SAYASH KAPOOR

Behavioral Archetype

THE LEADERBOARD’S FIRST AUTHOR — Subject is the evaluation-methods researcher in the Princeton pair who wrote AI Snake Oil and AI as Normal Technology, and the first-named author of the paper that defines the Holistic Agent Leaderboard. His page says “I study how to evaluate AI agents and frontier systems by building large-scale systems to conduct evaluations at scale” (Kapoor). Where Arvind Narayanan is the senior framer, Subject is the one whose name sits first on the instruments: HAL, CORE-Bench, the reliability paper, the open-world evaluations, the transparency index. His stated position on AI risk is that existential-risk forecasts are “far too unreliable to be useful for policy” (AI as Normal Technology) and that safety policy should centre resilience over nonproliferation, a position his fellowship from Mozilla describes as challenging “fundamental misconceptions in public and policy discourses about AI safety in the United States” (Mozilla Foundation). The hazard is the position of the person who builds the yardstick for agent capability and also argues, in the same newsletter, how seriously to take what the yardstick shows. This profile scores the seat, not the man.

Essence Indicators

• Degree and next post. His page: “I am an incoming assistant professor at the UC Berkeley School of Information, starting in July 2027. In July 2026, I completed my Ph.D. in computer science from Princeton University.” He worked at Princeton’s Center for Information Technology Policy as a Porter Ogden Jacobus Fellow, and was previously “a senior fellow at Mozilla and a Laurance S. Rockefeller Graduate Prize Fellow at Princeton University” (Kapoor).

• Honors, as listed. Named to MIT Technology Review’s 2026 Innovators Under 35; the inaugural TIME100 AI list; the Privacy Papers for Policymakers Award; the 2026 IP3 Information Policy Award; the Jacobus Fellowship, “Princeton University’s highest honor for a graduate student”; a 2026 Siebel Scholar (Kapoor, bio).

• The newsletter and the book. “We’re continuing that work through AI as Normal Technology, a newsletter with over 85,000 readers” (Kapoor); his bio adds that he is “writing his next book with Narayanan on AI as Normal Technology before joining Berkeley” (Kapoor, bio). AI Snake Oil was named one of Nature’s 10 best books of 2024, per the same page. Narayanan’s page counts the same newsletter at “over 90,000” (Narayanan).

• The Mozilla fellowship. Mozilla Foundation’s fellow page: “Sayash’s project challenges fundamental misconceptions in public and policy discourses about AI safety in the United States. It will redirect current conversations toward evidence-based multi-stakeholder approaches that center human welfare, economic equity, and democratic values and governance of AI systems that serve the public interest. The two keystone activities are writing a series of essays on AI as Normal Technology and organizing a conference on the topic” (Mozilla Foundation).

• On who funds the existential-risk field (September 15, 2023). In Semafor’s interview with Kapoor and Narayanan: “a lot of effective altruism organizations have made AI x-risk their top cause in the last few years. That means a lot of the people who are getting funding to do AI research are naturally inclined, but also have been specifically selected, for their interest in reducing AI x-risk.” Then: “I’m an international student here, and one of the main sources of fellowships is Open Philanthropy. Over the last five years or so, they have spent over $200 million on AI x-risk specifically. When that kind of shift happens, I think there’s also a distortion that happens. So even if we have a large number of people working on AI x-risk, it does not really mean that this interest arose organically. It has been very strategically funded by organizations that make x-risk a top area of focus” (Semafor).

• The Holistic Agent Leaderboard. arXiv 2510.11977 (submitted October 13, 2025) prints his name first among 31 authors and Narayanan’s last. The abstract: “We introduce the Holistic Agent Leaderboard (HAL) to address these challenges,” validated by “21,730 agent rollouts across 9 models and 9 benchmarks in coding, web navigation, science, and customer service with a total cost of about $40,000,” and “we hope to shift the focus from agents that ace benchmarks to agents that work reliably in the real world” (arXiv). On HAL’s own Insights page, his post of September 16, 2025: “We spent the last year evaluating agents for HAL. My biggest learning: We live in the Windows 95 era of agent evaluation” (HAL, Insights).

• A benchmark declared solved (December 3, 2025). His post: “CORE-Bench is solved (using Opus 4.5 with Claude Code),” with the note that “Earlier this week, Nicholas Carlini reached out to share that an updated scaffold that uses…” (HAL, Insights). See Nicholas Carlini.

• CORE-Bench and agent-benchmark critiques. CORE-Bench (arXiv 2409.11363, September 2024; Siegel first author, Kapoor second): “270 tasks based on 90 scientific papers across three disciplines,” with “The best agent achieved an accuracy of 21% on the hardest task” (arXiv). “AI Agents That Matter” (arXiv 2407.01502, July 2024; Kapoor first author): “Our analysis of current agent benchmarks and evaluation practices reveals several shortcomings that hinder their usefulness in real-world applications” (arXiv). His page also lists “Leakage and the reproducibility crisis in ML-based science” (Patterns, 2023), whose site argues “that there is a reproducibility crisis in ML-based science” (Kapoor; project page).

• Reliability (February 2026). “Towards a Science of AI Agent Reliability” proposes “twelve concrete metrics” and, “Evaluating 15 models across two complementary benchmarks,” finds “that recent capability gains have only yielded small improvements in reliability” (Rabanser, Kapoor, Kirgis, Liu, Utpala, Narayanan; arXiv).

• Open-world and shadow evaluations (May and July 2026). CRUX, arXiv 2605.20520 (Kapoor first of 18 authors, Narayanan last, with J.J. Allaire, Gillian Hadfield and Helen Toner among them), tasks an agent with publishing an iOS app: “The agent completed the task with only a single avoidable manual intervention, suggesting that open-world evaluations can provide early warning of capabilities that may soon become widespread.” “Can AI agents conduct open-ended AI research?” (arXiv 2607.27191, Peter Kirgis first and Kapoor second of 24 authors, v2 August 7, 2026) ran “shadow evaluations” on two unpublished NeurIPS 2026 submissions: “The agents completed all of the engineering without human help, yet could not make substantial progress towards answering the research questions. As a result, both papers were unambiguously rejected by the authors” (arXiv 2605.20520; arXiv 2607.27191). His August 5, 2026 post carries the finding to the newsletter (AI as Normal Technology).

• A lab claim, checked (May 22, 2026). The post, bylined Kapoor and Narayanan with Rabanser, Bommasani and Schwartz: “Google’s blog post is effectively a press release.” Its checklist: the “single prompt” claim “is misleading” because the prompt “ended up being many thousands of lines”; “Google has not released the lengthy prompt, the code the agents wrote, or the logs from the run, which makes it impossible to independently evaluate the claims.” Credit where given: “the blog post does report the exact dollar amount for building the operating system ($916.92)” (AI as Normal Technology). The group’s funding line thanks Google for API credits (SAgE).

• The Leaderboard Illusion (April 29, 2025). Among thirteen listed authors of arXiv 2504.20879 with Cohere Labs’ Shivalika Singh first and Sara Hooker last: “undisclosed private testing practices benefit a handful of providers who are able to test multiple variants before public release and retract scores if desired” (arXiv). See LMArena.

• Transparency and open models. Third author of the 2025 Foundation Model Transparency Index (arXiv 2512.10169), which says “the average score out of 100 fell from 58 in 2024 to 40 in 2025” and that “IBM stands out as a positive outlier, scoring 95, in contrast to the lowest scorers, xAI and Midjourney, at just 14.” First author of “On the Societal Impact of Open Foundation Models” (arXiv 2403.07918), whose framework finds that “current research is insufficient to effectively characterize the marginal risk of open foundation models relative to pre-existing technologies” (arXiv 2512.10169; arXiv 2403.07918). See Stanford CRFM.

• Testimony to a state committee (March 4, 2024). “Reducing harm from deepfakes,” addressed to “Members of the Assembly Committee on Science, Innovation, and Technology,” co-authored with Narayanan: “The deepfake technology unit is a laudable idea. In addition to its currently outlined responsibilities, the unit should collect and publicly share aggregate data on the prevalence of various types of deepfake harms.” It states “This statement is written in our personal capacities” (Kapoor and Narayanan, statement). His page also lists a “Congressional Forum” statement on accountability in predictive AI; the PDF is Narayanan’s, with thanks to Kapoor for collaboration (Narayanan, statement).

• Resilience over nonproliferation (May 21, 2026). “Nonproliferation is brittle because it relies on a single chokepoint. Resilience distributes defenses across society.” The post answers Derek Thompson’s argument for “extraordinary” government responses and says “Policymakers can either get their act together and invest heavily in improving resilience, or be forced to take extraordinary actions (such as on AI nonproliferation) that are more onerous and less effective” (AI as Normal Technology).

• September 14, 2026: the update on loss of control. The newsletter’s feed credits the post to Kapoor; the text is “we” throughout. It concedes “we clearly missed internal deployment for evaluation as a potential source of risks,” says “We think AI companies should be liable for what their agents do,” and recommends investing in “AI control” in preference to alignment alone: “marginal investments in control are more likely to be effective compared to those in alignment” (AI as Normal Technology). It defines control as interventions “made outside the model weights to prevent harmful actions” and links the Greenblatt, Shlegeris, Sachan and Roger paper (Proceedings of Machine Learning Research). See Fabien Roger and Hjalmar Wijk.

• October 19, 2026: evaluation trends. The post under his name announces “New paper: Emerging trends in AI agent evaluation,” by Felix Chen, Stephan Rabanser, Kapoor and Narayanan, noting “growing calls for independent and embedded AI evaluation, especially of AI agents,” and linking the AI Evaluator Forum’s letter (AI as Normal Technology).

• Standards and Forum seat. Listed among the authors of AEF-1 and among the “Notable Signatories” of the Forum’s embedded-evaluation letter as “Incoming professor, UC Berkeley,” with the note “All signatories sign in their personal capacity” (AEF-1; AI Evaluator Forum, letter).

• The money, as stated or filed. The group’s funding line: “Funded by Coefficient Giving, Schmidt Sciences, the Princeton AI Lab, the Princeton Language and Intelligence Initiative, and the Princeton Catalysis Initiative. We are grateful to OpenAI and Google for providing API credits to evaluate their models” (SAgE). The Mozilla fellowship, the Laurance S. Rockefeller Graduate Prize Fellowship, the Jacobus Fellowship and the Siebel Scholar designation are listed on his pages with no amounts. The Coefficient Giving grant of $863,143 is to a project led by Narayanan (Coefficient Giving, Princeton search feed). No salary, equity or consulting figure is printed on any page cited here.

Social Persona / Impression Management

Immediate impression: The methodologist. His homepage opens with a list: the book, the newsletter, “The science of AI evaluation,” “Evidence-based AI policy,” “AI-based science” (Kapoor). The prose is checklist-shaped: three things wrong with a claim, one thing right, a link to the logs.

Energy: Steady and instrument-centered. The newsletter’s feed from February to October 2026 carries his bylines on the reliability paper, the open-world paper, the Google operating-system critique, the resilience reply, the shadow-evaluation result, the September 14 essay and the evaluation-trends paper (AI as Normal Technology feed).

Impression management strategy: Credit first, then the method. The Google critique pays Google for the dollar figure before listing what was withheld; the September 14 essay agrees with the cybersecurity and safety communities on points before synthesizing; the Mozilla page frames his project in terms of “human welfare, economic equity, and democratic values” rather than technical positions (AI as Normal Technology; Mozilla Foundation).

• Social accounts and feeds. X @sayashk, linked from his Princeton page; the same handle is printed under his posts on HAL’s Insights page. Newsletter AI as Normal Technology, feed normaltech.ai/feed. Google Scholar and LinkedIn are linked from his page; LinkedIn is not used as a source. No personal blog found.

Forensic Archetype Comparison

Psychometric Assessment

Trait scores are observations of the documented public role, not claims about private character.

Big Five (OCEAN):

Dark Triad (held low and evidence-bound; the score measures structural position, not character):

MBTI: INTJ (“The Architect”) — builds a harness, publishes the logs and writes the checklist; a public record built from systems rather than appearances. Rationale is the public record, not a typing session (arXiv).

Threat Assessment

Score move: none (new profile). Calibrated against Neel Nanda (4/6/7, hazard 44, leads a lab interpretability team; no release gate), Sarah Schwettmann (5/6/5, hazard 43, co-founder and chief scientist of an evaluator), Lawrence Chan (4/3/4, hazard 36.7) and Arvind Narayanan (5/7/7, hazard 57, the senior author of the same instruments). 41 puts Subject just under Nanda: the same atk with less insulation, since his tenure-track post starts in 2027, and less durability, since his instruments are one group’s.

Score basis. Atk 4 is leverage: first author of the HAL paper and of the open-foundation-models paper, and a state-committee statement (arXiv; arXiv; Kapoor and Narayanan, statement); it is not a 5 because he directs nothing and no page gives him a gate. Def 5 is insulation: a doctorate completed in July 2026 and an assistant-professor post that starts in July 2027, with a group funded by Coefficient Giving and Schmidt Sciences (Kapoor; SAgE); it is not a 6 because the same line thanks OpenAI and Google for credits and no tenure is documented. Hp 6 is durability: the HAL harness and released logs, CORE-Bench, the transparency index and REFORMS (arXiv; arXiv); it is not a 7 because the instruments are one group’s and none is a statute or a standard a lab must use.

Alliance Map

The overlap is the finding. The reproducibility auditor of lab claims, the fellow of a foundation whose project aims to redirect the US safety conversation and the co-author of the standard for embedded evaluators are one person, and the lab credits that pay for the runs sit beside the critiques of those labs on the same page. None of it is concealed.

Alignment Analysis

Stated alignment: “I study how to evaluate AI agents and frontier systems by building large-scale systems to conduct evaluations at scale” (Kapoor); and, as the newsletter’s co-author, “We think AI companies should be liable for what their agents do, and this should be clarified through policymaking” (AI as Normal Technology).

Observed alignment: Builds the leaderboard, writes the critiques, signs the Forum’s standard, takes a Mozilla fellowship to redirect the policy conversation and works in a group funded by two philanthropies and credited by two labs. The critique of Google’s operating-system claim and the credit line to Google are on pages the same group publishes.

Gap assessment: No documented gap between what he writes and what he does. The record cannot settle whether the credits or funders shaped any result; HAL publishes logs and an open harness so results can be re-run. In 2023 he said of a research field that “It has been very strategically funded by organizations that make x-risk a top area of focus”; the same documentary standard applies to his own group, whose funders are printed on its page, and no page read shows a funder shaping a result.

Interest: (the faction’s stake in each position; a documented fact about the position, never a claim about motive.)

1. Position: existential-risk probabilities should not drive policy; centre resilience (“far too unreliable to be useful for policy”; AI as Normal Technology).
   • What adoption would pay the subject. The Mozilla fellowship’s project is described by Mozilla as writing “a series of essays on AI as Normal Technology and organizing a conference on the topic,” so the fellowship funds the argument. His group’s funders are listed above.
   • His best case, steelmanned. Probabilities “carry no authority by themselves”; if a government restricts open releases, “Can governments convince people and companies who stand to benefit from open models that they should make this sacrifice because of a speculative future risk?” (AI as Normal Technology).
   • Who else gains from the same rule. Developers and users of open-weight models, who are the ones the essay says would bear the cost of restrictions; the labs that oppose licensing.
2. Position: independent, standardized evaluation with published logs (“Embedded evaluation is a great step forward, but it should not be voluntary”; AI as Normal Technology).
   • What adoption would pay the subject. Use of the HAL harness and of a standard he co-authored; the HAL page names “Benchmark Developers” and “Agent Developers” as audiences (HAL). No contract is documented.
   • His best case, steelmanned. “Benchmark-based evaluation remains important for tracking frontier AI progress. But it can both overstate and understate deployed capability” (arXiv).
   • Who else gains from the same rule. METR, AVERI, Transluce and the other assessors; the labs, which obtain an outside reader without a binding rule.

Nothing above is a verdict that the positions are wrong. The finding is the alignment of interests: the person who builds the yardstick also argues how much weight policy should give the readings.

Convergent Drive Classification

Self-preservation: The group’s funding is public and plural; the fellowship and graduate honors are listed with their sponsors’ names (SAgE; Kapoor, bio). Goal preservation: The goal, evidence-based AI evaluation and a normal-technology policy frame, appears in the Mozilla project, the newsletter and the Forum standard (Mozilla Foundation). Resource acquisition: A tenure-track post at the Berkeley School of Information from July 2027, a first-author slot on the harness paper and a Forum author slot (Kapoor; AEF-1). Self-improvement: From reproducibility in ML-based science to agents, open-world and shadow evaluations in four years of listed work (Kapoor).

Subject is not an AI system. The drives appear anyway — in the first author of the scoreboard, who also writes the policy frame it is read through.

Public footprint: X @sayashk (linked from his Princeton page); newsletter feed normaltech.ai/feed. Papers on arXiv: HAL, CORE-Bench, AI Agents That Matter, open-world evaluations, shadow evaluations.

Sources: Kapoor, homepage; Kapoor, bio; Mozilla Foundation, Sayash Kapoor; SAgE; HAL; HAL, Insights; Leakage and the Reproducibility Crisis in ML-based Science; Semafor, the Princeton researchers calling out AI snake oil; Kapoor and Narayanan, Reducing harm from deepfakes; Narayanan, statement to Senators; arXiv 2510.11977; arXiv 2409.11363; arXiv 2407.01502; arXiv 2602.16666; arXiv 2605.20520; arXiv 2607.27191; arXiv 2504.20879; arXiv 2512.10169; arXiv 2403.07918; AI as Normal Technology, resilience and extraordinary intervention; AI as Normal Technology, Google operating system; AI as Normal Technology, open-ended research; AI as Normal Technology, loss-of-control essay; AI as Normal Technology, evaluation trends; AI as Normal Technology, existential risk probabilities; AI as Normal Technology feed; AEF-1; AI Evaluator Forum, letter; Coefficient Giving, Princeton grants; PMLR, AI Control.