← Evidence index · 中文 · Main post
Each section: Claim · Why · Evidence · Analogue · Would update if · Conf (H/M/L).
Parent: node12 rsi locus takeoff · Shared Ci spine
Supersedes: crosscut secondary cruxes §6
Correlates: Node 1 (cloud labor), Node 2 (bio DBTL), Node 4 (C10 timing), Node 7 (compute), Node 8 (cyber/finance), Hanson variant
Purpose: Claim | Why | Evidence | Analogue | Would update if | Conf — for every Node 12 probability.
1. Node definition — locus vs geometry
Claim: RSI locus (cloud vs bio vs embodied vs cyber/finance) determines which tail fires first by 2030; takeoff geometry (local foom vs Hanson absorption vs bio step-function) reweights my pdoom branch mix by ±3–6pp — a branch selector, not additive node delta.
Why: Node 7 = where FLOPs run; Node 12 = what domain compounds. Node 4 C10 = cloud-native superhuman AI researcher — bio/embodied/cyber may front-run C10 salience.
Evidence:
- node12 rsi locus takeoff — definition, four loci + L5
- crosscut secondary cruxes §6 — promoted 2026-07-04
- (internal note) §1–3 — RSI tiers A/B/C
Analogue: Pandemic origin debate — same global risk, different policy venue by locus (respiratory vs foodborne).
Would update if: Two loci hit C-tier within 6 mo with no ordering → L5 ↑.
Conf: H
2. P = 0.50 — L1 cloud/software RSI fires first (modal)
Claim: P(cloud/datacenter agent R&D loops produce first catastrophic policy-salient tail | by 2030) = 0.50 (range 0.42–0.58).
Why: Anthropic >80% prod code AI-written, 52× experiment speedup; METR horizon ~4 mo doubling; AI 2027 Agent-4 “genius country” beat = L1. Bio wet-lab closure months; embodied deployment years.
Evidence:
- (internal note) §2 — engineering + research metrics
- METR time horizons — 6–12h frontier 50%-horizon Feb–Mar 2026
- Node 12 master table — L1 0.50
Analogue: Software ate the world before robotics — digital compounding faster than physical.
Would update if: METR >1 week on AI R&D task + lab confirms successor experiment → L1 ↑ to 0.60+.
Conf: M–H
3. P = 0.22 — L2 bio lab / DBTL fires first
Claim: P(bio Tier 2→3 or autonomous DBTL stack produces first tail before cloud C-tier) = 0.22 (range 0.15–0.30).
Why: Different regulatory chokepoint (BMIA not compute); GeneBreaker + Evo-class FM; open-weight tamper (Node 5) accelerates bio path; screening delay if hidden near-misses (Node 10).
Evidence:
- node2 — Tier 2 window, Trigger E
- (internal note) — bio capability spine
- Node 12 — AI 2027 deliberate deviation 22% bio-first
Analogue: Nuclear first proliferation via different route than H-bomb timeline predicted.
Would update if: Documented Tier 3 attempt before C9 cloud RSI metrics → L2 ↑ to 0.35+.
Conf: M
4. P = 0.12 — L3 embodied / factory fires first
Claim: P(physical agent catastrophe or supply-chain embodied RSI headline before cloud C10 alignment scare) = 0.12 (range 0.07–0.18).
Why: Humanoids pilot-scale (<5,000 shift-work units by 2029); integration 12–24 mo/site; but injury/death more legible than API deception — T-PA2 embodied Trigger E2 path (Node 1).
Evidence:
- node physical ai embodied — verified deployment baseline
- (internal note) — NIST/CAISI embodied suite less mature than UK
- Node 12 conditional Ci table — L3 0.15 if capability stalls at C6
Analogue: Factory robot fatalities — local punctuations before “AGI” narrative.
Would update if: Humanoid fatality or >$100M loss headline before C10 → L3 ↑ to 0.20+.
Conf: M
5. P = 0.10 — L4 finance / critical infra fires first
Claim: P(agentic fraud, OT cascade, or Lobstar-class systemic near-miss headline first) = 0.10 (range 0.06–0.15).
Why: L4 capability A–B operational (Monterrey, Mythos, Interpol $442B fraud stat); policy venue = SEC/CFTC/export gating, not alignment whistleblower. Node 8 cyber/infra track.
Evidence:
- node8 cyber infra — capability/policy windows
- (internal note) — Lobstar Wilde
- Node 10 §18 — finance near-miss taxonomy gap
Analogue: Flash crash 2010 — finance first, AI narrative later.
Would update if: >$10B agentic fraud systemic near-miss before alignment scare → L4 ↑.
Conf: L–M
6. P = 0.06 — L5 multipolar simultaneous (<6 mo overlap)
Claim: P(≥2 loci hit C-tier within <6 months with no clear first — coordination-fail variant) = 0.06 (range 0.03–0.10).
Why: AI 2027 packs theft + alignment fail + race same year — narrative multipolar; empirical ordering usually exists. Reserved for true overlap (bio near-miss during cloud RSI essay month, etc.).
Evidence:
- Node 12 — L5 partition sums to 1.00
- (internal note) — simultaneous beats
- Hanson variant — anti-multipolar; L5 85% default / 15% Hanson mix
Analogue: 2020 COVID + financial shock — overlapping crises, hard to prioritize policy.
Would update if: Two loci C-tier Q same quarter 2027–28 with competing media cycles.
Conf: L–M
7. RSI tier A — scaffold self-edit (production)
Claim: Tier A (agent edits prompts/tools/repo; human sets goals) = production at all four loci in 2026 — necessary but insufficient for first-tail classification.
Why: Claude Code, bio agent protocols, warehouse AMR pilots, payment agents all A-tier live; Node 12 first-tail requires B→C escalation or harm at A/B.
Evidence:
- (internal note) §1 — tier definitions
- Node 12 RSI taxonomy table
- Anthropic RSI essay — >80% prod code = A/B not C
Analogue: Excel macros — automation without RSI.
Would update if: Lab publishes fully autonomous repo refactor without human review → A→B boundary blur.
Conf: H
8. RSI tier B — closed-loop optimization (SOTA + partial prod)
Claim: Tier B (auto-evaluator iterates; meta-framework human-fixed) = SOTA + partial prod on L1 cloud; A–B assist on L2 bio; A pilots on L3; A–B operational on L4.
Why: AlphaEvolve, 52× experiment loops (Anthropic); GeneBreaker + Evo FM on bio assist; embodied rare B-tier.
Evidence:
- Anthropic RSI §2.2 — 52× experiment speedup
- (internal note) §2–3
- OpenAI Preparedness — o3 not High on self-improvement (external check)
Analogue: AutoML — human defines search space.
Would update if: OpenAI/DeepMind mark internal model High on self-improvement per Preparedness Framework.
Conf: M–H
9. P = 0.12 — RSI tier C public by 2028
Claim: P(public evidence of Tier C — months-long autonomous AI R&D producing strictly better successor — by 2028) = 0.12 (range 0.06–0.20).
Why: No lab has published C-tier; Anthropic essay implies approaching B→C; OpenAI blueprint RSI indicators visibility not halt. 12% = tail, not modal.
Evidence:
- Node 12 load-bearing P #3
- (internal note) — C not public
- OpenAI Frontier Safety Blueprint (Jun 2026) — RSI indicators for CAISI
Analogue: Fusion breakeven — rumored internal milestones ≠ public proof.
Would update if: METR >1 week horizon on AI R&D task + lab confirms successor-training experiment.
Conf: L–M
10. P = 0.18 — bio Tier 3 before cloud C-tier
Claim: P(Tier 3 bio misuse attempt before cloud reaches published C-tier RSI) = 0.18 (range 0.12–0.26).
Why: Bio Tier 3 and cloud C-tier different evidence standards — screening failure visible before METR publishes week-long R&D horizon. Node 10 hidden near-miss ↑ conditional.
Evidence:
- Node 12 load-bearing P #4
- Node 2 — Tier 3 tail timing
- Node 10 — Tier 3 before screening 0.28 | high suppression (related channel)
Analogue: CRISPR DIY bio before GPT-4 — domain-specific tails decouple.
Would update if: Cloud C-tier public while bio still Tier 2 only → ↓ to <0.10.
Conf: L–M
11. P = 0.10 — embodied catastrophe before C10
Claim: P(kinetic embodied harm headline — fatality, major factory loss — before Node 4 C10 superhuman-AI-researcher salience) = 0.10 (range 0.06–0.15).
Why: Legibility of physical harm; GUARD/RESPONSIBLE pipeline (Node 1); still low vs cloud because deployment scale tiny.
Evidence:
- Node 1 §T-PA2 — embodied Trigger E2
- Node 12 load-bearing P #5
- Figure/Agility KPIs — pilots not fleets
Analogue: Uber AV pedestrian death 2018 — before “self-driving everywhere” narrative peak.
Would update if: ≥3 OEMs >500 shift-work units (T-PA1) → ↑ to 0.18+.
Conf: L–M
12. P = 0.14 — cyber/finance Trigger E before alignment Trigger E
Claim: P(Node 8-class CI/fraud emergency policy-salient before Node 4 alignment whistleblower/C10 scare) = 0.14 (range 0.10–0.20).
Why: L4 faster capability, slower kinetic escalation; Interpol fraud stats; export-control OT incidents — parallel track to alignment media cycle.
Evidence:
- node8 — Trigger taxonomy
- Node 12 cross-node matrix — N4 “CI panic competes”
- (internal note) — Fable-class recall (compute), adjacent
Analogue: SWIFT hacks — finance/infra salience without AGI frame.
Would update if: Grid OT agent incident headline Q before Oct 2027 AI 2027 whistleblower beat.
Conf: L–M
13. Anthropic RSI essay — primary L1 cloud evidence
Claim: Anthropic When AI Builds Itself (2026-06-04) is strongest public L1 signal: >80% prod code, 8× LOC, 52× experiments — same month Trump EO + GAAIA acceleration wins over conditional expect pause.
Why: Essay anchors cloud-first narrative; pause wording requires verification that does not exist — branch selector ↑ misalign if L1 resolves.
Evidence:
- (internal note) — §1 conditions, §2 metrics, §5.2 tension
- Node 11 §10 — E4 managed disclosure
- Node 12 AI 2027 mapping — Agent-4 Jun 2027 beat
Analogue: Moore’s Law press releases — capability disclosure ** accelerates** race narrative.
Would update if: Anthropic actually pauses named run citing RSI triggers.
Conf: H
14. METR ~4 mo doubling — cloud horizon driver
Claim: METR 50%-time-horizon doubling ~4 months (post-2024) binds L1 timing — C8–C10 cloud salience 2027 H2 – 2028 on hybrid track.
Why: Faster doubling → earlier L1 first-tail vs bio/embodied unless capability plateaus. AI 2027 Confirmed multipolar labs, Emerging on some agent beats.
Evidence:
- METR TH1.1 — 88.6-day doubling
- (internal note) — METR Ahead
- Node 1 — METR ≥8h already met
Analogue: GPU performance curve — predictable exponential until plateau.
Would update if: Doubling reverts to 7-month TH1.0 rate through 2027 → L1 ↓, Hanson ↑.
Conf: M–H
15. Hanson variant — L1 B-plateau path
Claim: If cloud RSI stalls at B-tier (Amdahl review bottleneck; no public C-tier by 2028), Hanson mixture weight ↑ to 55% default / 45% Hanson; P(extinction) 8–10% Hanson column.
Why: Anthropic essay admits code review bottleneck — human-in-loop limits C-tier; Hanson predicts institutional absorption not local FOOM.
Evidence:
- hanson variant human action — 60/40 default mix; misalign tail 25%→10%
- Anthropic RSI §3 — Amdahl’s law internal
- Node 12 Hanson table — L1 B-plateau 45% default / 55% Hanson
Analogue: Hanson–Yud 2008 — speed after AGI disputed, not whether agents improve.
Would update if: Single lab >50% frontier FLOPs + internal RSI >10× without sharing — Hanson falsified.
Conf: M
16. Hanson variant — L3/L4 first increases Hanson weight
Claim: If L3 embodied or L4 cyber/finance fires first, P(Hanson mix >50%) = 0.55 — slow takeoff, liability/insurance absorption.
Why: Physical/finance tails = whimper/severe before extinction; institutions already regulate OT/finance better than alignment.
Evidence:
- Node 12 Hanson integration table — L3 40/60 default/Hanson; L4 50/50
- Hanson variant node table — N8 modal = Hanson absorption
- Node 12 non-extinction reweight — L3 +3–5pp whimper
Analogue: Industrial robot safety standards — decades of OSHA absorption.
Would update if: Embodied incident triggers federal training moratorium — anti-Hanson.
Conf: M
17. Node 2 link — bio locus regulatory chokepoint
Claim: L2 bio tail uses BMIA/synthesis venue — orthogonal to Node 7 compute KYC/BIS — so bio can fire first while cloud FLOPs unchanged.
Why: GeneBreaker, Tier 2 emerging, screening politics (Node 2 P(BMIA)=0.45) — different Kingdon stream than agent labor (Node 1).
Evidence:
- node2 — BMIA, Trigger E, Tier 2→3
- Node 12 cross-node matrix — “Tier 3 after cloud C” vs “Front-run screening failure”
- Node 10 — hidden near-miss delays BMIA → ↑ L2 first-tail risk
Analogue: Fentanyl precursors — chemistry chokepoint ≠ internet chokepoint.
Would update if: BMIA passes before Tier 3 + cloud still B-tier → L2 ↓, extinction ↓ 2pp (Node 12 falsifier).
Conf: M
18. Node 7 link — compute co-location with L1
Claim: L1 cloud RSI co-locates with Node 7 triopoly (AWS/Azure/GCP) — same hyperscaler chokepoint for FLOPs and agent R&D loops; Node 7 ≠ Node 12 but correlates ρ≈0.5+ on L1 modal.
Why: Stargate, Korea 260k GPU, Genesis — ↑ total L1 capacity; export controls slow not stop B-tier. Embodied not on triopoly.
Evidence:
- node7 compute cloud — triopoly, KYC
- node7
- Node 12 — “Distinct from Node 7” explicit
Analogue: Cloud concentrates digital RSI; factories distribute embodied.
Would update if: Major lab trains >50% cycles off hyperscaler (sovereign-only) → L1/N7 decouple.
Conf: M
19. Node 8 link — cyber/finance parallel salience track
Claim: Node 8 fires capability 2025 H2–27; policy 2027–29 — can front-run Node 4 C10 without changing Ci cloud spine.
Why: Monterrey OT, Mythos agent fraud, Interpol stats — L4 first-tail 0.10; T8-B vetting tail overlaps Node 1 GUARD.
Evidence:
- node8 cyber infra — hybrid timing table
- node8
- Node 12 — L4 cyber/finance row
Analogue: Ransomware wave — cyber policy before AGI governance mature.
Would update if: Node 8 Trigger E before 2027-H1 with >2 mo media → L4 ↑.
Conf: M
20. Branch reweight — L1 cloud resolves (+2–4pp extinction)
Claim: Resolving L1 as first tail: Δw_m −5pp, Δw_b −3pp, Δw_x +8pp → P(extinction) ≈21–22% (from ~17% doom baseline (emergent sim)) — apply one column only.
Why: Cloud RSI first ↑ misalignment branch (25%×40%) if no halt; worked example in Node 12 parent.
Evidence:
- Node 12 reweight formulas + worked example
- (internal note) — branch model
- Crosscut §6 — “+3–5pp misalign if cloud 10+ pp faster than bio”
Analogue: Bayesian model selection — one locus resolution, not stack.
Would update if: Full Phase 4b re-stitch with correlation discount.
Conf: M
21. Branch reweight — L3 embodied resolves (−2–3pp extinction)
Claim: L3 first tail: Δw_m +5pp, Δw_x −5pp → P(extinction) ≈15.5% — ↓ short-run extinction, ↑ whimper/severe.
Why: Hanson-compatible; kinetic harm legible but not paperclip path; factory vetting not training pause.
Evidence:
- Node 12 worked example L3
- Hanson variant — embodied ↑ whimper to 28%
- Node 1 physical wave — GUARD 12–14% embodied tail
Analogue: Industrial automation — harm without extinction.
Would update if: Embodied RSI compounds to C-tier on factory floor before cloud C — rare joint tail.
Conf: M
22. AI 2027 mapping — L1-only RSI vs our weights
Claim: AI 2027 scenario is L1-only for RSI compounding (Agent-2→4, whistleblower Oct 2027); our 22% bio-first + 12% embodied-first are deliberate deviations — physical tails can front-run narrative order.
Why: Scenario useful for L1 timing; insufficient for sim path reweighting. Tracker Behind on Agent-3 superhuman coder (~0.70×).
Evidence:
- AI 2027 scenario
- (internal note)
- Node 12 AI 2027 mapping table
Analogue: IEA scenarios — one demand narrative, multiple supply paths.
Would update if: AI 2027 adds explicit bio-first beat with dates → reconcile weights.
Conf: M
23. Physical limits — cosmic scale does not bind Node 12 horizon
Claim: Even L1 C-tier cloud RSI does not break Landauer/Bekenstein on 2030 policy horizon — Tier 1 solar-system optimizer still 10¹–10³ yr engineering (superintelligence physical limits §2).
Why: Node 12 is Earth-surface first-tail selector, not galaxy-fill timing. Energy/data caps slow C-tier, not block B-tier on 2026–28 window.
Evidence:
- (internal note) §2
- Physical economy limits
- Node 12 takeoff geometry table — “Physical limits bind” row
Analogue: Moore’s Law eventually hits physics — after policy-relevant RSI tiers.
Would update if: Datacenter build hard-stopped by grid/energy policy 2027 → L1 ↓, L3/L4 ↑.
Conf: H (for 2030 scope)
24. Conditional Ci — capability stall shifts locus weights
Claim: If global Ci ≤ C6 by 2030, L1 cloud first-tail ↓ to 0.35; L2 bio ↑ to 0.32; L4 cyber ↑ to 0.14 — cloud RSI narrative overpriced.
Why: Without cloud C-tier, physical-layer tails gain share; Hanson mixture ↑. Tracker Behind/Emerging on several C7+ beats.
Evidence:
- Node 12 conditional Ci table
- (internal note) — mixed Ahead/Behind
- Node 12 falsifier — METR flat <8h through 2028
Analogue: Clean energy — if solar stalls, coal reweights in mix scenarios.
Would update if: C9+ Confirmed on tracker 2027 → restore L1 0.52+.
Conf: M
25. L5 multipolar — Yud-coordination-fail sub-variant
Claim: L5 resolution → misalign branch +17pp weight delta; P(extinction) 22–24%; 85% default / 15% Hanson mixture — anti-Hanson.
Why: Simultaneous salience without ordering = coordination failure — Yudkowsky tail 30–40% sub-variant in crosscut §6.
Evidence:
- Node 12 TAIL-L5 branch — +4–6pp total
- Hanson falsifiers — kinetic datacenter strike
- Node 3 + Node 9 — multipolar diplomacy fails
Analogue: 1914 July crisis — linked events, no single first cause in policy.
Would update if: Clear ordering established post-hoc → collapse L5 into L1–L4.
Conf: L
26. Open-weight (Node 5) — accelerates L1 + L2 tamper
Claim: Node 5 open-weight tails accelerate both L1 cloud tamper and L2 bio cascade — does not change first-tail partition unless T5-B bio near-miss public before cloud C.
Why: Persona vectors, GeneBreaker on open weights — dual locus acceleration; Node 12 cross-node matrix.
Evidence:
- node5 — T5-B bio tamper
- Crosscut §6 Node 5 link
code/typebits/— persona-axis steering (Node 4 ext overlap)
Analogue: Dual-use biotech export — two pathways, one material.
Would update if: Open-weight bio FM release precedes next cloud frontier → L2 ↑.
Conf: M
27. p(doom) summary — branch selector discipline
Claim: Node 12 point estimates: L1 21%, L2 20%, L3 16%, L4 19%, L5 23% vs ~17% doom baseline (emergent sim) — do not add linearly to N7–N11 node deltas; apply one reweight when locus evidence resolves.
Why: Methodology guardrail from Node 12 parent + my pdoom scenario branches.
Evidence:
- Node 12 p(doom) summary table
- correlation matrix — avoid multiplying marginals
- Crosscut §9 rank #3 RSI locus ±3–6pp
Analogue: Scenario planning — pick one world, not sum.
Would update if: User completes Phase 4b full re-stitch with mixture α from Hanson table.
Conf: M
Related files
- node12 rsi locus takeoff
- hanson variant human action
- node2
- node7
- node8
- Anthropic When AI Builds Itself 2026
- main forecast
Update log
| Date | Change |
|---|---|
| 2026-07-04 | Initial — 27 evidenced sections from crosscut §6 + Hanson + N2/7/8 links |