Jared Kaplan
Last reviewed
May 24, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 ยท 5,017 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 24, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 ยท 5,017 words
Add missing citations, update stale details, or suggest a clearer explanation.
Jared Daniel Kaplan is a theoretical physicist and artificial intelligence researcher, a co-founder and Chief Science Officer of anthropic, the AI safety company that develops the claude family of language models.[1][2] He is also an associate professor in the Department of Physics and Astronomy at Johns Hopkins University, where his academic research centered on quantum gravity, the AdS/CFT correspondence, and the conformal field theory bootstrap before he pivoted into AI.[1][3] Kaplan is the lead author of the 2020 paper "Scaling Laws for Neural Language Models," which established that transformer language model loss follows smooth power-law functions of model size, dataset size, and training compute, an empirical observation that guided OpenAI's decision to scale GPT-3 to 175 billion parameters and shaped frontier model development across the industry.[4][5] In October 2024 he assumed the additional role of Anthropic's Responsible Scaling Officer, taking charge of the company's pre-deployment safety evaluations under its Responsible Scaling Policy.[6]
| Field | Detail |
|---|---|
| Education | B.S. physics and mathematics, Stanford (2005); Ph.D. physics, Harvard (2009) |
| Doctoral advisor | Nima Arkani-Hamed |
| Doctoral thesis | "Aspects of Holography" |
| Academic post | Associate Professor, Johns Hopkins University Department of Physics and Astronomy (since 2012) |
| Industry roles | Researcher at OpenAI (2019 to 2020); co-founder and Chief Science Officer, Anthropic (2021 to present); Anthropic Responsible Scaling Officer (October 2024 to present) |
| Notable works | "Scaling Laws for Neural Language Models" (2020); "Language Models are Few-Shot Learners" (2020); "Constitutional AI: Harmlessness from AI Feedback" (2022) |
| Honors | Hertz Fellow (2005), NSF CAREER Award (2015), Sloan Research Fellowship (2014) |
Kaplan attended Stanford University as an undergraduate, completing a bachelor's degree in physics and mathematics in 2005.[7] During his senior year at Stanford he was named a Fannie and John Hertz Foundation Fellow, a competitive graduate fellowship that funds doctoral study in the applied physical, biological, and engineering sciences and counts numerous prominent scientists among its alumni.[7] The Hertz Foundation supported his doctoral studies at Harvard, where he entered the physics Ph.D. program and worked in theoretical particle physics with advisors Howard Georgi and Nima Arkani-Hamed, two of the most influential figures in modern high-energy theory.[1][8] His 2009 dissertation, titled "Aspects of Holography," addressed topics in the AdS/CFT correspondence and quantum gravity, including holographic descriptions of black holes in anti-de Sitter space and questions about whether the holographic principle extends naturally to asymptotically flat spacetimes.[8][9]
After completing his Ph.D., Kaplan held a postdoctoral fellowship jointly affiliated with the SLAC National Accelerator Laboratory and Stanford University from 2009 through 2012, working on questions at the intersection of effective field theory, the conformal bootstrap, and quantum gravity.[3][10] During the SLAC years he became a regular collaborator of A. Liam Fitzpatrick, with whom he would publish many of his most influential physics papers over the next decade.[10][17] He joined the faculty of Johns Hopkins University in 2012 as an assistant professor in the Department of Physics and Astronomy, was tenured as an associate professor a few years later, and has remained on the Hopkins faculty since while continuing his research collaborations.[1][3] At Hopkins his teaching has included graduate courses in quantum field theory, conformal field theory, and, in later years, the foundations of deep learning.[1]
For roughly the first fifteen years of his research career Kaplan worked exclusively as a theoretical physicist.[11] His academic publications cover effective field theory, particle physics, cosmology, scattering amplitudes, the conformal field theory bootstrap, AdS/CFT correspondence, and quantum gravity.[3][10] At Johns Hopkins he taught graduate-level quantum field theory; the lecture notes he prepared for that course, totalling several hundred pages and circulating informally as "QFT Lectures Notes," are widely used by graduate students as a free study reference and combine standard textbook material from Weinberg and Schwartz with less conventional pedagogical choices such as introducing effective field theory through ball-and-spring models in the first lecture.[12] The notes work through preliminaries on creation and annihilation operators, perturbation theory and scattering, simple interactions, the emergence of classical fields, locality, semiclassical methods, symmetries, electromagnetic fields, special relativity in quantum mechanics, relativistic spinless and spin-half particles, quantum electrodynamics, radiative corrections and renormalization, bound states, and path integrals for fields, with a second-semester treatment of Wilsonian renormalization, gauge symmetries, Nambu-Goldstone bosons, the Higgs mechanism, the Standard Model, and anomalies.[12] He also distributed lecture notes titled "Lectures on AdS/CFT from the Bottom Up," covering the conformal group, conformal partial wave expansions, and the analytic bootstrap, which serve as an accessible introduction to a technically demanding area of mathematical physics.[13]
A significant strand of Kaplan's physics output examined how the conformal bootstrap, an axiomatic approach to conformal field theory that uses crossing symmetry and unitarity to constrain operator dimensions and operator product expansion coefficients, intersects with the AdS/CFT correspondence. A 2012 paper with A. Liam Fitzpatrick, David Poland, and David Simmons-Duffin, "The Analytic Bootstrap and AdS Superhorizon Locality," demonstrated that every unitary conformal field theory above two dimensions containing a scalar operator must possess an infinite tower of operators whose twists approach specific accumulation points as their spin grows; this result connected the consistency of CFT four-point functions to the locality of bulk physics in anti-de Sitter space and is one of the founding works of the modern analytic-bootstrap program.[14] The paper has been cited more than a thousand times in the subsequent literature on conformal field theory.[14] A follow-up 2014 paper with Fitzpatrick and Matthew Walters, "Universality of Long-Distance AdS Physics from the CFT Bootstrap," extended the analysis to show that the leading interactions between widely separated objects in AdS gravity are universal consequences of the bootstrap, providing a CFT-side derivation of the long-distance behavior usually obtained from bulk effective field theory.[15]
Kaplan also co-authored a 2011 paper, "A Natural Language for AdS/CFT Correlators," with Fitzpatrick, Joao Penedones, Suvrat Raju, and Balt van Rees, which argued that Mellin space provides a particularly natural representation for holographic correlators because CFT correlators in Mellin space have poles whose residues encode the operator product expansion in a way that closely mirrors the factorization channels of bulk scattering amplitudes.[16] In the regime where correlators are computable by tree-level Witten diagrams in AdS, the authors derived explicit formulae for Mellin amplitudes and showed that they satisfy algebraic finite-difference equations, giving simple diagrammatic rules for constructing Mellin amplitudes in any bulk scalar theory.[16] A 2012 companion paper, "Unitarity and the Holographic S-Matrix," explored the analytic structure of these Mellin amplitudes and their relation to the bulk S-matrix.[17] Other physics papers addressed the eikonal limit of conformal blocks, pure quantum gravity in three-dimensional anti-de Sitter space (where the theory is exactly solvable), and the implications of modular invariance for two-dimensional CFTs at large central charge.[16][17]
In 2014 the Alfred P. Sloan Foundation named Kaplan a Sloan Research Fellow, an award given annually to about 126 early-career researchers in recognition of distinguished work and the potential to make substantial contributions to their fields; the fellowship carried a $50,000 two-year award.[18] In 2015 the National Science Foundation awarded him a CAREER grant (PHY-1454083) supporting research and teaching activities in theoretical particle physics; the CAREER program is the NSF's most prestigious award for junior faculty.[7] He is also a principal investigator within the Simons Foundation Collaboration on the Nonperturbative Bootstrap, a multi-institution program funding work on the conformal bootstrap that brings together researchers from leading physics departments and institutes.[3] Kaplan has given invited talks on quantum gravity and the bootstrap at venues including the Philosophical Society of Washington, the Princeton Institute for Advanced Study, and various Aspen and KITP physics workshops.[10]
By the late 2010s Kaplan had grown interested in the rapid empirical progress of deep learning. He began following work on transformer language models after the 2017 publication of "Attention Is All You Need" and the subsequent appearance of GPT-2 in 2019, and saw in the early results an empirical pattern that resembled the scaling regimes familiar from condensed-matter physics and critical phenomena.[11][32] In 2019 he started a research consulting engagement with OpenAI while retaining his Hopkins faculty position, traveling between San Francisco and Baltimore to participate in research meetings.[11] His arrival at OpenAI placed him in close working relationships with Dario Amodei, then OpenAI's vice president of research, and Sam McCandlish, then a research scientist; together they formed the core of what would become OpenAI's "scaling" research team.[19] Kaplan brought to the engagement a physicist's intuition for scaling phenomena: in many physical systems, dimensionless observables follow power-law behavior across many orders of magnitude when no other characteristic scale is present, and Kaplan's instinct was that neural network training loss, viewed as a function of compute and data, might exhibit similar regularities if measured carefully across a wide enough range.[11]
The collaborative work that emerged from this intuition appeared on arXiv on January 23, 2020, under the title "Scaling Laws for Neural Language Models," with Kaplan as first author followed by Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, alec radford, Jeffrey Wu, and dario amodei.[4] The paper reported that the cross-entropy test loss of decoder-only transformer language models is, to high accuracy, a smooth power-law function of three quantities considered separately: the number of non-embedding parameters N in the model, the size of the training dataset D in tokens, and the amount of compute C spent during training, with the trends spanning more than seven orders of magnitude.[4] Architectural details such as the ratio of network width to depth and the number of attention heads had only small effects within a wide range, suggesting that the scaling laws reflected a deep property of the loss landscape rather than artefacts of a particular hyperparameter choice.[4] Overfitting was shown to be governed by a simple ratio of model size to data size, and the dependence of training speed on model size was described by another simple functional form.[4]
The paper's most consequential prescriptive claim concerned compute-optimal training. Given a fixed compute budget, the authors argued that the optimal allocation pushed most of the budget into model size and trained the resulting very large model for a relatively short period on a modest number of tokens; in their words, larger models were "significantly more sample-efficient" and stopping training before convergence was preferable to using a smaller model that ran to convergence.[4] This conclusion directly motivated OpenAI's subsequent decision to train gpt-3 at 175 billion parameters, ten times larger than any prior dense language model.[20][21] The paper became one of the most influential works in modern machine learning and has accumulated more than 8,500 direct citations on Google Scholar by 2026, with countless more citations through derivative works.[22] In a 2024 retrospective interview with Y Combinator, Kaplan described the scaling-laws result as a "guidepost" that allowed researchers to predict in advance how much capability could be expected from a planned training run, transforming research planning from a guessing game into a budgeted exercise.[34]
The compute-optimal prescription was later partially revised. In 2022 DeepMind's "Training Compute-Optimal Large Language Models" paper introduced the chinchilla model and argued that Kaplan and colleagues had underestimated the importance of data: when more carefully tuned cosine learning-rate schedules were used and embedding parameters were counted, the optimal allocation balanced parameters and tokens roughly equally, at about twenty training tokens per parameter.[23] Subsequent work showed that the discrepancy arose mainly from Kaplan and coauthors having used models up to about one billion parameters with relatively short training horizons and learning-rate schedules tuned for that regime, rather than from a fundamental error in the scaling-law functional form.[23] The chinchilla scaling revision is now standard practice for frontier pretraining, but the qualitative finding of the Kaplan paper, that loss scales as predictable power laws in compute, data, and parameters, has been confirmed repeatedly and remains the central organizing principle of frontier model development.[23][34] The broader research field that grew out of these results is now referred to as the study of scaling laws for neural networks.[34]
While at OpenAI, Kaplan contributed to two further landmark papers. He is a co-author of "Language Models are Few-Shot Learners," posted on arXiv on May 28, 2020, which introduced gpt-3 and reported that scaling a transformer language model to 175 billion parameters produced strong in-context learning across a wide range of natural language tasks without parameter updates.[24] The paper's contributions list attributes the demonstration that larger models learn more quickly from in-context examples specifically to Kaplan and Sam McCandlish, an observation that linked the scaling-laws framework directly to the few-shot learning phenomenon that made GPT-3 commercially attractive.[24] The paper has accumulated more than 73,000 citations and is by some measures the most cited machine-learning paper of the early 2020s.[22]
Kaplan is also a co-author of the 2021 paper "Evaluating Large Language Models Trained on Code," which introduced openai codex (the model that initially powered GitHub Copilot) and the HumanEval benchmark for assessing functional correctness in code generation.[25] Codex was the first widely deployed application of GPT-style models to code, and HumanEval has since become the canonical entry-level coding benchmark cited by virtually every subsequent code-generation system.[25]
In late 2020 and early 2021 a group of senior OpenAI researchers and leaders, including dario amodei (then vice president of research), daniela amodei (then vice president of safety and policy), Kaplan, Sam McCandlish, Tom Brown, Chris Olah, Jack Clark, and Ben Mann, departed to found anthropic as a public benefit corporation focused on AI safety research.[26][27] Anthropic launched publicly in 2021 and raised an initial $124 million Series A round backed by investors including Jaan Tallinn, Dustin Moskovitz, and Eric Schmidt.[27] Kaplan took the title of Chief Science Officer.[1][2]
As Chief Science Officer, Kaplan oversees scientific direction across pretraining, alignment research, mechanistic interpretability, and policy-relevant evaluation work, sitting alongside CEO Dario Amodei, President Daniela Amodei, and CTO Sam McCandlish on Anthropic's executive team.[1][11] In interviews he has described the founding scientific bet of the company as a continuation of the scaling story: that AI capabilities will continue to improve predictably with compute and data, that this trajectory is likely to reach human-level performance within roughly a decade, and that responsible development requires preparing safety techniques and policies that scale with capability rather than waiting until capability arrives.[11][34] He projected to interviewer Dwarkesh Patel that an AGI-level training run might require on the order of 10^29 to 10^30 floating-point operations and that this scale could be reached by approximately 2030 given continued growth in compute investment.[28]
Kaplan is a co-author of several foundational Anthropic alignment papers. He is a listed author on "Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback" (April 2022), an early study using rlhf to train large language models toward helpfulness and harmlessness simultaneously, which introduced the now-standard "HH" preference framework used throughout the industry.[29] He is also a listed author on "Constitutional AI: Harmlessness from AI Feedback" (December 2022), which introduced the technique of training a model to critique and revise its own outputs against a written list of principles, a "constitution," and then using AI-generated preference labels rather than human ones in the reinforcement-learning stage; this method is now known as RLAIF (reinforcement learning from AI feedback).[30] constitutional ai became the public name for Anthropic's alignment methodology and was used in training successive generations of claude, including Claude 1, Claude 2, claude 3 opus, and later models.[30][31] In a 2024 interview Kaplan described the founding insight behind the work as the realization that "we're just gonna write a constitution for a language model and that'll change all of its behavior," contrasting the inspectable, principle-based approach with the harder-to-audit alternative of relying solely on human labels for adversarial behaviors.[32]
Kaplan also oversees the company's mechanistic interpretability effort, the project of reverse-engineering trained neural networks into human-understandable algorithmic components. In public talks he has described the work as analogous to performing a brain scan on the model and has discussed Anthropic's research on sparse autoencoders, dictionary learning, and the discovery of interpretable reasoning "circuits" inside large language models.[33] He has highlighted Anthropic's May 2024 release of work mapping millions of interpretable features inside the Claude 3 Sonnet model as a milestone for the program.[33] Beyond interpretability, his oversight portfolio includes Anthropic's alignment science team, the safety evaluations and red-teaming groups, and the model behaviors and policy research groups.[1]
On October 15, 2024, Anthropic announced an updated version of its responsible scaling policy and named Kaplan the company's Responsible Scaling Officer, succeeding co-founder and CTO Sam McCandlish, who had held the position during the policy's first year of implementation.[6] In that role Kaplan is responsible for determining whether models pass the safety evaluations the policy requires before release, and for deciding on the deployment safeguards and "AI Safety Levels" applied to new model capabilities.[6] The updated policy introduced AI Safety Level (ASL) thresholds tied to specific capability tests in domains such as chemical, biological, radiological, and nuclear (CBRN) risk, autonomous AI research and development, and persuasion or manipulation.[6] Kaplan also has authority under the policy to recommend halting deployment of any model whose capabilities trigger a higher ASL than the company is prepared to deploy safely.[6] Anthropic announced at the same time that it was hiring a Head of Responsible Scaling to coordinate the various teams involved in implementing the updated policy under Kaplan's overall direction.[6]
Kaplan represents Anthropic in many policy and public-affairs settings. In 2023 he submitted a written statement to the U.S. Senate AI Insight Forum on "Risk, Alignment, and Guarding Against Doomsday Scenarios," describing the responsible-scaling approach and Anthropic's view of catastrophic AI risk.[35] He has appeared on the TechCrunch "Equity" podcast and at TechCrunch Sessions: AI, given a Y Combinator Startup Library interview on "Scaling and the Road to Human-Level AI," spoken on the Life with Machines podcast about Constitutional AI, and participated in industry conversations at venues including Salesforce TrailblazerDX.[32][34] He was named as a witness or declarant in several legal proceedings involving Anthropic, including a 2026 declaration in a federal district court matter in the Northern District of California involving the company.[38]
Kaplan's central named scientific contribution to AI is the establishment of empirical scaling laws for neural language models, encapsulated in scaling laws paper. The result, that loss decreases as a smooth power law in model size, dataset size, and compute, supplied a quantitative basis for treating language-model development as a budgeted scaling exercise rather than an architectural search, and is widely credited with shifting both research culture and industry investment toward ever-larger pretraining runs.[4][21][34] The general framework is now referred to simply as "Kaplan scaling laws" in distinction to the later "chinchilla scaling laws" revision; both are part of the broader field of scaling laws.[23] The framework has since been extended in several directions, including a 2021 OpenAI paper on "Scaling Laws for Transfer" co-authored by Kaplan and collaborators, which examined how scaling continues to behave when models trained on one distribution are fine-tuned on a related distribution.[37]
The Kaplan paper's empirical findings have been treated as design rules in industry practice. Successive frontier models (including GPT-3, GPT-4, the Claude family, Google DeepMind's Gemini family, and others) have all been planned with explicit reference to projected loss curves derived from scaling-law fits, and frontier labs routinely run small "ladder" experiments to fit scaling exponents before committing to a large training run.[22][34] The economic implication, that returns to scale are smooth and predictable, has also influenced AI investment patterns and has been cited in policy discussions of compute governance.[35]
Within Anthropic, Kaplan is a co-author of and frequent public spokesperson for constitutional ai, the alignment training method that uses a model to revise and re-rank its own outputs against a written set of principles, replacing some or all of the human preference labels conventionally used in reinforcement learning from human feedback.[30] The technique substantially reduces the volume of human-labeled adversarial data required for harmlessness training and provides a written specification of the model's intended values that can be inspected and updated as policies evolve.[30] In 2023 Anthropic published "Claude's Constitution," a public document describing the specific principles, drawn from sources including the Universal Declaration of Human Rights and other widely recognized normative documents, that the company uses to guide Claude's training.[31] Constitutional AI has been replicated and extended by researchers at other organizations, and a sizable academic literature has emerged comparing its outcomes with conventional RLHF on safety and helpfulness benchmarks.[30]
Kaplan oversees the Anthropic team that pursues mechanistic interpretability research aimed at identifying interpretable circuits inside trained models, an approach he has publicly compared to performing a brain scan on the network.[33] He has framed the work as a long-term safety bet: if researchers can read out what a trained model is computing in human-interpretable terms, they can verify whether a model's actual reasoning matches its stated reasoning, an audit that pure black-box behavioral evaluations cannot provide.[33] The interpretability team's milestones during his tenure include the 2024 release of "Scaling Monosemanticity," which used sparse autoencoders to extract millions of human-interpretable features from a frontier-scale model.[33]
He has spoken on Capitol Hill about AI risks and policy, including the 2023 written statement to the U.S. Senate AI Insight Forum on "Risk, Alignment, and Guarding Against Doomsday Scenarios" mentioned above, and has consistently emphasized that responsible-scaling commitments and pre-deployment evaluations are complements to, rather than substitutes for, government oversight of frontier AI.[35]
Beyond the headline scaling-laws and Constitutional AI work, Kaplan is a listed co-author on numerous other Anthropic and OpenAI research outputs, including studies of red-teaming and adversarial robustness, alignment evaluations, model behavior under reward hacking, and dataset curation effects on capability. His Google Scholar page lists roughly two hundred publications between physics and machine learning, with an h-index of approximately 80 by 2026.[22]
Kaplan has produced roughly two hundred research outputs across theoretical physics and machine learning. His Google Scholar profile lists more than 150,000 total citations as of 2026, with an h-index of approximately 80.[22] A representative selection of his most cited or otherwise significant works follows.
Kaplan is among the more publicly visible Anthropic co-founders. Y Combinator hosted him in a 2024 Startup Library interview titled "Scaling and the Road to Human-Level AI," in which he discussed the company's research roadmap and his projections about when transformative AI might arrive.[34] He appeared in a Salesforce conversation series during TrailblazerDX 2024 and on the TechCrunch "Equity" podcast in 2025, and Dwarkesh Patel's 2025 book "The Scaling Era: An Oral History of AI, 2019 to 2025" devotes a chapter-length conversation to Kaplan, drawing on previously unpublished interview material.[28][34] He has also contributed opinion pieces and short essays for TechCrunch under his byline.[1] Within the AI research community he is often credited, alongside Sam McCandlish and Dario Amodei, with introducing the scaling-laws framework that organizes how the field thinks about pretraining budgets, and he is sometimes referred to by science journalists as a "patron saint" of the scaling hypothesis.[34]
Kaplan resides in Pacifica, California, and has a son.[1] As of 2026, Forbes estimated his net worth at approximately $3.7 billion, reflecting his founding equity stake in anthropic after successive funding rounds valued the company in the hundreds of billions of dollars.[1] In 2024 Kaplan was among the seven Anthropic co-founders, alongside Dario Amodei, Daniela Amodei, Tom Brown, Jack Clark, Sam McCandlish, and Chris Olah, who collectively pledged to commit approximately eighty percent of their personal fortunes to addressing AI-driven inequality and to philanthropic efforts aligned with the company's stated mission.[36] Kaplan has continued to hold his Johns Hopkins faculty appointment in parallel with his Anthropic role, and he has continued to advise graduate students in physics at Hopkins while leading research at the company.[1][3]