Holography says a universe with gravity can be described completely by a theory living on its edge. Here are 15 physicists placed inside that picture: the founders deep in the bulk, the people who took the ideas into AI out at the boundary. Select a name to see the paper to start with, how that person connects to the others, what each connection rests on, and a talk or playlist to watch.
Height is the year of one landmark work per person, from 1915 to 2021; the ticks on the left are years. The year can be checked. Which work counts as the landmark is a choice, and the scale is stretched: 1997 to 2021 takes more than half the height.
Depth is a reading
How far in or out someone sits is an editorial judgment, not a measurement, and it comes in three bands only. Foundational figures sit deep in the bulk, where gravity lives; holography’s core sits at middle depth; people who crossed from physics into AI sit near the wall, the boundary, where the theory becomes the product.
Distance means nothing
Where a name sits around the cylinder was chosen so labels stay readable. Two people close together are not closer in any sense, and a long curve is not a weaker link.
The curves
Each connection is a hyperbolic geodesic: seen from above, an arc that meets the boundary circle at a right angle. The geometry is exact; using it for people is an analogy.
What a line rests on
Every connection says whether it is documented, found in the text of a paper, or only a shared technique. Only the first means two people were in contact.
Fainter names
Bridge figures who are not among the fifteen.
15 people, 49 sourced connections
The same people and connections as the diagram, in era order.
Albert Einstein
1915 · General relativity · deep in the bulk
General relativity, and both halves of ER = EPR: the 1935 wormhole paper and the 1935 entanglement paper.
Connects through the 1939 letter to Roosevelt: the letter Einstein signed opens by citing recent work on uranium by Fermi and Szilard. It is a citation, not joint work.
Connects through general relativity: Hawking and Penrose’s 1970 theorem shows that, under broad conditions, Einstein’s theory leads to spacetime singularities.
Connects through ER = EPR: Maldacena and Susskind proposed that Einstein’s two 1935 ideas, the Einstein–Rosen bridge and the EPR pair, describe the same thing.
Connects through the 1939 letter to Roosevelt: the letter Einstein signed opens by citing recent work on uranium by Fermi and Szilard. It is a citation, not joint work.
Both worked at wartime Los Alamos: Feynman led a group in the Theoretical Division from 1943, and Fermi moved there in September 1944 to head F Division.
Connects through a unit: Hofstadter is credited with naming the femtometre, the length scale of the nucleus, the fermi in his 1956 review of electron scattering.
Connects through a method, not a meeting: in a January 2020 lecture Kaplan sizes the brain at a petaflop from neurons, synapses and firing rate, then compares a lifetime with a training run. He never says Fermi; the name is this site’s.
Both worked at wartime Los Alamos: Feynman led a group in the Theoretical Division from 1943, and Fermi moved there in September 1944 to head F Division.
Connects through the path integral: Hartle and Hawking’s Wave Function of the Universe gives the ground state of a closed universe as a path integral over geometries, Feynman’s method applied to spacetime.
Connects by contrast: The Amplituhedron starts from the observation that scattering amplitudes have a simplicity that Feynman diagrams hide, and looks for a formulation without them.
Connects through method: The Principles of Deep Learning Theory treats neural networks with the perturbative field-theory toolkit, and its first and last chapters open with epigraphs about Feynman. The book itself draws no Feynman diagrams.
Rests on: shared techniqueSource: arXiv:2106.10165Outside source
Connects through a unit: Hofstadter is credited with naming the femtometre, the length scale of the nucleus, the fermi in his 1956 review of electron scattering.
Connects through general relativity: Hawking and Penrose’s 1970 theorem shows that, under broad conditions, Einstein’s theory leads to spacetime singularities.
Connects through the path integral: Hartle and Hawking’s Wave Function of the Universe gives the ground state of a closed universe as a path integral over geometries, Feynman’s method applied to spacetime.
Connects through the black hole information problem: Hawking argued that an evaporating black hole destroys information, and Susskind’s black hole complementarity starts from the postulate that quantum theory stays valid.
Connects through Hawking’s 2005 paper Information Loss in Black Holes, which argues information is preserved and finds the question well posed only in asymptotically AdS spacetimes, the setting of Maldacena’s correspondence.
Connects through a citation: Soft Hair on Black Holes, by Hawking, Perry and Strominger, cites Pasterski’s papers on gravitational and electromagnetic memory.
Connects through the black hole information problem: Hawking argued that an evaporating black hole destroys information, and Susskind’s black hole complementarity starts from the postulate that quantum theory stays valid.
Connects through the word in the title: Witten’s Anti De Sitter Space and Holography cites Susskind’s The World as a Hologram and gives AdS/CFT a precise holographic dictionary.
Coauthors of Complexity Equals Action, which conjectures that the quantum complexity of a holographic state is dual to the action of a region of spacetime.
Both at the Stanford Institute for Theoretical Physics in 2015: McCandlish was a graduate student there, and the acknowledgments of his paper Tensor Networks from Kinematic Space thank Susskind for discussions.
Connects through the dictionary: Witten’s Anti de Sitter Space and Holography took Maldacena’s conjecture and stated precisely how boundary correlation functions are computed from the bulk.
Connects through Hawking’s 2005 paper Information Loss in Black Holes, which argues information is preserved and finds the question well posed only in asymptotically AdS spacetimes, the setting of Maldacena’s correspondence.
Connects through Kaplan’s thesis topic: Aspects of Holography works inside the AdS/CFT correspondence Maldacena proposed in 1997. No joint paper is claimed.
Faculty colleagues in the School of Natural Sciences at the Institute for Advanced Study; the Institute announced their Fundamental Physics Prizes together in 2012.
Connects through ER = EPR: Maldacena and Susskind proposed that Einstein’s two 1935 ideas, the Einstein–Rosen bridge and the EPR pair, describe the same thing.
Connects through the holographic principle: Pasterski’s Snowmass review cites Maldacena’s 1997 paper as a concrete realization and sets out to find the counterpart for asymptotically flat space.
Connects through the dictionary: Witten’s Anti de Sitter Space and Holography took Maldacena’s conjecture and stated precisely how boundary correlation functions are computed from the bulk.
Connects through the word in the title: Witten’s Anti De Sitter Space and Holography cites Susskind’s The World as a Hologram and gives AdS/CFT a precise holographic dictionary.
Connects through twistor space: Arkani-Hamed’s 2009 paper The S-Matrix in Twistor Space (with Cachazo, Cheung and Kaplan) says its construction was inspired by Witten’s 2003 twistor string theory.
Faculty colleagues in the School of Natural Sciences at the Institute for Advanced Study; the Institute announced their Fundamental Physics Prizes together in 2012.
Faculty colleagues at the Institute for Advanced Study, where Arkani-Hamed’s appointment began in 2008; both were among its four 2012 Fundamental Physics Prize winners.
Connects through an endorsement: Witten’s note for The Principles of Deep Learning Theory says it is very interesting, for a physicist, to see deep learning approached from statistical physics.
Connects through twistor space: Arkani-Hamed’s 2009 paper The S-Matrix in Twistor Space (with Cachazo, Cheung and Kaplan) says its construction was inspired by Witten’s 2003 twistor string theory.
Faculty colleagues at the Institute for Advanced Study, where Arkani-Hamed’s appointment began in 2008; both were among its four 2012 Fundamental Physics Prize winners.
Connects by contrast: The Amplituhedron starts from the observation that scattering amplitudes have a simplicity that Feynman diagrams hide, and looks for a formulation without them.
Connects through celestial amplitudes: Arkani-Hamed coauthored Celestial Amplitudes from UV to IR with Pate, Raclariu and Strominger, and Pasterski’s 2021 review cites it. No joint paper is claimed.
Adam Brown, Daniel A. Roberts, Leonard Susskind, Brian Swingle and Ying Zhao · 2015 · arXiv:1509.07876
The conjecture that the complexity of a holographic state equals the action of a spacetime region, and the suggestion that black holes are the fastest computers nature allows.
Coauthors, with Susskind, Swingle and Zhao, of Complexity Equals Action, published in Physical Review Letters as Holographic Complexity Equals Bulk Action?
Both gave the Stanford Institute for Theoretical Physics as their affiliation on papers posted ten weeks apart in 2015: Complexity Equals Action, and Tensor Networks from Kinematic Space. No joint paper is claimed.
Dyer joined Google’s Blueshift team in 2018, the group Adam Brown leads; both are listed as organizers of the Aspen winter conferences on theoretical physics for machine learning in 2023 and for AI in 2026.
Connects through a citation: Soft Hair on Black Holes, by Hawking, Perry and Strominger, cites Pasterski’s papers on gravitational and electromagnetic memory.
Connects through celestial amplitudes: Arkani-Hamed coauthored Celestial Amplitudes from UV to IR with Pate, Raclariu and Strominger, and Pasterski’s 2021 review cites it. No joint paper is claimed.
Connects through the holographic principle: Pasterski’s Snowmass review cites Maldacena’s 1997 paper as a concrete realization and sets out to find the counterpart for asymptotically flat space.
Kaplan, by then at Anthropic, gave a virtual talk at Perimeter’s Theory + AI Symposium in April 2025; Pasterski was one of its three scientific organizers.
Connects through Kaplan’s thesis topic: Aspects of Holography works inside the AdS/CFT correspondence Maldacena proposed in 1997. No joint paper is claimed.
Coauthors first on An Empirical Model of Large-Batch Training in 2018, then equal-contribution lead authors of Scaling Laws for Neural Language Models in 2020, both at OpenAI.
Connects through a method, not a meeting: in a January 2020 lecture Kaplan sizes the brain at a petaflop from neurons, synapses and firing rate, then compares a lifetime with a training run. He never says Fermi; the name is this site’s.
Connects through a citation: a footnote in the epilogue of Roberts and Yaida’s book cites the scaling-laws paper and guesses that representation group flow may explain it. In 2022 Roberts, Maloney and Sully published a solvable model of those laws.
Kaplan, by then at Anthropic, gave a virtual talk at Perimeter’s Theory + AI Symposium in April 2025; Pasterski was one of its three scientific organizers.
Both at the Stanford Institute for Theoretical Physics in 2015: McCandlish was a graduate student there, and the acknowledgments of his paper Tensor Networks from Kinematic Space thank Susskind for discussions.
Both gave the Stanford Institute for Theoretical Physics as their affiliation on papers posted ten weeks apart in 2015: Complexity Equals Action, and Tensor Networks from Kinematic Space. No joint paper is claimed.
Coauthors first on An Empirical Model of Large-Batch Training in 2018, then equal-contribution lead authors of Scaling Laws for Neural Language Models in 2020, both at OpenAI.
Connects through a model of his measurement: A Solvable Model of Neural Scaling Laws, by Maloney, Roberts and Sully, sets out to reproduce the laws McCandlish co-led. Sully, by then at Anthropic, was McCandlish’s coauthor on the kinematic-space papers.
Coauthors, with Susskind, Swingle and Zhao, of Complexity Equals Action, published in Physical Review Letters as Holographic Complexity Equals Bulk Action?
Coauthors of Complexity Equals Action, which conjectures that the quantum complexity of a holographic state is dual to the action of a region of spacetime.
Connects through a citation: a footnote in the epilogue of Roberts and Yaida’s book cites the scaling-laws paper and guesses that representation group flow may explain it. In 2022 Roberts, Maloney and Sully published a solvable model of those laws.
Connects through method: The Principles of Deep Learning Theory treats neural networks with the perturbative field-theory toolkit, and its first and last chapters open with epigraphs about Feynman. The book itself draws no Feynman diagrams.
Rests on: shared techniqueSource: arXiv:2106.10165Outside source
Connects through a model of his measurement: A Solvable Model of Neural Scaling Laws, by Maloney, Roberts and Sully, sets out to reproduce the laws McCandlish co-led. Sully, by then at Anthropic, was McCandlish’s coauthor on the kinematic-space papers.
Connects through an endorsement: Witten’s note for The Principles of Deep Learning Theory says it is very interesting, for a physicist, to see deep learning approached from statistical physics.
Dyer joined Google’s Blueshift team in 2018, the group Adam Brown leads; both are listed as organizers of the Aspen winter conferences on theoretical physics for machine learning in 2023 and for AI in 2026.
Every fact here is on the public record. The arrangement is the contribution, and it shows one thing that is hard to see from separate biographies. In January 2020 two holography PhDs at OpenAI, Jared Kaplan and Sam McCandlish, measured how language models improve with scale. Within two years that measurement had drawn two theoretical accounts from former holographers at other organisations: one written with Kaplan by Ethan Dyer and colleagues at Google, and one by Daniel Roberts with Alexander Maloney and James Sully, who had been McCandlish’s coauthor in holography. Follow the lines out of Kaplan and McCandlish and you can watch it happen. The page on Kaplan and Roberts lays out the evidence.
The map also shows how thin the join between two groups is. Among the 15, the people who crossed into AI are the Anthropic founders on one side, and on the other Adam Brown and Roberts, who wrote Complexity Equals Action with Susskind. No byline joins one side to the other. What joins them is a building and a literature. In 2015 McCandlish was a graduate student at the Stanford Institute for Theoretical Physics while Adam Brown and Susskind, at the same institute, were writing that paper. Since 2021 the second group’s theory has cited the first group’s measurement. Dyer, a bridge figure, is the only person here who has published with both sides.
Three things on the map cut against its own premise, and they are drawn in anyway. Dario Amodei is a physics PhD at the boundary with no line to the middle of the diagram: his physics was the statistical physics of neurons, not holography. Adam Brown’s crossing is documented as a job and not as a paper. And the line from Fermi to Kaplan is a shared technique, the weakest kind of connection there is; it has a page explaining exactly what it rests on. The map does not show that holography caused anything in AI. It shows who was trained where, who wrote what with whom, and which papers answer which.
How to read the diagram
The shape is anti-de Sitter space drawn the usual way: a cylinder. Each horizontal slice is a Poincaré disk, a map of hyperbolic space in which the edge is infinitely far away, and the slices are stacked along a vertical time axis. The wall of the cylinder is the boundary, where the conformal field theory lives. The inside is the bulk, where gravity lives. In the AdS/CFT correspondence the two are the same physics written twice.
Depth is a judgment, not a measurement. Height on the cylinder is a year you can check; depth is my reading, in three bands. Einstein, Fermi and Feynman sit deep in the bulk because the rest is built on them. Maldacena, Witten, Arkani-Hamed and Pasterski sit at middle depth: holography’s core. Kaplan, McCandlish, Amodei, Adam Brown and Roberts sit near the boundary, because the boundary is where the theory becomes the product.
Height is a year, on a stretched scale. Each person sits at the year of one landmark work, from general relativity in 1915 to The Principles of Deep Learning Theory in 2021. The year is checkable; which work counts as the landmark is my choice. Roberts sits at his 2021 book and not at his 2015 physics. The axis is also uneven on purpose: the 24 years from 1997 to 2021 take more than half the height, so that the 82 years before them do not swallow the diagram.
Distance is not a measurement. The angle at which a person sits around the cylinder was chosen to keep names from colliding. Two people near each other are not closer in any sense, and a long curve is not a weaker connection than a short one.
The curves are real geodesics. Seen from above, every connection is the shortest path between two points of the hyperbolic disk: an arc of a circle that meets the boundary at a right angle. The faint pattern on the active slice is the {7,3} tiling, seven-sided tiles meeting three to a corner, all the same size in hyperbolic terms. The geometry is exact. Applying it to people is my visual analogy, and nothing in the physics says a career has a geodesic.
Five kinds of connection
advisorPhD advisor and student.
coauthorNames on the same paper.
same placeSame place at the same time: an institute, a lab, a lecture room.
idea lineageOne body of work builds on another; an interpretation, not a collaboration.
crossed into AIThe link runs through an AI lab or an AI event.
Three things a line can rest on
The type of a connection says what kind of relationship it is. Its basis says how much weight it can bear, and every row names it.
documentedA byline, thesis record, staff page or programme puts the two people together. 27 of 49.
in the textOne person’s published words cite, name or answer the other’s. The citation can be checked; how much it mattered is a reading. 19 of 49.
shared techniqueThe later work uses a method the earlier person is known for. No contact and no citation is claimed; the pairing is this site’s. 3 of 49.
analogyThe cylinder itself: depth, angle and the curved lines. None of the 49 connections rests on it, and none should.
The paper trail
Each person’s entry opens with one paper to start with and up to two more: 48 in all. They were chosen to explain why that person is on this map, which is not always their most cited work. Titles, bylines and years were checked against the arXiv listing or the journal’s DOI record on November 15, 2026; work with no preprint links to its DOI or to an institutional record. The sentence under each one is my description of what to look for. Pages from this site are listed separately under each person as coverage, because my reporting is not evidence for my own map.
Three close readings
Three of the people at the boundary get a longer look. Each reading separates what the record shows from what I make of it, and ends with a question only that person can answer.
Daniel Roberts: the footnote and the follow-up
What a profile would miss
The cover of The Principles of Deep Learning Theory credits three people in two ways: it is by Daniel A. Roberts and Sho Yaida, based on research in collaboration with Boris Hanin. Its central technical choice gets lost when the book is summarised as physics for neural networks. It does not stop at infinite width, where the mathematics is exact and, as the book shows, the network learns no representations. It expands in a network’s depth divided by its width, and that one ratio controls the whole theory. The authors call the layer-to-layer flow representation group flow and apologise for the name in a footnote: it is really a semigroup, which they say repeats a historic mistake in physics.
What the physics and the AI work share
Before the book, Roberts worked on quantum chaos. With Pavan Hosur, Xiao-Liang Qi and Beni Yoshida he defined scrambling in terms of information; with Adam Brown, Leonard Susskind, Brian Swingle and Ying Zhao he wrote Complexity Equals Action. The book uses the working method of that physics: start from a Gaussian theory that can be solved, add interactions as small corrections, and follow how the couplings change with scale, which here means with depth.
Where the comparison stops
The worked theory is for multilayer perceptrons. The preface says that was a teaching choice and not a technical limit, but the scaling laws the epilogue points to were measured on transformers. And when Roberts did model those laws in 2022, with Alexander Maloney and James Sully, the solution came from a random-feature model and random matrix theory, not from the representation group flow the footnote had guessed at. That paper’s outlook describes how the book’s tools might be brought back in.
Which part is my reading
The book mentions Feynman in epigraphs, not in equations, so the line from Feynman to Roberts is a shared technique and nothing more. Placing him at 2021, the book, instead of 2015, the physics, is also a choice made here.
A question for Daniel Roberts
Your epilogue footnote guessed that representation group flow might explain the Kaplan and McCandlish scaling laws. The 2022 model with Maloney and Sully used random matrix theory. Did you try the flow first, and is that explanation still open?
On arXiv the title is a statement: Complexity Equals Action. In Physical Review Letters the same five-author paper is a question: Holographic Complexity Equals Bulk Action? Its abstract says the conjecture replaced an earlier one, that complexity is dual to volume, and raises the hypothesis that black holes are the fastest computers in nature. Three years earlier, writing alone, he took a famous thought experiment about lowering boxes toward a black hole and asked what the rope is made of. His answer in Tensile Strength and the Mining of Black Holes: the rope sets the limit, and the boxes are superfluous.
What the physics and the AI work share
Both papers put an engineer’s question to a physicist’s object: what is the most any device could do? His Aspen Center for Physics biography says he leads Blueshift, a Google DeepMind team working on the scientific and reasoning abilities of AI, and that he joined Google in 2018. In a 2024 interview he said he uses his own graduate general relativity exam as a private benchmark, and that models went from scoring zero on it to nearly acing it in about three years.
Where the comparison stops
This is where the premise of the atlas is weakest. Quantum complexity counts the gates needed to prepare a quantum state. It is not the compute used to train a model, and no paper found for this page connects the two. His crossing into AI is documented as a job, not as a result: no machine-learning paper is listed on his node because none could be verified under his name. In the same interview he said he had not yet seen a model do anything like inventing general relativity.
Which part is my reading
Reading the fastest computers in nature next to an AI lab is a juxtaposition made here. Nothing found shows him drawing that line himself.
A question for Adam Brown
Aspen’s biography says you joined Google in 2018, the year Ethan Dyer joined Blueshift. Did Blueshift exist before you arrived, or did you start it, and in what year? The atlas says you lead it and cannot yet say since when.
Scaling Laws for Neural Language Models marks Kaplan and McCandlish as equal contributors, and its contributions note says the two of them led the research. Calling him a coauthor undersells that. Thirteen months earlier, An Empirical Model of Large-Batch Training lists him first, with a footnote: work done as an OpenAI Fellow. It defines one statistic, the gradient noise scale, and shows that it predicts the largest useful batch size for image classifiers, a language model, Atari and the Dota 2 agents. An appendix relates it to a temperature of training.
What the physics and the AI work share
His physics was also about reading an interior from outside. With Bartlomiej Czech, Lampros Lamprou and James Sully he worked on kinematic space: arrange a boundary theory’s intervals into a geometry of their own, and the length of any curve in the bulk can be computed from it.
Where the comparison stops
No paper of his carries holography into machine learning, and the scaling-laws paper cites none. The thesis, Depth Perception in Holography, and the batch-size paper share an author and a habit of measurement. They do not share a method.
Where his work argues with this page
His most revealing physics paper argues against a picture like this one. Tensor Networks from Kinematic Space takes a tensor network that physicists had read as a drawing of curved space and argues it is best viewed as a discretization of the space of bulk geodesics rather than of the bulk geometry itself. That is a warning about mistaking a diagram that looks hyperbolic for the space. This atlas draws people inside a hyperbolic diagram. Its curves are true geodesics of the picture. They are not a geometry of the people.
Which part is my reading
The common thread, finding the one simple quantity a messy system obeys, is a reading made here. He has not been found saying it.
A question for Sam McCandlish
The 2018 large-batch paper lists you as an OpenAI Fellow and Kaplan at Johns Hopkins and OpenAI. Did you two know each other from physics before OpenAI, or meet there? The atlas dates your connection from that paper because it has nothing earlier.
There are 49 connections. 16 are backed by a page on this site, mostly the FounderFiles on Jared Kaplan, Juan Maldacena, Sam McCandlish, Sabrina Pasterski and Ethan Dyer. The other 33 rest on outside sources: arXiv and INSPIRE-HEP records, institute pages and U.S. Department of Energy history pages. Every row links to its source.
Connections marked idea lineage are interpretation. They say one body of work builds on another, not that two people worked together. Where a link could not be sourced it was left out, and a person with two solid connections was left with two. Current roles for Adam Brown and Daniel Roberts were checked on November 12, 2026, and Sam McCandlish’s on November 15. The Adam Brown here is the physicist who leads Blueshift at Google DeepMind. He is not the GPT-3 author with the same surname who has a FounderFile on this site.
Ethan Dyer and Leo Szilard appear as bridge figures. They are not among the 15, but each joins two people who are. For the longer story of why holographers ended up at AI labs, read The Data Desert: Physics, AI, and the Diaspora.