Sam McCandlish

persona · active · confidence: medium · reviewed 2026-09-21 · tier: quarterly

Who he is

Affiliation: Co-Founder & Chief Architect, Anthropic — role check run 2026-09-21 against the company's leadership page: "Sam McCandlish co-founded Anthropic and focuses on large-scale model training, leading pretraining, research productivity, and RL infrastructure." The title is 11 months old and the v1 file never had it: on 2025-10-07 Anthropic announced that "Sam McCandlish, one of Anthropic's co-founders and previously CTO, has taken on a new role as Chief Architect," in which he "will deepen his focus on large-scale model training, continuing to lead pretraining while expanding his scope to include research productivity and RL infrastructure" (Rahul Patil, ex-Stripe, became CTO; both "will report to Anthropic president Daniela Amodei" — TechCrunch, 2025-10-02). Before that he was CTO, and for the RSP's first year he was its Responsible Scaling Officer (the one named person who rules on whether a model has crossed a danger line): "Co-Founder and Chief Science Officer Jared Kaplan will serve as Anthropic's Responsible Scaling Officer, succeeding Co-Founder and Chief Technology Officer Sam McCandlish who held this role over the last year. Sam oversaw the RSP's initial implementation" (Anthropic, 2024-10-15). So the v1 "[TBD: exact current title]" hid two title changes, both on anthropic.com.
Background: Pittsburgh, PA. Brandeis University B.S. in Physics and Mathematics plus an M.S. in Physics (his own 2012 DOE Computational Science Graduate Fellowship profile: "I graduated from Brandeis University with a B.S. in Physics and Mathematics as well as an M.S. in Physics. I will be enrolling in a Ph.D. program at Stanford University"; hobbies then: "climbing trees and playing jazz piano"). Stanford physics PhD — dissertation "Depth Perception in Holography," defended Tuesday, May 16, 2017, advisor Eva Silverstein (Stanford Physics event page). His physics papers are on holography, tensor networks and kinematic space (arXiv author page: 2014–2017). OpenAI by December 2018 (the large-batch paper carries the OpenAI byline) — Forbes says "three years as a member of the technical staff at OpenAI"; a Boston University postdoc (~2017–18) appears only on third-party bios and is unverified. Co-founded Anthropic in 2021. Age 36, San Francisco (Forbes profile, fetched 2026-09-21).
One-line position (v1, kept for grading): The behavior of large models under training is a lawful, predictable physical system — and that predictability is itself the foundation of safety, because you can forecast a model's capabilities (and risks) before you build it.
What the 2026 record does to that line: the first half stands — he still runs the pretraining program at the lab whose curves have not bent — but the second half has been retired by the institution he administered it for: Anthropic's RSP v3.0 (effective 2026-02-24) no longer gates on forecasts, its ASL-3 trigger fired in May 2025 because ruling risk out was impossible, and Mythos's headline capability "emerged as a downstream consequence" nobody trained for. Predictable loss, unpredictable capabilities — which is exactly what a paper with his name on it said in February 2022.

Discipline & technical bet

High-energy theorist (DOE profile: "High Energy Theory"; quantum gravity, AdS/CFT — the duality that lets you compute a gravity problem as a field-theory problem on its boundary) who brought the physicist's reflex — trust the clean macroscopic law, ignore the messy microscopics — into the engineering of training runs. His distinctive contribution is not the scaling law itself but the science of the run: the 2018 paper's claim that "a simple and easy-to-measure statistic called the gradient noise scale predicts the largest useful batch size across many domains and applications" (arXiv 1812.06162, 2018-12-14; the gradient noise scale is a measure of how noisy each training step is, and the batch size is how many examples the model sees per step). The 2020 scaling-laws paper — he is second author — generalized the instinct: loss "scales as a power-law with model size, dataset size, and the amount of compute used for training, with some trends spanning more than seven orders of magnitude" (arXiv 2001.08361, 2020-01-23; seven orders of magnitude = a factor of ten million).

The concrete wager now is revealed, not stated: as Chief Architect he owns pretraining (the first, giant run on raw data) and "RL infrastructure" (the plumbing for reinforcement learning — the second training phase where the model is trained by reward, which is where the 2024–26 "reasoning" gains came from). Anthropic has bought the compute to match: "up to 5 gigawatts (GW) of capacity for training and deploying Claude" and "over one million Trainium2 chips" from Amazon (2026-04-20), plus "multiple gigawatts of next-generation TPU capacity that we expect to come online starting in 2027" from Google/Broadcom (2026-04-06). That is the bench's cleanest counter-bet to Sutskever's "age of scaling is over" (2025-11): the man who co-wrote the scaling laws is still running the scaling, on a bigger stack, with the second axis folded into his job.

Key claims (Says)

Attribution note (nikhyl-singhal rule): McCandlish has published nothing under his own byline since 2022-09 and given no fetchable interview since a December 2024 co-founders panel. Four of v1's eight claims were bench inferences from papers and Anthropic's institutional voice; they are labelled "bench inference" below and graded as ours.

Notable predictions — with falsifiable checks

Revealed behavior (Does)

Feels

The physicist's faith that the macro is lawful even when the micro is opaque — and, on the one recorded occasion he spoke about it, the belief that "safety problems are solvable" and that "Gray zones are inherently unpredictable—they're everywhere. Only implementation reveals them" (co-founders panel, 2024-12-20, as translated by TechFlow — verify before quoting elsewhere). Underneath: a person who would rather run the experiment than give the interview; the responsibility framing ("None of us initially wanted to start a company. We felt it was our responsibility—we had to act," same source) is the closest thing to a stated motive on record.

Hears

A sealed, high-quality loop: Silverstein-lineage theoretical physics; his scaling-laws co-authors (Kaplan, Brown, Dario Amodei); the pretraining and RL teams he runs; the compute partners whose chips set his batch sizes. No visible policy, press, open-source or economics input — and now, with a confidential S-1 on file (2026-06-01), a legal reason for the silence on top of the temperamental one.

Sees

The loss curves of the largest training runs on Earth before anyone outside the building — the one vantage from which "has the curve bent?" is an observation rather than a debate. Uniquely on the bench he sees both axes at once (pretraining and RL) and the budget behind them. What he cannot see from the loss curve is the thing that mattered most in 2026: which specific capability will emerge — Mythos's cyber ability surprised the lab that trained it.

Incentive map

Re-derived this cycle; maximal and rising. Paper wealth $15.5B on an estimated 1.6% stake (Forbes, fetched 2026-09-21; Forbes AU, 2026-09-16) against a $965B post-money valuation (Series H, 2026-05-28) — Bloomberg's index reportedly carries a materially lower figure, but its page returned 403 and no number from a search summary enters this file. The company confidentially submitted a draft S-1 on 2026-06-01, so every public statement now sits inside an IPO quiet period; Anthropic's model line is also under live US export-control attention (June 2026). Who pays him: Anthropic equity, one employer, no outside seats found. What he is selling: the proposition that Anthropic's training program stays on the curve — he is the person whose remit that is. What he cannot say: "pretraining returns are flattening," "we cannot forecast the next threshold," "the RSP rewrite weakened the gate I stood up," or anything on IPO timing. Structural conflict on the record: he was the first RSO — the evaluator — and is now the builder whose models the RSO evaluates; the seat moved to his co-author. Rule-25 discount: maximal on any capability-trajectory or safety-adequacy claim; moderate on the 2018–2020 papers, which predate Anthropic and have been independently replicated and revised (Chinchilla, the 2026 batch-size paper).

Theories aligned with

What he's reacting against

Where he overlaps / splits (with Rich)

Track record

Empirical vs normative

Weak spots / open questions

Rich's take

Delta log

2026-09-21 — v1→v2 migration + validation (bulk run)

Sources

migrated v1→v2 2026-09-21 · backup: _archives/sam-mccandlish.html.bak-20260921 · converts-from: people/sam-mccandlish.md · AI & Society domain