{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-10-08-002",
    "source_story_id": "tmp-feature-fault-tolerant-foundation-models",
    "edition_id": "mp-2026-10-08-morning-0090",
    "edition_url": "https://themachinepress.com/edition/2026-10-08",
    "position": 2,
    "story_type": "secondary",
    "section": "chips-infrastructure",
    "editorial_classification": "editorial",
    "headline": "The Larger Model Learned to Compute Around Broken Hardware",
    "slug": "the-larger-model-learned-to-compute-around-broken-hardware",
    "dek": "Training runs on simulated faulty digital hardware suggest that error resilience can rise with model scale, opening a speculative route to lower-energy inference.",
    "summary": "Training runs on simulated faulty digital hardware suggest that error resilience can rise with model scale, opening a speculative route to lower-energy inference.",
    "body_text": "Efficient hardware often becomes less reliable as voltage and safety margins fall. A study built modified neural scaling laws from 40,000 GPU-hours of language-model training with simulated digital faults and found that appropriately trained models became more resilient as they grew rather than more fragile. The authors interpret the behavior as learned computation inside good error-correcting codes whose relative overhead need not expand with model size. That interpretation remains a conjecture: the experiments simulate faulty hardware, do not establish formal fault tolerance and do not demonstrate an energy-saving production chip. Yet the result reframes the systems bargain. Instead of forcing every transistor to behave perfectly and treating model computation as brittle, future hardware and training might divide the reliability work between them. The immediate contribution is evidence that scale can absorb faults under training—not proof that unreliable accelerators are ready for deployment.",
    "why_it_matters": "Training runs on simulated faulty digital hardware suggest that error resilience can rise with model scale, opening a speculative route to lower-energy inference.",
    "limitations": [
      "The authors interpret the behavior as learned computation inside good error-correcting codes whose relative overhead need not expand with model size.",
      "That interpretation remains a conjecture: the experiments simulate faulty hardware, do not establish formal fault tolerance and do not demonstrate an energy-saving production chip.",
      "The immediate contribution is evidence that scale can absorb faults under training—not proof that unreliable accelerators are ready for deployment."
    ],
    "importance": 9,
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-08-002/the-larger-model-learned-to-compute-around-broken-hardware",
    "json_url": "https://themachinepress.com/story/mp-2026-10-08-002.json",
    "first_published_at": "2026-10-08T09:00:00.000-04:00",
    "modified_at": "2026-10-08T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-10-08-002-001",
        "text": "Training runs on simulated faulty digital hardware suggest that error resilience can rise with model scale, opening a speculative route to lower-energy inference.",
        "source_ids": [
          "source-2026-10-08-002"
        ],
        "qualification": "The authors interpret the behavior as learned computation inside good error-correcting codes whose relative overhead need not expand with model size."
      }
    ],
    "source_ids": [
      "source-2026-10-08-002"
    ],
    "tags": [
      "fault tolerance",
      "language models",
      "energy efficiency"
    ],
    "image_url": "https://themachinepress.com/issues/2026-10-08/feature-fault-tolerant-foundation-models.png",
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-10-08-002",
      "title": "arXiv preprint 2610.10311",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2610.10311",
      "canonical_url": "https://arxiv.org/abs/2610.10311",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-10-07T12:05:48.000-04:00",
      "accessed_at": "2026-10-08T08:26:48.909-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-10-08-002-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "The Larger Model Learned to Compute Around Broken Hardware",
    "publisher": "The Machine Press",
    "published_at": "2026-10-08T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-08-002/the-larger-model-learned-to-compute-around-broken-hardware"
  }
}