Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Verifying International Agreements on AI: Six Layers of Verification for Rules on Large-Scale AI Development and Deployment

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Six largely independent verification layers—on-chip, off-chip, and personnel-based—could let states verify compliance with international AI agreements, with the main open questions being engineering and infrastructure rather than…

desk verdict A candid, well-structured roadmap for AI verification, but the 'six independent layers' claim overstates redundancy until correlated failure modes are analyzed. read the letter →

arxiv 2507.15916 v2 pith:5MR5ABIJ submitted 2025-07-21 cs.CY

Mauricio Baker , Gabriel Kulp , Oliver Marks , Miles Brundage , Lennart Heim This is my paper
classification cs.CY
keywords AIverificationinternationalagreementslarge-scalecomputelayersaccountingConfidentialComputinghardwaresecuritywhistleblowerprograms
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Frontier AI risks may require international agreements, but states will only sign and keep such agreements if they can verify each other's compliance. This paper argues that compliance with rules on large-scale AI development and deployment could eventually be verified through six largely independent verification layers: security features built into AI chips, network taps and analog sensors attached to those chips, and three personnel-based mechanisms (whistleblower programs, interviews, and national intelligence). The authors claim that every verification subgoal can be completed by plausible mechanisms, so that a stacked regime would provide substantial redundancy. They also argue that the main obstacles are not fundamental impossibility but unsolved engineering, hardware-security, and infrastructure problems, plus guardrails to prevent verification itself from enabling abuse or power concentration.

What carries the argument

The load-bearing device is the verification layer: a collection of similar mechanisms that can complete every subgoal end-to-end without needing another layer. Six such layers are assembled from over twenty mechanisms, with the on-chip layer using hardware-backed workload certificates and compliance-locked chips, the two off-chip layers using network taps and analog-sensor compute accounting, and the personnel layers applying whistleblower programs, interviews, and intelligence to all subgoals. The framework's compute accounting idea makes large-scale AI compute use the audited resource, on the grounds that compute has a distinctive physical footprint and is less intrusive to account for than power or algorithm design.

What would settle it

A concrete falsifier: show that a leading AI chip's Confidential Computing or secure boot can be broken by a team with design-level knowledge and physical access, or that a spoofed training declaration passes partial re-execution within the error bars allowed by analog-sensor accounting; either demonstration would remove an entire layer and cut the promised redundancy from six to four.

Watch

Extended reading notes

Core claim

The report's central claim is that international verification of rules on large-scale AI compute is feasible in principle, with enough redundancy that no single failure mode need be fatal. Verification is decomposed into four subgoals: declared large-scale AI compute uses are accurate (1A), declared uses have the required properties (1B), no undeclared large-scale uses occur on declared clusters (2A), and no large undeclared clusters exist (2B). The paper proposes six verification layers that each address all subgoals: on-chip secure boot and Confidential Computing; off-chip network taps; off-chip analog sensors (power and other physical measurements) with partial workload re-execution; whistleblower programs; personnel interviews; and national intelligence. The paper concludes that all subgoals may be completed by plausible verification mechanisms, forming up to six redundant layers, while acknowledging that the on- and off-chip layers will likely remain circumventable until substantial R&D progress is made.

Load-bearing premise

The whole scheme rests on the assumption that hardware security features such as secure boot, Confidential Computing, and tamper-evidence can be hardened enough to survive a state-level adversary who controls chip design, manufacturing, and the operating environment; the paper itself calls this challenge severe and unsolved.

Editorial extensions

If this is right

  • If the six-layer claim holds, an international AI agreement can be designed so that no single defeated mechanism eliminates verification, because the remaining layers still provide separate evidence for each subgoal.
  • The framework gives concrete structure to which R&D should be funded first: secure boot and Confidential Computing strong enough for state-level adversaries, network-tap hardware that can handle inter-chip bandwidth, and analog-sensor methods with small error bars.
  • Personnel-based layers could provide some verification capacity almost immediately, before hardware layers are stress-tested, because they require no new core technology.
  • The paper argues that verifiable use of exported chips could allow export controls to become more targeted, preserving national security goals while letting compliant chip uses proceed.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable next step would be a defender-versus-adversary red-team exercise on a simulated verification stack, measuring how many independent layers survive when the adversary controls chip design, data-center operations, and personnel vetting; the paper lists red teaming as a next step but does not run such an exercise.
  • If on-chip secure boot matures, the same hardware attestation could become a general infrastructure layer for domestic AI regulation and independent auditor access, not just international agreements.
  • The paper's focus on thousands of chips leaves open the question of decentralized training across many smaller clusters; its personnel layers would partly cover this, but a quantitative threshold for 'thousands' would need calibration against future capability trends.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper develops a conceptual framework for verifying compliance with international agreements on large-scale AI development and deployment. It decomposes the verification goal into four subgoals (accurate declared uses, compliant declared uses, no undeclared uses of declared clusters, and no undeclared clusters) and then proposes six "layers" of verification: on-chip security features, off-chip network taps, off-chip analog sensors, whistleblower programs, personnel interviews, and national intelligence activities. The central claim is that these layers are largely independent and provide substantial redundancy, so that states could eventually verify compliance if the listed R&D and infrastructure challenges are addressed. The paper includes detailed implementation appendices, a list of open problems, and draws on literature plus 18 expert interviews.

Significance. If its central claim is accepted, the paper makes a useful contribution by giving policy audiences a structured map of verification options and by identifying concrete R&D gaps. Its strengths include a clear subgoal decomposition, unusually detailed implementation sketches in the appendices, an honest enumeration of open problems (e.g., network tap analysis, hardware security), and a sequenced R&D agenda. However, the "six largely independent layers" claim is the load-bearing result, and it is not established by the current analysis because failure correlations across layers are not examined and several layers are admittedly based on unsolved technical problems. The paper is therefore valuable as a scoping and agenda-setting document, but its headline redundancy claim needs to be substantially qualified or defended.

major comments (3)
  1. [Section 4 (Table 1 and Figure 2); Section 3.2] The claim that the six layers provide "largely independent" verification with "substantial redundancy" is not supported by any analysis of failure correlation. Layers 1-3 (on-chip, network taps, analog sensors) all ultimately depend on the integrity of hardware against a Prover that controls chip design and manufacturing; layers 4-6 all depend on human leakiness and can plausibly be defeated by compartmentalized, small, loyal groups. A single failure mode can therefore disable several layers at once. The "weakest link" heuristic in Section 3.2 applies to subgoals within a regime, not to the independence of layers, so it does not establish the effective redundancy of the six-layer stack. The authors should either provide an argument that the failure modes are substantially uncorrelated, or weaken the claim to two or three independent verification families with sub-variants.
  2. [Appendix A.3; Section 4.2] The central claim that "all subgoals may be completed by plausible verification mechanisms" is directly undercut by the paper's own statement that analyzing network tap data to verify that a cluster executed only a claimed workload is "an unsolved problem." Because this analysis is required for the network-tap layer to complete Subgoals 1.A and 2.A, Layer 2 cannot currently deliver end-to-end verification. The paper needs to state explicitly whether the central claim is conditional on solving this problem, and to give some concrete evidence or argument that the problem is plausibly solvable rather than merely listing it as an open challenge.
  3. [Section 4.1.2 and Table 6] The on-chip layer is described as facing "severe, unsolved hardware security challenges," with the paper noting that current hardware security features have mixed track records and that on-chip mechanisms "will likely have irreparable design flaws that allow them to be circumvented" absent replacement of millions of chips. Counting Layer 1 as one of six redundant layers therefore requires an explicit assumption about future hardware security that is not currently met. The paper should either downgrade this layer to a conditional or aspirational layer, or provide a concrete technical pathway under which secure boot, Confidential Computing, and tamper-evidence can be made robust against state-level adversaries who control the chip supply chain.
minor comments (5)
  1. [Section 5] The text refers to "Appendix c.3" but should refer to "Appendix C.3".
  2. [Section 4.1.2] The phrase "non-complaint AI data centers" should read "non-compliant AI data centers."
  3. [Section 4.3] The parenthetical cross-reference "Section 4.3: Section 4.4" should use a semicolon or comma instead of a colon.
  4. [Abstract] The phrase "confidentially overseeing AI development" is awkward; "confidential oversight of AI development" would be clearer.
  5. [Figure 2] Figure 2 is extremely dense and difficult to read at normal print size; consider splitting it into per-layer panels or increasing font sizes.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the six-layer framework is a definitional taxonomy, and the feasibility claims rest on mechanism-level analysis, expert input, and explicitly open R&D challenges, not on the framework itself.

full rationale

The paper's central contribution is a decomposition of a verification goal into subgoals and an assembly of mechanisms into six 'layers.' The subgoal decomposition is explicitly definitional: 'if all the claims in the framework are true, then it will logically follow that the Prover cannot have used large-scale AI compute in a non-compliant manner' (Section 3.1). This is a valid analytical decomposition, not a prediction derived from fitted inputs, and it is not presented as empirical evidence. The six layers are defined by the authors' own criterion ('A verification layer is a collection of similar verification mechanisms, with one mechanism for each verification subgoal,' Section 4), so the number six is a counting result of that taxonomy, not a circular derivation of feasibility. The load-bearing feasibility claim—that states 'could eventually verify compliance'—is supported by mechanism-level analyses, 18 expert interviews, and a candid list of unsolved problems (e.g., Section 4.1.2: 'On-chip verification faces severe, unsolved hardware security challenges'; Table 6). The paper cites prior work by its own authors (Brundage et al., 2020; Kulp et al., 2024; Heim et al., 2024) for specific mechanisms such as offline licensing and compute accounting, but these are used as inputs or prior proposals, not as the sole justification of the six-layer redundancy claim, and the paper supplements them with independent technical discussion and external references (e.g., FlexHEGs from Petrie et al.). The skeptical concern that layers 1–3 share a hardware-root-of-trust failure mode and layers 4–6 share a human-leakage failure mode is a substantive challenge to the 'largely independent' characterization, but it is a question about the strength of evidence and failure-correlation analysis, not a demonstration that any conclusion is equivalent to its inputs by construction. No fitted parameter is relabeled as a prediction, no uniqueness theorem is imported from the authors' prior work to forbid alternatives, and no known result is merely renamed. The report is self-contained as a conceptual taxonomy and roadmap, with its main limitations explicitly acknowledged.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests mainly on the chosen scale threshold, cooperation assumptions, and the feasibility of securing hardware against advanced adversaries. No new physical entities are introduced. The framework's internal logic is definitional rather than empirical.

free parameters (1)
  • large-scale compute threshold = thousands of high-end AI chips over multiple months
    The paper's scope definition, chosen to track near-frontier training, determines what falls under verification and what is considered intrusive. The authors note it could be set higher, e.g., hundreds of thousands of chips, for practicality.
assumptions (4)
  • domain assumption States will agree on rules and cooperate with verification, including declarations, inspections, and whistleblower access.
    The Prover/Verifier framework assumes an international agreement exists and that parties self-report and allow verification. Section 3.1 establishes this context.
  • domain assumption Large-scale AI compute use is a meaningful and stable proxy for the AI activities that pose international security risks.
    Justifies the compute-accounting approach and the focus on thousands of chips. The paper acknowledges small-scale deployment may also be dangerous, noting this as a scope limitation.
  • ad hoc to paper Hardware security primitives (secure boot, Confidential Computing, tamper-evidence) can be made robust against nation-state adversaries.
    The on-chip layer requires this, but Section 4.1.2 lists severe unsolved challenges and mixed track records, so this is a load-bearing and currently unmet assumption.
  • domain assumption Personnel involved in large-scale AI development are numerous enough and not fully automatable, making whistleblower and interview layers effective.
    Supports layers 4 and 5. The report itself notes AI automation may reduce personnel, weakening these layers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Verifying International Agreements on AI: Six Layers of Verification for Rules on Large-Scale AI Development and Deployment." pith.science (2026). https://pith.science/paper/5MR5ABIJ

@misc{pith2026250715916,
  author       = {Pith},
  title        = {Pith review of: Verifying International Agreements on AI: Six Layers of Verification for Rules on Large-Scale AI Development and Deployment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5MR5ABIJ}},
  note         = {Machine review of arXiv:2507.15916}
}
read the original abstract

The risks of frontier AI may require international cooperation, which in turn may require verification: checking that all parties follow agreed-on rules. For instance, states might need to verify that powerful AI models are widely deployed only after their risks to international security have been evaluated and deemed manageable. However, research on AI verification could benefit from greater clarity and detail. To address this, this report provides an in-depth overview of AI verification, intended for both policy professionals and technical researchers. We present novel conceptual frameworks, detailed implementation options, and key R&D challenges. These draw on existing literature, expert interviews, and original analysis, all within the scope of confidentially overseeing AI development and deployment that uses thousands of high-end AI chips. We find that states could eventually verify compliance by using six largely independent verification approaches with substantial redundancy: (1) built-in security features in AI chips; (2-3) separate monitoring devices attached to AI chips; and (4-6) personnel-based mechanisms, such as whistleblower programs. While promising, these approaches require guardrails to protect against abuse and power concentration, and many of these technologies have yet to be built or stress-tested. To enable states to confidently verify compliance with rules on large-scale AI development and deployment, the R&D challenges we list need significant progress.

Figures

Figures reproduced from arXiv: 2507.15916 by the authors.

Figure 1
Figure 1. Framework of verification subgoals. We decompose a broad verification goal into subgoals, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Verification layers consist of distinct mechanisms for each verification subgoal. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Assumed context for verification: verification is preceded by declarations and can lead to [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Framework of verification subgoals. for compliance. The substance of these tests would depend on what rules are being verified (Section 2.1), though as we will discuss, the infrastructure for running tests could be rule-agnostic. • Subgoal 2: verifying that there are n…
Figure 5
Figure 5. Figure 5: Summary of how the on-chip verification layer would complete each subgoal. [PITH_FULL_IMAGE:figures/full_fig_p020_5.png]
Figure 6
Figure 6. Figure 6: Summary of how off-chip verification layers would complete each subgoal. Note “verifying [PITH_FULL_IMAGE:figures/full_fig_p024_6.png]
Figure 7
Figure 7. Figure 7: Summary of how personnel-based verification layers would complete each subgoal. Each [PITH_FULL_IMAGE:figures/full_fig_p029_7.png]
Figure 8
Figure 8. Figure 8: Supplementary verification mechanisms, each under the verification subgoal(s) it could [PITH_FULL_IMAGE:figures/full_fig_p031_8.png]
Figure 9
Figure 9. Figure 9: How AI computing hardware is organized. AI data centers are facilities that host AI [PITH_FULL_IMAGE:figures/full_fig_p066_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. How to Catch a GPU: A Taxonomy of Verification and Enforcement Mechanisms for International AI Agreements

    cs.CY 2026-06 conditional novelty 6.0 of 10

    Verification of international AI agreements will fail first at detecting hidden compute facilities, around the 10,000-H100-equivalent scale, before other enforcement mechanisms break.

Reference graph

Works this paper leans on

21 extracted references · 11 canonical work pages · cited by 1 Pith paper

  1. [7]

    Lennart Heim

    URL https://www.thonking.ai/p/strangely-matrix-multiplications. Lennart Heim. A Trusted AI Compute Cluster for AI Verification and Evaluation, March 2024. URL https://blog.heim.xyz/a-trusted-ai-compute-cluster/. Lennart Heim. The Rise of DeepSeek: What the Headlines Miss, January 2025a. URL https: //www.rand.org/pubs/commentary/2025/01/the-rise-of-deepsee...

    work page 2024

  2. [8]

    Dan Hendrycks, Eric Schmidt, and Alexandr Wang

    URL https://www.oxfordmartin.ox.ac.uk/publications/governing-through-the-cloud-the- intermediary-role-of-compute-providers-in-ai-regulation/. Dan Hendrycks, Eric Schmidt, and Alexandr Wang. Superintelligence Strategy: Expert Version,

    work page

  3. [10]

    URL https://www.semanticscholar.org/paper/Su rvey-of-Recent-Developments-for-Hardware-Trojan-Jain-Zhou/45d190fc5ab153d9beca3b392 cdd141fbc66fcf2

    doi: 10.1109/ISCAS51556.2021.9401143. URL https://www.semanticscholar.org/paper/Su rvey-of-Recent-Developments-for-Hardware-Trojan-Jain-Zhou/45d190fc5ab153d9beca3b392 cdd141fbc66fcf2. Hengrui Jia, Mohammad Yaghini, Christopher A. Choquette-Choo, Natalie Dullerud, Anvith Thudi, Varun Chandrasekaran, and Nicolas Papernot. Proof-of-Learning: Definitions and ...

    arXiv 2021

  4. [12]

    Caliptra

    URL https://www.rand.org/pubs/working_papers/WRA3056-1.html. KVM. Kernel Virtual Machine (KVM). URL https://linux-kvm.org/page/Main_Page. Thomas Kwa, Ben West, Joel Becker, Amy Deng, Katharyn Garcia, Max Hasin, Sami Jawhar, Megan Kinniment, Nate Rush, Sydney V on Arx, Ryan Bloom, Thomas Broadley, Haoxing Du, Brian Goodrich, Nikola Jurkovic, Luke Harold Mi...

    arXiv 2025

  5. [13]

    Kevin Roose

    URL https://www.theverge.com/2024/5/17/24159095/openai-jan-leike-superalignment- sam-altman-ai-safety. Kevin Roose. OpenAI Insiders Warn of a ‘Reckless’ Race for Dominance. The New York Times,

    work page 2024

  6. [14]

    Michael D

    URL https://www.nytimes.com/2024/06/04/technology/openai-culture-whistleblowers.html. Michael D. Rosenthal, Leslie G. Fishbone, Linda Gallini, Allan Krass, Myron Kratzer, Jonathan Sanborn, Warren Stern, Barclay Ward, and Norman A. Wulf.Deterring Nuclear Proliferation: The Importance of IAEA Safeguards. 2019. URL https://www.bnl.gov/nns/iaea-textbook/. Gir...

    arXiv 2024

  7. [15]

    Jaime Sevilla, Lennart Heim, Marius Hobbhahn, Tamay Besiroglu, Anson Ho, and Pablo Villalobos

    URL https://epoch.ai/blog/training-compute-of-frontier-ai-models-grows-by-4-5x-per-year. Jaime Sevilla, Lennart Heim, Marius Hobbhahn, Tamay Besiroglu, Anson Ho, and Pablo Villalobos. Estimating Training Compute of Deep Learning Models, 2022. URL https://epoch.ai/blog/estimati ng-training-compute. Agam Shah. NVIDIA Shipped 3.76 Million Data-center GPUs in...

    work page 2022

  8. [16]

    77 Yonadav Shavit

    URL https://www.hpcwire.com/2024/06/10/nvidia-shipped-3-76-million-data-center-gpus- in-2023-according-to-study/. 77 Yonadav Shavit. What does it take to catch a Chinchilla? Verifying Rules on Large-Scale Neural Network Training via Compute Monitoring, 2023. URL https://arxiv.org/abs/2303.11341. Matt Sheehan. Untitled Thread on X, July 2023. URL https://t...

    arXiv 2024

Show all 21 references
  1. [17]

    Anton Shilov

    URL https://www.rand.org/content/dam/rand/pubs/conf_proceedings/CFA3400/CFA3429- 1/RAND_CFA3429-1.pdf. Anton Shilov. Nvidia to Reportedly Triple Output of Compute GPUs in 2024: Up to 2 Million H100s,

    2024

  2. [18]

    Charles Slayman

    URL https://www.tomshardware.com/news/nvidia-to-reportedly-triple-output-of-compute- gpus-in-2024-up-to-2-million-h100s. Charles Slayman. Soft Error Trends and Mitigation Techniques in Memory Devices. Technical report, Ops A La Carte LLC, 2011. URL https://pdfs.semanticscholar...

    2024 arXiv

  3. [19]

    Trieu Trinh and Thang Luong

    URL https://openmined.org/blog/secure-enclaves-for-ai-evaluation/. Trieu Trinh and Thang Luong. AlphaGeometry: An Olympiad-level AI system for geometry, January

  4. [20]

    Two Sigma

    URL https://deepmind.google/discover/blog/alphageometry-an-olympiad-level-ai-system- for-geometry/. Two Sigma. A Workaround for Non-Determinism in TensorFlow, 2017. URL https://www.twosigma .com/articles/a-workaround-for-non-determinism-in-tensorflow/. Mark Tyson. Chinese engi...

    2001

  5. [21]

    James David Vance

    URL https://www.sec.gov/files/fy23-annual-report.pdf. James David Vance. Remarks by Vice President Vance at the Artificial Intelligence (AI) Action Summit, 2025. URL https://lk.usembassy.gov/remarks-by-vice-president-vance-at-the-artificial- intelligence-ai-action-summit/. Kap...

    2025 arXiv

  6. [1985]

    Gabriel Kulp, Daniel Gonzales, Everett Smith, Lennart Heim, Prateek Puri, Michael J

    URL https://www.sipri.org/sites/default/files/files/books/SIPRI85Krass/SIPRI85Krass.pdf. Gabriel Kulp, Daniel Gonzales, Everett Smith, Lennart Heim, Prateek Puri, Michael J. D. Vermeer, and Zev Winkelman. Hardware-Enabled Governance Mechanisms. Working paper, RAND, January

  7. [2014]

    URL https://doi.org/10.1109/jproc.2014.2334493

    doi: 10.1109/JPROC.2014.2334493. URL https://doi.org/10.1109/jproc.2014.2334493. Joseph R. Biden Jr. Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, 2023. URL https://www.whitehouse.gov/briefing-room/presidential- actions/20...

    2014

  8. [2015]

    Claire Dennis, Sam Manning, Stephen Clare, Boxi Wu, Jake Okechukwu Effoduh, Chinasa T

    URL https://theconversation.com/why-companies-have-little-incentive-to-invest-in- cybersecurity-37570. Claire Dennis, Sam Manning, Stephen Clare, Boxi Wu, Jake Okechukwu Effoduh, Chinasa T. Okolo, Lennart Heim, and Katya Klinova. Options and Motivations for International AI Be...

    2025 arXiv

  9. [2019]

    ISBN 978-1-939133-06-9

    USENIX Association. ISBN 978-1-939133-06-9. URL https://www.usenix.org/conference/ usenixsecurity19/presentation/de-oliveira-nunes. Gerrit De Vynck, Rachel Lerman, and Nitasha Tiku. Microsoft’s AI chatbot is going off the rails,

  10. [2021]

    Stuart Armstrong, Nick Bostrom, and Carl Shulman

    URL https://www.armscontrol.org/policy-white-papers/2021-02/reinforcing-global-norm- against-chemical-weapons-use. Stuart Armstrong, Nick Bostrom, and Carl Shulman. Racing to the Precipice: A Model of Artificial Intelligence Development. Technical Report 2013-1, Future of Huma...

    2021 arXiv

  11. [2023]

    Benjamin Dean

    URL https://www.washingtonpost.com/technology/2023/02/16/microsoft-bing-ai-chatbot- sydney/. Benjamin Dean. Why companies have little incentive to invest in cybersecurity. The Conversation,

    2023

  12. [2024]

    Predictable

    URL https://epoch.ai/blog/introducing-the-distributed-training-interactive-simulator. Ege Erdil and David Schneider-Joseph. Data movement limits to frontier model training, 2024. URL https://arxiv.org/abs/2411.01137. European Union. AI Act. URL https://digital-strategy.ec.euro...

    2024 arXiv

  13. [2025]

    Revision date: 2025-04-14

    URL https://arxiv.org/abs/2503.05628. Revision date: 2025-04-14. 72 Will Henshall. The 3 Most Important AI Policy Milestones of 2023, 2023. URL https://time.com/651 3046/ai-policy-developments-2023/. Will Henshall. The Billion-Dollar Price Tag of Building AI, 2024. URL https:/...

    2025 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.