{
  "question_id": "W10",
  "slug": "lab-runs-design-its-staff-cannot-understand",
  "title": "A lab's best system runs on an AI design its staff don't fully understand",
  "type": "binary",
  "resolution": {
    "levels": [],
    "values": [],
    "annotations": []
  },
  "forecasts": [
    {
      "question_id": "W10",
      "forecast_date": "2026-09-20",
      "evidence_cutoff": "2026-09-20",
      "method": "wide-v2",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": "A substantive AI-designed inference algorithm or integrated agent runtime can qualify, but generic model opacity, unexplained benchmark gains alone, and ordinary AI-written implementation cannot substitute for an explicit collective lack of adequate explanation about that design.",
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-09-30",
          "p_yes": 0.003
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.015
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.0327
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.054
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.0786
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.1062
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.1357
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.166
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.1968
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.2281
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.259
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.2887
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.3171
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.3444
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.3706
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.3958
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.42
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.4432
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.4655
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.4868
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.5074
        }
      ],
      "passes": []
    },
    {
      "question_id": "W10",
      "forecast_date": "2026-09-27",
      "evidence_cutoff": "2026-09-27",
      "method": "wide-v2",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": "An AI-designed agent harness or inference algorithm can qualify as an integrated computational design if it governs the explicitly identified overall most capable system, but AI-written implementation or generic model opacity alone cannot.",
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-09-30",
          "p_yes": 0.002
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.01
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.0219
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.0375
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.0568
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.0794
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.1052
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.1338
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.165
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.1984
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.2305
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.2612
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.2908
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.3192
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.3498
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.3791
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.407
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.4337
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.4592
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.4835
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.5067
        }
      ],
      "passes": []
    },
    {
      "question_id": "W10",
      "forecast_date": "2026-09-30",
      "evidence_cutoff": "2026-09-30",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": "Counted developer phrasing like 'our most capable model' (including internal models in risk reports) as the most-capable identification; required an explicit admission that staff lack an explanation of how/why a specific AI-designed architecture/algorithm (or substantive part) works, excluding generic black-box, code-review, or verification-only statements.",
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-09-30",
          "p_yes": 0.001
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.006
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.013
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.023
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.035
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.05
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.066
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.083
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.102
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.121
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.139
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.157
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.175
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.192
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.207
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.221
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.235
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.249
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.261
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.273
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.284
        }
      ],
      "passes": []
    },
    {
      "question_id": "W10",
      "forecast_date": "2026-10-07",
      "evidence_cutoff": "2026-10-07",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "summary": "AI labs now use AI to do much of their own research, but no lab has said its top system runs on an AI-made design its own staff cannot explain. Labs have strong reasons not to admit that, and the conditions are strict. I put the chance at about 4% by the end of 2027, rising to about 20% by the end of 2031. The development that would most raise it: a major lab credits a core part of its flagship model to its AI and calls it not yet understood.",
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-12-31",
          "p_yes": 0.004
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.011
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.02
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.03
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.04
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.053
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.066
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.079
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.092
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.104
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.116
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.126
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.137
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.145
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.154
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.162
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.171
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.178
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.185
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.193
        },
        {
          "period_end": "2031-12-31",
          "p_yes": 0.2
        }
      ],
      "passes": []
    },
    {
      "question_id": "W10",
      "forecast_date": "2026-10-14",
      "evidence_cutoff": "2026-10-14",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "summary": "I put the chance at about one in five by the end of 2031, and under 4% by the end of 2027. No major AI developer has said anything like this yet. Labs have strong reasons not to admit they run designs they don't understand, and the wording required is strict. But AI is rapidly taking over research at these companies, which are unusually open about it. The biggest change would be a lab announcing that an AI-discovered architecture powers its top model.",
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-12-31",
          "p_yes": 0.004
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.009
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.016
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.026
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.036
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.047
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.059
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.071
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.083
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.096
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.108
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.12
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.131
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.142
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.153
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.163
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.173
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.183
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.192
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.201
        },
        {
          "period_end": "2031-12-31",
          "p_yes": 0.21
        }
      ],
      "passes": []
    }
  ],
  "declined": []
}