{
  "question_id": "W18",
  "slug": "lab-runs-100000-research-instances",
  "title": "A lab runs 100,000 model instances on its own research",
  "type": "binary",
  "resolution": {
    "levels": [],
    "values": [],
    "annotations": []
  },
  "forecasts": [
    {
      "question_id": "W18",
      "forecast_date": "2026-09-20",
      "evidence_cutoff": "2026-09-20",
      "method": "wide-v2",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-09-30",
          "p_yes": 0.015
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.1332
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.2806
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.4173
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.528
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.6129
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.6749
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.7204
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.7539
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.7785
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.7963
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.8115
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.8247
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.8361
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.846
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.8544
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.8617
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.8679
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.8732
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.8783
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.8832
        }
      ],
      "passes": []
    },
    {
      "question_id": "W18",
      "forecast_date": "2026-09-27",
      "evidence_cutoff": "2026-09-27",
      "method": "wide-v2",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-09-30",
          "p_yes": 0.008
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.127
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.267
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.399
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.507
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.591
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.652
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.701
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.74
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.771
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.796
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.817
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.833
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.846
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.857
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.866
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.874
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.881
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.887
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.893
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.898
        }
      ],
      "passes": []
    },
    {
      "question_id": "W18",
      "forecast_date": "2026-10-01",
      "evidence_cutoff": "2026-10-01",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": "Pooled platform counts (e.g., Anthropic's 'agents at any one time on our most-used platform') and sandbox-infrastructure counts (DeepSeek DSec) are treated as non-qualifying for lack of single-model attribution; RL training rollouts are treated as doubtful 'research work'; 'on the order of 100,000 concurrent agents' of one model is treated as an approximate point estimate that likely qualifies.",
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-09-30",
          "p_yes": 0.03
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.079
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.157
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.241
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.325
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.399
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.465
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.521
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.569
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.61
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.645
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.675
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.701
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.724
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.743
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.76
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.774
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.787
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.797
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.807
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.817
        }
      ],
      "passes": []
    },
    {
      "question_id": "W18",
      "forecast_date": "2026-10-08",
      "evidence_cutoff": "2026-10-08",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "summary": "Labs have already disclosed running about 10,000 (OpenAI, on one problem) and 30,000 (Anthropic, platform-wide) AI agents at once, and usage is growing fast. So a 100,000-copy statement is fairly likely within a few years. The rule that the count must be tied to one specific model and to research holds the odds down, as do pauses and possible secrecy. The biggest swing would be Anthropic's next update showing 100,000 or more with a named model.",
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-12-31",
          "p_yes": 0.04
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.107
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.188
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.261
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.327
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.381
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.431
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.473
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.513
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.54
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.565
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.587
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.607
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.625
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.642
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.656
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.67
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.681
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.693
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.702
        },
        {
          "period_end": "2031-12-31",
          "p_yes": 0.711
        }
      ],
      "passes": []
    },
    {
      "question_id": "W18",
      "forecast_date": "2026-10-15",
      "evidence_cutoff": "2026-10-15",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "summary": "I put the chance at about 34% by the end of 2027 and about 71% by the end of 2031. In September 2026 OpenAI reported about 10,000 agents working at once on one math problem, and Anthropic about 30,000 on its internal platform. So 100,000 looks reachable within a year or two. The main obstacle is that most disclosures don't tie the count to a single model. A lab report naming one model running 100,000 or more copies at once on research would settle it.",
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-12-31",
          "p_yes": 0.04
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.107
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.188
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.269
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.342
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.404
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.458
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.499
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.534
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.562
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.586
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.607
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.624
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.639
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.652
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.664
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.674
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.684
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.692
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.7
        },
        {
          "period_end": "2031-12-31",
          "p_yes": 0.707
        }
      ],
      "passes": []
    }
  ],
  "declined": []
}