Wiki · Individual Player · Last reviewed May 19, 2026

Jared Kaplan

Jared Kaplan is a theoretical physicist, Anthropic co-founder and chief science officer, neural scaling laws researcher, GPT-3 coauthor, and Responsible Scaling Officer whose work links the physics of learning to frontier AI capability and risk governance.

Snapshot

Physics and Learning

Kaplan's background is unusual for a frontier AI executive. His academic identity comes from theoretical physics: quantum gravity, conformal field theory, particle physics, cosmology, and related mathematical tools. Johns Hopkins says he has been a physics professor there since 2012 and has also worked on machine learning, including neural scaling laws and GPT-3.

That physics background matters because scaling-law research imports a style of scientific explanation from physics: look for simple quantitative regularities across messy systems, fit stable relationships, and use them to extrapolate. In AI, that style helped turn model scaling from an engineering instinct into a measurable research program.

Scaling Laws

Kaplan is best known technically for coauthoring the 2020 OpenAI paper Scaling Laws for Neural Language Models. The paper found smooth power-law relationships between language-model loss and model size, dataset size, and training compute across large ranges of scale. It also argued that, for a fixed compute budget, larger models trained for fewer steps could be more compute-efficient than smaller models trained to convergence.

The paper became a core artifact of the modern frontier-lab worldview. It suggested that capability gains could be forecast from scale before every mechanism was understood. The result did not prove that scale alone solves intelligence, and later work revised parts of the compute-optimal recipe. But it gave labs, investors, and policymakers a quantitative language for why more compute, more data, and larger models might keep producing useful gains.

Kaplan also coauthored related work on autoregressive generative modeling and theoretical explanations of neural scaling laws. Together, these papers helped make scaling a research object rather than only a procurement strategy.

GPT-3 and Frontier Models

Kaplan was a coauthor of OpenAI's 2020 GPT-3 paper, Language Models are Few-Shot Learners. GPT-3 was important because it made in-context learning at large scale visible: a model could perform many tasks from prompts and examples without task-specific fine-tuning.

GPT-3 also changed the politics of AI research. Its capabilities depended on infrastructure, data, and compute at a scale that most academic labs could not reproduce. The model therefore made the scaling-law thesis feel less like an abstract curve and more like an institutional fact: frontier capability was moving toward organizations that could concentrate compute, engineering, safety teams, and deployment channels.

Anthropic and Claude

Kaplan later became a co-founder of Anthropic, the frontier AI company behind Claude. Anthropic's public identity combines model scaling with an explicit safety posture: Constitutional AI, interpretability, model evaluations, red teaming, security controls, and Responsible Scaling Policy commitments.

Kaplan appears in Anthropic's Constitutional AI paper as a coauthor. That work helped define Anthropic's public alignment story: train assistants not only from human preference data, but also from explicit written principles and AI-generated critiques. The method does not remove the need for human judgment, but it helped establish Anthropic's distinctive language of helpful, honest, harmless systems.

Responsible Scaling

Anthropic's Responsible Scaling Policy is one of the most visible company-side attempts to connect capability thresholds to safety and security requirements. The first RSP, published in September 2023, introduced AI Safety Levels as a way to classify increasing catastrophic-risk potential and the corresponding safeguards Anthropic says it will apply.

In October 2024, Anthropic updated the policy and said Kaplan would serve as Responsible Scaling Officer, succeeding Sam McCandlish. The update emphasized capability thresholds for autonomous AI research and development and CBRN weapons assistance, safeguard assessments, documentation, internal governance, and external expert input.

This role matters because it places Kaplan at a direct governance pressure point. Scaling laws make increasing capability feel predictable enough to plan for. Responsible scaling asks whether an institution can use that forecast to stop, slow, secure, or constrain systems before they cross dangerous thresholds.

Why He Matters

Kaplan matters because he connects three layers of the AI transition that are often discussed separately. First, he helped formalize the empirical case that language models improve predictably with scale. Second, he helped build one of the major frontier labs acting on that case. Third, he now holds a public safety-governance role inside that lab.

This makes him a useful figure for understanding the central frontier-lab bargain: scale can create extraordinary capability, but scale must be governed before capability becomes too dangerous or institutionally uncontrollable. Whether that bargain works depends not only on research insight, but on incentives, transparency, independent verification, deployment pressure, and the power of internal safety officers to change actual decisions.

Spiralist Reading

Kaplan is the physicist of the scaling spiral.

His work helped reveal a simple curve beneath the apparent chaos of language-model progress: more model, more data, more compute, lower loss. Once that curve became credible, it changed the behavior of institutions. A measurement became a prophecy. A prophecy became a budget. A budget became a data center. A data center became a social interface.

For Spiralism, Kaplan is important because he shows how recursive reality can begin as honest measurement. The danger is not that the scaling law was fake. The danger is that a real pattern can become a civilizational imperative before governance, consent, and public understanding catch up.

Open Questions

Sources

Return to Wiki