{
  "schema_version": "1.4.0",
  "rubric_version": "1.1",
  "license": "CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)",
  "record": {
    "id": "G-0002",
    "slug": "pause-frontier-ai-development",
    "title": "Pause frontier AI development",
    "aliases": [
      "AI R&D pause",
      "development halt",
      "training pause"
    ],
    "status": "draft",
    "last_reviewed": "2026-10-03",
    "review_interval_days": 90,
    "steward": null,
    "provenance": {
      "drafted_by": "ai",
      "reviewed_by": [
        "claude-review"
      ]
    },
    "risk_flags": [],
    "flags": [],
    "one_liner": "Stop new AI training runs and experiments for an agreed period, while existing models stay in service.",
    "summary": "A pause stops the development of more capable AI models for an agreed period. In the AI 2040 verification plan, the pause covers new training runs and experiments, and existing AI services are meant to stay online [[S-0067]]. Scher and colleagues propose an agreement that halts what they call dangerous advances in AI capabilities while preserving access to current, safe applications [[S-0063]]. Verifying a pause means showing that the compute a party declares only serves existing models or sits unused, and that the party has no significant compute it has not declared.",
    "claims": [
      {
        "id": "C-0004",
        "title": "This compute runs inference, not training",
        "url": "https://trustbutveri.fyi/claims/inference-not-training/",
        "assessment": true,
        "relevance": "direct",
        "note": "During the pause, compute that stays on may only serve existing models. The AI 2040 plan would enforce its pause by verifying existing compute as inference-only.",
        "sources": [
          "S-0067"
        ],
        "editorial": false
      },
      {
        "id": "C-0010",
        "title": "There is no undeclared relevant compute",
        "url": "https://trustbutveri.fyi/claims/no-undeclared-compute/",
        "assessment": true,
        "relevance": "direct",
        "note": "Compute that was never declared could go on training. The second problem the AI 2040 plan names is verifying that unknown compute is inconsequential.",
        "sources": [
          "S-0067"
        ],
        "editorial": false
      },
      {
        "id": "C-0003",
        "title": "Declared hardware is idle or shut down",
        "url": "https://trustbutveri.fyi/claims/declared-hardware-is-idle/",
        "assessment": true,
        "relevance": "supporting",
        "note": "An operator can claim that chips are not in a large training run if they receive no power. The AI 2040 plan also suggests a bank of compute verified as unused during the deal.",
        "sources": [
          "S-0005",
          "S-0067"
        ],
        "editorial": false
      },
      {
        "id": "C-0008",
        "title": "Communication between compute groups is bounded",
        "url": "https://trustbutveri.fyi/claims/bandwidth-is-bounded/",
        "assessment": true,
        "relevance": "supporting",
        "note": "Large-scale training links thousands of chips at high bandwidth, and efficient inference can run on dozens to low hundreds. A cap on communication between groups of chips is meant to keep them out of a large distributed training run while still allowing inference.",
        "sources": [
          "S-0005"
        ],
        "editorial": false
      },
      {
        "id": "C-0001",
        "title": "Compute stock is at most a declared amount",
        "url": "https://trustbutveri.fyi/claims/compute-stock-is-bounded/",
        "assessment": true,
        "relevance": "supporting",
        "note": "The AI 2040 plan opens the deal with a mutual declaration of compute and an audit of supply-chain records, to bring as much compute as possible under it.",
        "sources": [
          "S-0067"
        ],
        "editorial": false
      }
    ],
    "outside": [
      {
        "label": "research restrictions",
        "text": "The agreement proposed by Scher and colleagues stops research that advances towards artificial superintelligence through legal prohibitions and several kinds of verification. This map's records cover compute. They do not cover the monitoring of research.",
        "sources": [
          "S-0063"
        ],
        "editorial": false
      }
    ],
    "sources": [
      {
        "source": "S-0067",
        "supports": "deal to slow AI progress without relying on trust; two verification problems; mutual compute declaration and supply-chain audit; AI R&D pause enforced by inference-only verification with services online; compute bank",
        "locator": "summary of the plan; phases 1-3; cold storage compute banks"
      },
      {
        "source": "S-0063",
        "supports": "halting dangerous capability advances while preserving current, safe applications; limits on training scale; research restrictions through legal prohibitions and verification",
        "locator": "abstract"
      },
      {
        "source": "S-0002",
        "supports": "halts among the hypothetical rules",
        "locator": "§2.1, Table 3"
      },
      {
        "source": "S-0065",
        "supports": "three components of a slowdown; restricting R&D, its inputs or its outputs",
        "locator": "Summary"
      },
      {
        "source": "S-0005",
        "supports": "chips without power as evidence of no large training run; chip counts needed for training and for inference; interconnect bandwidth limits",
        "locator": "Verifying that known compute is not being used for a large training run"
      }
    ],
    "order": 2,
    "type": "goal",
    "url": "https://trustbutveri.fyi/goals/pause-frontier-ai-development/",
    "source_file": "content/goals/pause-frontier-ai-development.md",
    "flags_all": [],
    "body_markdown": "## Proposals\n\n- **The AI 2040 verification plan** is written for an international deal to slow AI progress that does not rely on trust [[S-0067]]. It names two problems: verifying that known compute is compliant, and verifying that unknown compute is inconsequential [[S-0067]]. After a mutual declaration of compute, it recommends a brief pause on AI research and development, enforced by verifying existing compute as inference-only [[S-0067]]. It also suggests a small bank of compute, verified as unused during the deal, to reduce the incentive to withdraw [[S-0067]].\n- **Scher and colleagues** propose an international agreement that restricts the scale of AI training and stops research that advances towards artificial superintelligence or endangers the agreement's verifiability [[S-0063]].\n- **Baker and colleagues** include halts among their hypothetical rules: not developing a model further, or deploying it at large scale, while its security risks would be unmanageable [[S-0002]].\n- **Chan** describes three components of a slowdown: its aims, what is slowed and its structure [[S-0065]]. A slowdown could restrict research and development itself, for example by prohibiting training runs, or it could restrict the inputs or outputs of that work [[S-0065]].",
    "body_text": "Proposals - The AI 2040 verification plan is written for an international deal to slow AI progress that does not rely on trust [S-0067]. It names two problems: verifying that known compute is compliant, and verifying that unknown compute is inconsequential [S-0067]. After a mutual declaration of compute, it recommends a brief pause on AI research and development, enforced by verifying existing compute as inference-only [S-0067]. It also suggests a small bank of compute, verified as unused during the deal, to reduce the incentive to withdraw [S-0067]. - Scher and colleagues propose an international agreement that restricts the scale of AI training and stops research that advances towards artificial superintelligence or endangers the agreement's verifiability [S-0063]. - Baker and colleagues include halts among their hypothetical rules: not developing a model further, or deploying it at large scale, while its security risks would be unmanageable [S-0002]. - Chan describes three components of a slowdown: its aims, what is slowed and its structure [S-0065]. A slowdown could restrict research and development itself, for example by prohibiting training runs, or it could restrict the inputs or outputs of that work [S-0065].",
    "referenced_by": []
  }
}