whole field has fixated on, a product of there being more companies than ideas, and said there is something better to build: an AI robustly aligned to care about sentient life. On emotions, he used them as an analogy for value functions, mused that they might be simple enough to map in a human-readable way, and stopped. Nothing he said describes SSI engineering machine emotion.

What he said about alignment lands in the table’s last empty space. He framed the difficulty of learning human values, and then of optimizing them, as instances of unreliable generalization. And he admitted he has no good hypothesis for how evolution manages to hardcode high-level social desires into a genome. That is the open problem stated plainly: the column the list never filled is the one SSI says it exists to fill.

§07Who holds the leverage

SSI sits on the strangest power map in frontier AI: maximum capital, minimum disclosure. The last priced round, led by Greenoaks in April 2025, valued the company at $32B. Sutskever confirmed that figure on the podcast, explaining that Meta offered to acquire SSI during the fundraise, that he said no, and that co-founder Daniel Gross took Meta’s offer instead. There is still no product, no paper and no model.

On July 27, 2026, NVIDIA announced a long-term partnership and an undisclosed investment that gives SSI access to Vera Rubin systems and, by SSI’s own account, ten times its compute within 12 months. Sutskever’s framing was that SSI now has research it considers worth scaling up. Reuters and Bloomberg, citing people familiar with the matter, put NVIDIA’s check at $5 billion; neither company has confirmed it. Headcount is unofficial too: about 20 in March 2025, and around 50 in most 2026 reports.

That framing matters for reading the list. The headline our own FounderFiles N°005 ran, that the age of scaling is dead, overstates him. His actual position is that the field is back in an age of research, only with big computers. A lab that says it has found something worth scaling is not against scale. It is against the recipe.

§08What can’t be checked from outside

Several parts of this story cannot be verified, and should be read that way. Sutskever has never confirmed the list or its date. The missing papers Carmack remembers are unknown. SSI’s technical approach is undisclosed; Sutskever said on the podcast that not all machine-learning ideas are discussed freely, and that this one isn’t. The Israeli outlet Globes reported in August 2026, citing unnamed experts, that SSI’s model may update its weights while it reads. That is speculation. A widely shared rumor of an August 2026 launch did not pan out, and as of late October SSI had released nothing.

Two corrections to our own record belong here as well. The Atlantic places the Feel the AGI chant at OpenAI’s 2022 holiday party, not at the Tenaya Lodge offsite, and accounts of the effigy burning differ on its framing and rest on anonymous sources. The 52-page memo, by contrast, is now confirmed under oath.

§09The white space

Here is the gap worth watching. Every frontier lab has spent five years filling the left side of the table: more operators, bigger compositions, steeper power laws. The right side, where an agent judges its own trajectory and keeps learning after it ships, is still nearly empty in the public literature, and empty on the list.

If Sutskever’s theory holds, the next element is not another architecture block. It is a learning principle that compresses experience well enough to generalize from a teenager’s worth of data, with a simple, robust value signal riding on top, aligned to care for sentient life. Mendeleev could describe eka-silicon’s density before anyone held germanium. Sutskever has described the shape of his missing element in more detail than he has described anything SSI has built.

The list told Carmack what mattered in 2020. Its blank columns tell the rest of us what Sutskever thinks will matter next.