{
  "question_id": "W16",
  "slug": "frontier-model-learns-continuously",
  "title": "A frontier model updates its own weights continuously in use",
  "type": "binary",
  "resolution": {
    "levels": [],
    "values": [],
    "annotations": []
  },
  "forecasts": [
    {
      "question_id": "W16",
      "forecast_date": "2026-09-20",
      "evidence_cutoff": "2026-09-20",
      "method": "wide-v2",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": "I count explicitly documented session-local adapter or neural-memory parameter learning, but do not treat ordinary recurrent-state updates as qualifying merely because they can be mathematically reinterpreted as fast weights.",
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-09-30",
          "p_yes": 0.01
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.0447
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.0829
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.1287
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.181
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.2383
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.2993
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.3553
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.4069
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.4543
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.498
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.5382
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.5751
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.6091
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.6365
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.6619
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.6856
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.7076
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.7251
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.7416
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.7571
        }
      ],
      "passes": []
    },
    {
      "question_id": "W16",
      "forecast_date": "2026-09-27",
      "evidence_cutoff": "2026-09-27",
      "method": "wide-v2",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": "I treat changes to recurrent attention state alone as excluded activation changes rather than learned-parameter updates, while allowing explicit session-local fast-weight or integrated neural-memory learning.",
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-09-30",
          "p_yes": 0.005
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.0398
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.083
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.1335
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.1898
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.2506
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.3105
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.3691
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.4259
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.4776
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.5272
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.5721
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.6128
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.6496
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.6811
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.7098
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.7345
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.7557
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.774
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.7898
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.8035
        }
      ],
      "passes": []
    },
    {
      "question_id": "W16",
      "forecast_date": "2026-09-29",
      "evidence_cutoff": "2026-09-29",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": "Per-session test-time-training or neural-memory modules count only if the developer explicitly describes them as updating weights or parameters during inference; linear-attention recurrent states (GDN/KDA) described only as attention state, and automated periodic checkpoint pipelines (e.g. Cursor-style real-time RL), do not count.",
      "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": 0.023
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.047
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.071
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.099
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.126
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.157
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.186
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.215
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.242
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.269
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.294
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.319
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.343
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.363
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.382
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.4
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.418
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.436
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.453
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.469
        }
      ],
      "passes": []
    },
    {
      "question_id": "W16",
      "forecast_date": "2026-10-06",
      "evidence_cutoff": "2026-10-06",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "summary": "No leading AI company has yet launched a top model that changes its own internal settings while people use it. Today's assistants \"learn\" by saving notes and memories instead. Research on models that train themselves as they work is moving fast, and one of the ten companies could plausibly ship such a model within five years. Cost, safety worries and secrecy about designs may delay it or keep it undisclosed. A convincing launch of such a model by a rival lab would raise this forecast the most.",
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-12-31",
          "p_yes": 0.008
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.026
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.047
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.071
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.097
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.126
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.156
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.186
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.215
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.244
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.271
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.297
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.322
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.345
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.368
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.389
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.41
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.429
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.447
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.465
        },
        {
          "period_end": "2031-12-31",
          "p_yes": 0.482
        }
      ],
      "passes": []
    },
    {
      "question_id": "W16",
      "forecast_date": "2026-10-13",
      "evidence_cutoff": "2026-10-13",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "summary": "No major lab has yet released a top model that it says updates its own internal settings while in use. Every recent \"learns as it works\" feature from these labs relies on saved notes and memories, and that does not count. Research on models that adjust themselves while running is very active, and ten labs have chances over five years. My estimate is about 12% by end-2027 and 51% by end-2031. A flagship model adopting self-updating memory layers would most change this.",
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-12-31",
          "p_yes": 0.013
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.036
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.061
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.088
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.116
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.146
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.177
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.208
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.238
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.267
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.294
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.321
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.347
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.371
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.394
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.416
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.437
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.457
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.475
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.493
        },
        {
          "period_end": "2031-12-31",
          "p_yes": 0.51
        }
      ],
      "passes": []
    }
  ],
  "declined": []
}