Jared Kaplan

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

Who he is

Affiliation: Co-Founder & Chief Science Officer, Anthropic — role check run 2026-09-21 against the company's leadership page ("co-founded Anthropic and heads its technical research efforts, leading foundational work on AI safety and development"). Also Anthropic's Responsible Scaling Officer (RSO — the one named person who rules on whether a model has crossed a danger threshold) since 2024-10-15, succeeding Sam McCandlish; last independently confirmed in that seat by TIME on 2025-08-26. RSP v3.x (2026) keeps the RSO role, but none of the v3 texts fetched this cycle name the holder — presumed still Kaplan, not re-verified after 2025-08. Associate professor of physics, Johns Hopkins (since 2012), on leave (Forbes profile, 2026-09-21). Harvard physics PhD 2009 (thesis "Aspects of Holography"); Stanford BS in physics and mathematics; Hertz Fellow 2005; NSF CAREER Award 2015. Joined OpenAI in 2019; first author of "Scaling Laws for Neural Language Models" (arXiv, 2020-01-23); fifth of 31 authors on GPT-3 (2020-05-28); senior (last) author of Constitutional AI (2022-12-15). Paper wealth: $15.5B (Forbes, as of 2026-09-21; ~1.6% stake; Forbes 400 debut alongside five co-founders, Forbes Australia 2026-09-16).
One-line position: AI progress follows smooth, physics-like power laws in compute, data, and parameters — making capability growth forecastable, timelines short, and safety work urgent because the trend lines are not slowing. Still his frame in 2026: Mythos is "really just kind of the culmination of the trend" (The Free Press, 2026-04-16). What changed is the safety half: in February 2026 he personally defended Anthropic dropping its pledge to pause when safeguards lag — "We felt that it wouldn't actually help anyone for us to stop training AI models" (TIME, 2026-02-24). The curves stayed; the gate moved.

Discipline & technical bet

Theoretical physicist (effective field theory of inflation, the conformal bootstrap — a way of pinning down quantum field theories from consistency conditions alone — and AdS/CFT; INSPIRE, fetched 2026-09-21) who brought the physicist's habit of trusting a clean macroscopic law over messy microscopic detail into machine learning. The 2020 paper's core sentence is still the bet: 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). The current, concrete wager (YC AI Startup School, 2025-06-16): "you can scale up the compute in both pre-training and RL and get better and better performance"; the length of tasks models can do "is doubling roughly every seven months"; current software tasks sit at "a time scale of hours" and "over the next few years, we may reach a point where AI models can do tasks that don't just take us minutes or hours, but days, weeks, months, years." The missing ingredients he names are organizational knowledge, memory, and fine-grained oversight — engineering gaps, not a new science. Contrast Sutskever, who declared the "age of scaling" over in 2025-11: Kaplan is the bench's clearest remaining scaling loyalist, and he now runs the lab that just raised $65B on that premise.

Key claims (Says)

Notable predictions — with falsifiable checks

Revealed behavior (Does)

Feels

Wants the physics to be true — that intelligence really is a clean curve and he read it first. Underneath: the referee's unease. He chose the ASL-3 "limb" and then signed off on removing the promise to stop; the Guardian's "scary process... you don't know where you end up" is the most unguarded thing he has said on record.

Hears

Anthropic's internal evals and training runs first; Dario and the founding cohort (McCandlish, Brown); METR's horizon data (he cites it by name); and since 2026, bankers and the IPO calendar. Almost no academic-economics input — the diffusion literature (Acemoglu, Cowen) does not reach him.

Sees

Every frontier Claude before anyone else, and — uniquely on the bench — the raw danger evals (bio, cyber) that decide whether it ships. He saw Mythos find zero-days in every major OS and browser months before the public did. That vantage is why his timeline claims cannot be dismissed and why they cannot be independently checked.

Incentive map

Re-derived from scratch this cycle; the v1 note ("CSO of a frontier lab — rule 25 applies; academic past partially mitigates") is now too mild. Paid by: ~1.6% of Anthropic, $15.5B on paper at the $965B Series H (2026-05-28), with a confidential S-1 filed 2026-06-01 — priced daily by public markets within months if the listing proceeds. Selling: frontier models and agentic coding (Claude Code); run-rate >$47B (2026-05). Structural conflict unique to him: as RSO he adjudicates whether the products that set his net worth are safe enough to ship — the referee is a 1.6% owner, and under RSP v3 the thresholds he referees are "public goals," not commitments. Cannot say: "we cannot make this generation safe" (blocks the revenue line); "timelines are long" (imminence sells the round); "RSP v3 weakened safety" (he defended it on the record). Thesis-is-the-product pairing (andrew-ng rule): his loudest 2025 claim — most white-collar work in 2–3 years — is the demand forecast for the product he rations compute toward. Mitigants, logged honestly: the academic record predates the venture; Mythos Preview was withheld at real cost; and he is the one executive on the bench whose title makes a restrictive call his job rather than a betrayal of it. Rule-25 discount: maximal on timelines and on any "the model is safe enough" claim; moderate on scaling-law science, which is public, replicated, and older than his equity.

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/jared-kaplan.html.bak-20260921 · converts-from: people/jared-kaplan.md · AI & Society domain