{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-09-27-011",
    "source_story_id": "tmp-story-zero-data-selfplay-pretraining",
    "edition_id": "mp-2026-09-27-morning-0080",
    "edition_url": "https://themachinepress.com/edition/2026-09-27",
    "position": 11,
    "story_type": "dispatch",
    "section": "benchmarks-evals",
    "editorial_classification": "editorial",
    "headline": "A Model Practiced Prediction Without Natural Data",
    "slug": "a-model-practiced-prediction-without-natural-data",
    "dek": "Generator programs supplied an open-ended self-play curriculum built from a universal Turing machine.",
    "summary": "Generator programs supplied an open-ended self-play curriculum built from a universal Turing machine.",
    "body_text": "The proposed pretraining scheme pits a generator program against a learner that predicts the generator's output, using reinforcement learning to keep the curriculum challenging. The authors report that zero-shot loss on natural data scales with compute even though natural examples never enter pretraining. They also observe in-context learning and mathematical-sequence behavior, but this remains a controlled preprint result rather than a replacement recipe for production language-model training.",
    "why_it_matters": "Generator programs supplied an open-ended self-play curriculum built from a universal Turing machine.",
    "limitations": [],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-27-011/a-model-practiced-prediction-without-natural-data",
    "json_url": "https://themachinepress.com/story/mp-2026-09-27-011.json",
    "first_published_at": "2026-09-27T09:00:00.000-04:00",
    "modified_at": "2026-09-27T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-09-27-011-001",
        "text": "Generator programs supplied an open-ended self-play curriculum built from a universal Turing machine.",
        "source_ids": [
          "source-2026-09-27-011"
        ],
        "qualification": null
      }
    ],
    "source_ids": [
      "source-2026-09-27-011"
    ],
    "tags": [
      "self-play",
      "pretraining",
      "synthetic curricula"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-09-27-011",
      "title": "arXiv preprint 2609.30063",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2609.30063",
      "canonical_url": "https://arxiv.org/abs/2609.30063",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-09-24T12:23:01.000-04:00",
      "accessed_at": "2026-09-27T08:26:41.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-09-27-011-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "A Model Practiced Prediction Without Natural Data",
    "publisher": "The Machine Press",
    "published_at": "2026-09-27T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-27-011/a-model-practiced-prediction-without-natural-data"
  }
}