GPT5-caliber model could shrink to 30-40B parameters, potentially as low as 1-2B or millions.

GPT5-caliber model could shrink to 30-40B parameters, potentially as low as 1-2B or millions. | Alex, you said this. I think you're the first person I ever heard from it saying, look, you know, the equivalent of a GPT5 is going to be maybe 30 40 billion parameters, but it could get as low as one or two billion truly because right now when they train that caliber of model, it has all this junk knowledge in it too.

Alex, you said this. I think you're the first person I ever heard from it saying, look, you know, the equivalent of a GPT5 is going to be maybe 30 40 billion parameters, but it could get as low as one or two billion truly because right now when they train that caliber of model, it has all this junk knowledge in it too.

Beneficiaries (24)

Adverse (6)

{
  "nia": false,
  "qty": "30-40B parameters, potentially 1-2B or millions",
  "url": "https://www.youtube.com/watch?v=T8X6kp-pcKs",
  "mode": "PREDICTION",
  "role": "Host",
  "caveats": "After stripping junk knowledge",
  "context": "look, you know, the equivalent of a GPT5 is going to be maybe 30 40 billion parameters, but it could get as low as one or two billion truly because right now when they train that caliber of model, it has all this junk knowledge in it too. I I could imagine scenarios where it's only a few million parameter equivalent that's sort of the core micro kernel of AGI or super intelligence",
  "to_year": 2028,
  "cited_by": "Dave Blundin",
  "verbatim": "Alex, you said this. I think you're the first person I ever heard from it saying, look, you know, the equivalent of a GPT5 is going to be maybe 30 40 billion parameters, but it could get as low as one or two billion truly because right now when they train that caliber of model, it has all this junk knowledge in it too.",
  "conv_cues": "going to be; could get",
  "direction": "DOWN",
  "from_year": 2026,
  "timeframe": "unspecified future",
  "conv_level": "MEDIUM",
  "milestones": [
    {
      "kind": "prereq",
      "label": "Nvidia became the world's first $5 trillion company (late 2025), operating a near-monopoly on advanced AI chips.",
      "status": "hit",
      "weight": 0.5,
      "ordinal": -8,
      "source_id": "SEM_011",
      "expected_date": "2026-04-29",
      "observed_date": "2026-04-29"
    },
    {
      "kind": "prereq",
      "label": "Nvidia Data Center revenue +66% YoY, contributing ~90% of $57B fiscal Q3 revenue; >$4.5T market cap entirely underpinned by AI silicon.",
      "status": "hit",
      "weight": 0.5,
      "ordinal": -7,
      "source_id": "SEM_027",
      "expected_date": "2026-04-29",
      "observed_date": "2026-04-29"
    },
    {
      "kind": "prereq",
      "label": "Nvidia's Arizona-based TSMC factory successfully fabricated cutting-edge semiconductors on US soil for first time in decades (October 2025).",
      "status": "hit",
      "weight": 0.5,
      "ordinal": -6,
      "source_id": "SEM_014",
      "expected_date": "2026-04-29",
      "observed_date": "2026-04-29"
    },
    {
      "kind": "prereq",
      "label": "Blackwell RTX PRO 5000 (72GB) engineered with 50% memory boost over previous generation — deliberate architectural concession for larger AI ",
      "status": "hit",
      "weight": 0.5,
      "ordinal": -5,
      "source_id": "SEM_029",
      "expected_date": "2026-04-29",
      "observed_date": "2026-04-29"
    },
    {
      "kind": "prereq",
      "label": "Nvidia quadrupled chip production output while only doubling human headcount — achieved by deploying AI coding tools (Cursor, Claude Code) a",
      "status": "hit",
      "weight": 0.5,
      "ordinal": -4,
      "source_id": "SEM_012",
      "expected_date": "2026-04-29",
      "observed_date": "2026-04-29"
    },
    {
      "kind": "llm_pre_event",
      "label": "Open-weights 30-40B model matches GPT-5 on aggregated benchmarks",
      "source": "deep_research_enriched",
      "status": "overdue",
      "weight": 0.4,
      "ordinal": -3,
      "source_id": null,
      "confidence": 0.6,
      "source_url": "https://www.marktechpost.com/2026/02/01/nvidia-ai-brings-nemotron-3-nano-30b-to-nvfp4-with-quantization-aware-distillation-qad-for-efficient-reasoning-inference/",
      "expected_date": "2026-05-01",
      "miss_emitted_at": "2026-05-09T22:14:10.596691+00:00",
      "miss_emitted_by": "metadata_milestone_sweep",
      "research_origin": "training",
      "expected_date_range": {
        "to": "2026-09-30",
        "from": "2025-12-01"
      },
      "measurement_criterion": "An open or proprietary 30-40B parameter model achieves parity (within 2pp) with GPT-5 on MMLU-Pro / GPQA / HumanEval composite."
    },
    {
      "kind": "llm_pre_event",
      "label": "Sub-10B parameter model achieves prior-frontier (GPT-4 class) per
... (truncated)