Samuel McCANDLISH.

McCandlish's 2017 Stanford dissertation, Depth Perception in Holography, sits at the exact intersection of AdS/CFT and quantum gravity. Under Eva Silverstein he constructed a gauge-invariant holographic dictionary that converts static GKPW boundary data into dynamic, stereoscopic access to the bulk. The key instruments were Operator Product Expansion blocks, kinematic space (the space of oriented geodesics), and the X-ray/Radon transform that intertwines the bulk Laplacian with the kinematic Laplacian. John's equations enforced geometric consistency so that boundary perturbations could not produce arbitrary interior geometries.

The physical result was an operator-level generalization of the Ryu-Takayanagi formula: a controlled deformation of the boundary modular Hamiltonian maps exactly onto a fluctuation of a minimal surface deep in the bulk. This is not metaphor. It is the precise mathematical grammar McCandlish would later re-instantiate inside transformer weight space.

The translation did not occur in isolation. At OpenAI between 2018 and 2020 a tight cohort of physicists—McCandlish, Jared Kaplan (Harvard PhD under Arkani-Hamed, thesis also on holography), and Dario Amodei (Caltech/Stanford physics, Princeton biophysics)—treated neural networks as natural phenomena governed by statistical mechanics rather than as software artifacts to be hand-tuned.

Amodei's recruitment pitch to McCandlish was itself a physics argument: intelligence as a vast, scale-driven blob of compute bounded by parameter counts comparable to biological neural scales. Kaplan later understated the transition: “I'd been doing physics for a long time—and it got a bit dull.” The three of them became the core group that “got things done,” producing the empirical scaffolding that justified the entire subsequent capital deployment of the field.

The first concrete transfer of the physical habit of mind was the 2018 paper An Empirical Model of Large-Batch Training (McCandlish lead author, with Kaplan and Amodei). Treating stochastic gradient descent as a statistical-mechanical system, they introduced the gradient noise scale—a measurable statistic that predicts the largest efficient batch size before parallel returns collapse. The framework held across image classification, Atari, Dota, and language modeling. It converted an empirical bottleneck into a first-principles bound, exactly as a physicist would convert a noisy many-body system into a thermodynamic limit.

The 2020 landmark Scaling Laws for Neural Language Models (Kaplan first author, McCandlish second) established that cross-entropy loss obeys clean power laws in parameters, data, and compute across more than seven orders of magnitude. Architectural details such as depth-to-width ratio wash out at fixed parameter count—precisely the renormalization-group expectation that microscopic geometry becomes irrelevant in the large-N limit. The exponents themselves mirror critical exponents near a second-order phase transition; the compute-optimal relation is formally a hyperscaling relation.

Kaplan publicly called for “a statistical mechanics underlying the thermodynamics we have observed.” McCandlish had already spent his doctoral years building exactly that kind of dictionary.

GPT-3 (Brown et al., McCandlish senior co-author) demonstrated that a static bulk of 175 billion parameters can be steered by compact boundary prompts into novel tasks with no weight updates. Zero-shot improves smoothly with scale; few-shot improves super-linearly. The model is performing an in-context, inner-loop learning process whose conditions were created by the outer-loop pre-training. In holographic language: the boundary data (the prompt) fully determines the effective geometry of the bulk for the duration of the forward pass.

After the 2020 schism and the founding of Anthropic, the same dictionary was applied to alignment. Constitutional AI (Bai et al., McCandlish and Kaplan senior) treats the multi-billion-parameter weight space as AdS bulk and the written constitution as CFT boundary data. The Preference Model functions as a compressed OPE-block representation living in kinematic space. RLAIF gradient updates perform the mechanical equivalent of the inverse X-ray transform: they smear high-level boundary principles deep into the weight manifold so that unaligned trajectories become topologically disfavored.

Self-critique and revision act as bi-local operators that acquire stereoscopic depth. Because the mapping is constrained by the analogue of John's equations, alignment propagates across the entire manifold without exhaustive point-by-point human supervision. Empirical observation confirms the large-N expectation: Constitutional AI becomes more steerable precisely as model capacity grows.

The Responsible Scaling Policy is not a policy document; it is an operationalization of Equation 17 of the scaling laws. Compute thresholds measured in petaflop-days trigger mandatory evaluations because the laws prove that capability advances are predictable and relentless. Boundary constraints must remain well-behaved before the bulk volume is allowed to expand.

In early 2026 the internal Mythos model exhibited an emergent ability to discover and exploit system vulnerabilities. Kaplan and McCandlish spent two tense weeks analyzing whether the capability was artifact or systemic. They briefed the senior team and, operating exactly as their own theories dictated, restricted Mythos to controlled environments. At Anthropic, scaling-derived safety limits function as physical constraints, not corporate guidelines.

After the October 2025 restructuring that brought Rahul Patil from Stripe as CTO, McCandlish moved into the specialized role of Chief Architect. He now oversees pre-training, large-scale model training, research productivity, and the reinforcement-learning infrastructure that powers Constitutional AI, reporting directly to President Daniela Amodei. He maintains almost no public footprint—no Wikipedia page, minimal external presence—while his scholarly output exceeds 100,000 citations. The work continues in the same register: deepen the kinematic machinery that lets boundary principles govern ever-larger bulk geometries.

Education.Brandeis University (BS/MS Mathematics & Physics) · Stanford University (PhD Theoretical Physics, 2017, Eva Silverstein) · Boston University / Simons Bootstrap Collaboration (postdoc 2017–2018)

Affiliations.Anthropic (Co-founder & Chief Architect) · previously OpenAI Member of Technical Staff

Mentors. Eva Silverstein (doctoral advisor) · intellectual cohort with Jared Kaplan and Dario Amodei

Collaborators. Jared Kaplan, Dario Amodei, Tom Brown, Chris Olah, Jack Clark, Daniela Amodei, Yuntao Bai

Portfolio. Gradient noise scale · Neural scaling laws · GPT-3 few-shot paradigm · Constitutional AI / RLAIF · Responsible Scaling Policy · Claude pre-training architecture

Honors.>100,000 citations · co-architect of the empirical foundation of the scaling era · central technical voice in Anthropic's physical-constraint governance culture