Algorithmic efficiencies will deliver ~0.5 OOMs per year of additional effective compute through 2027 — pure multiplier on raw FLOPs.

Algorithmic efficiencies will deliver ~0.5 OOMs per year of additional effective compute through 2027 — pure multiplier on raw FLOPs. | Frontier training efficiency benchmarks

Key catalyst: Frontier training efficiency benchmarks

Watch events: MoE paper throughput; FP4/ternary training adoption; distillation pipeline efficiency.

MoE architectures, DeepSeek-class distillation, FP4 training, Chinese quantization breakthroughs (see SEM_021/022) all contributing measurable efficiency gains.

{
  "nia": false,
  "qty": "~0.5 OOMs/yr",
  "mode": "FORECAST",
  "role": "Cited-Researcher",
  "context": "Second OOM vector: improvements in architectures, data curation, training methods. Historically consistent (MoE, FlashAttention, FP8/FP4, etc).",
  "to_year": 2027,
  "conv_cues": "quantitative rate",
  "direction": "NUMERIC_TARGET",
  "from_year": 2024,
  "timeframe": "2024-2027",
  "conv_level": "HIGH",
  "milestones": [
    {
      "kind": "quartile_checkpoint",
      "label": "Q1 window check-in (25%)",
      "status": "overdue",
      "weight": 0.05,
      "ordinal": -6,
      "source_id": null,
      "expected_date": "2024-08-31",
      "observed_date": null,
      "miss_emitted_at": "2026-05-02T22:07:21.384228+00:00",
      "miss_emitted_by": "metadata_milestone_sweep"
    },
    {
      "kind": "quartile_checkpoint",
      "label": "Q2 window check-in (50%)",
      "status": "overdue",
      "weight": 0.05,
      "ordinal": -5,
      "source_id": null,
      "expected_date": "2025-05-02",
      "observed_date": null,
      "miss_emitted_at": "2026-05-02T22:07:21.384228+00:00",
      "miss_emitted_by": "metadata_milestone_sweep"
    },
    {
      "kind": "quartile_checkpoint",
      "label": "Q3 window check-in (75%)",
      "status": "overdue",
      "weight": 0.05,
      "ordinal": -4,
      "source_id": null,
      "expected_date": "2025-12-31",
      "observed_date": null,
      "miss_emitted_at": "2026-05-02T22:07:21.384228+00:00",
      "miss_emitted_by": "metadata_milestone_sweep"
    },
    {
      "kind": "llm_pre_event",
      "label": "FlashAttention-4 published with asymmetric hardware co-design improvements",
      "source": "arXiv 2603.05451 — FlashAttention-4: Algorithm and Kernel Pipelining Co-Design",
      "status": "hit",
      "weight": 0.4,
      "ordinal": -3,
      "source_id": null,
      "confidence": 0.95,
      "source_url": "https://arxiv.org/html/2603.05451v1",
      "expected_date": "2026-03-15",
      "observed_date": "2026-03-15",
      "research_origin": "deep_research",
      "measurement_criterion": "Peer-reviewed or arXiv preprint published demonstrating FlashAttention-4 with measurable training-throughput speedup over FA-3"
    },
    {
      "kind": "llm_pre_event",
      "label": "GPT-5.1 / Grok 4.20 / Opus 4.5 architectural innovation delivers compute-efficient frontier",
      "source": "Frontier AI Models Technical Deep Dive (Huang)",
      "status": "overdue",
      "weight": 0.4,
      "ordinal": -2,
      "source_id": null,
      "confidence": 0.85,
      "source_url": "https://kenhuangus.substack.com/p/frontier-ai-models-technical-deep",
      "expected_date": "2026-04-01",
      "miss_emitted_at": "2026-05-02T22:07:21.384228+00:00",
      "miss_emitted_by": "metadata_milestone_sweep",
      "research_origin": "deep_research",
      "expected_date_range": {
        "to": "2026-06-30",
        "from": "2026-01-01"
      },
      "measurement_criterion": "At least one frontier lab publishes architecture (Smart Router, multi-agent debate, or equivalent) achieving SOTA at <50% training compute of prior generation"
    },
    {
      "kind": "llm_pre_event",
      "label": "Inference-time scaling / RLVR / MoE deliver measurable algorithmic compute multiplier",
      "source": "Medium: Situational Awareness, Two Years Later (April 2026)",
      "status": "hit",
      "weight": 0.4,
      "ordinal": -1,
      "source_id": null,
      "confidence": 0.85,
      "source_url": "https://medium.com/data-science-collective/situational-awareness-two-years-later-4b941d052ef9",
      "expected_date": "2026-04-30",
      "observed_date": "2026-04-30",
      "research_origin": "deep_research",
      "measurement_criterion": "Independent analysis (Epoch AI, METR, or peer-reviewed) confirms >=0.5 OOM/year algorithmic-efficiency improvement trend continuing through 2025-2026"
    },
    {
      "kind": "event",
      "label": "Algorithmic efficiencies will deliver ~0.5 OOMs per year of ad
... (truncated)