← Evidence index · 中文 · Main post
Each section: Claim · Why · Evidence · Analogue · Would update if · Conf (H/M/L).
Parent: Shared Ci spine
Covers: Hybrid time rule, capability spine C0–C10 dating, cross-node synthesis table, shared method claims
Node-specific P’s: see node1 … node4
Hybrid time rule (locked: C)
P(method) — Fast track ≈ AI 2027 / METR for capabilities
Claim: Agent autonomy, METR horizon doubling, coding-agent revenue, internal RSI discourse track ~AI 2027 calendar or faster on some metrics.
Why: AI 2027 Tracker Jun 2026: agent autonomy Confirmed/Ahead; METR TH1.1 doubling ~89 days since 2024 (faster than AI 2027 implied ~131-day all-time). Claude Code / Opus 4.6 pulled AI Futures Model timelines 1.5–2 years earlier (Q1 2026 update).
Evidence:
- (internal note) — agent Ahead, METR-related Ahead
- https://metr.org/blog/2026-1-29-time-horizon-1-1/
- (internal note) §METR + AI Futures Q1 2026
website/src/content/writing/agi-timeline-prediction-methods.md
Analogue: 2023–24 “unhobbling” step-change (chatbot → agent) matched Aschenbrenner OOM narrative timing.
Would update if: METR doubling reverts to >180 days for 4 consecutive frontier releases and agent revenue stalls → slow capability track to 0.85× or below.
Conf: M–H
P(method) — Slow track ≈ +30% for politics / public salience
Claim: Mass job shock, DC protests, mandatory screening enacted, NYT whistleblower mainstream cycle lag capability milestones by ~6–18 months vs AI 2027 drama calendar.
Why: Tracker Jun 2026: economic job-shock nodes Behind; governance protest Emerging not Confirmed; BMIA introduced Jan 2026 but not law; Kokotajlo/Saunders 2024 salience zero pause — policy follows coalitions slowly. COVID Western policy lag ~2–3 mo even for obvious tail risk ((internal note) uses this invariant).
Evidence:
- (internal note) — Economic Behind on some predictions; governance protest Emerging
- (internal note) — FLI letter zero effect; SB 1047 veto timeline
- (internal note) — Canaries micro shock without macro collapse
- Yale Budget Lab 2026 — no national AI unemployment spike 33 mo post-ChatGPT
Analogue: Y2K remediation preceded visible crisis by years; Montreal Protocol followed ozone hole discovery with ~2–3 yr treaty lag.
Would update if: Federal training moratorium or BMIA signed within 3 mo of a single capability headline → compress slow track to +10%.
Conf: M
P(method) — Anchor human nodes by Ci milestone, not AI 2027 plot beats
Claim: Node 3 (theft) binds to C5+, not Feb 2027; Node 4 binds to C10, not Oct 2027 memo.
Why: Weight exfiltration value scales with checkpoint worth (10²⁷–10²⁸ FLOP class); tracker has no public 10²⁸ yet → C5 late. Whistleblower policy effect in 2024 preceded C10 capability — we separate salience (early) from full Node 4 (C10 + Trigger E). AI 2027 is scenario color, not forecast calendar.
Evidence:
- (internal note) — theft beat is narrative stress-test
- (internal note) §timing
- (internal note) — partial Node 4s 2024–26
- (internal note) — SL3/S L4 upgrade lags capability
Analogue: Nuclear test ban debates lagged warhead design capability; policy anchors on deployed systems.
Would update if: Confirmed 10²⁸ run + weight theft attempt in 2027 → pull Node 3 forward to AI 2027 calendar.
Conf: M–H
Capability spine C0–C10 (tracker status claims)
| Ci | Claim (2026-07) | Why | Evidence | Conf |
|---|---|---|---|---|
| C0 | Agents niche → mainstream Confirmed | OSWorld, coding agents, METR hours up | Tracker agent autonomy Confirmed; (internal note) | H |
| C1 | 10²⁸ FLOP run Emerging, not public | Epoch frontier ~10²⁶·⁵–10²⁷; no lab confirmed 10²⁸ | (internal note) §Epoch; tracker C1 Emerging | M–H |
| C2 | 1.5× R&D multiplier mixed | Internal algo speedups real but hard to verify publicly | Tracker economic/agent mixed | L–M |
| C3 | China CDZ Emerging | Geopolitical mobilization rhetoric ↑; hard to verify CDZ equivalent | Tracker geopolitics Emerging | L–M |
| C4 | Junior labor shock partially live; protest Behind | Canaries, Block, Challenger; no 10k DC march | Node 1 rationale; tracker economic Behind on some | M |
| C5–C10 | Not yet testable / Emerging | Continuous learning, neuralese, AGI announce, Agent-4 — no public confirms | Tracker Not Yet Testable / Emerging | M (for “not yet”) |
Each Ci human-response node timing is derived in node-specific rationale files once capability trigger fires.
Cross-node synthesis table (modal path)
Phase: Now → 2027 Q1 (Node 1 partial + Node 2 insider)
Claim: Agents scale; junior hiring hurt; screening coalition forms without public bio panic.
Why: Node 1 triggers met (Canaries, Challenger); Node 2 insider salience met (Amodei Jun 2026, screendna, BMIA intro). Public CBRN salience still needs Trigger E (Node 2 rationale §Trigger E).
Evidence: node1 triggers; node2 insider timing; (internal note)
Conf: M–H
Phase: 2027 Q2 – 2028 (Node 1 peak + Node 2 modal)
Claim: Labor hearings + BMIA/EU screening likely; still no pause.
Why: Modal branch masses on Node 1 (P hearings 0.75) × Node 2 (P BMIA 0.45) with independence approximate — different coalitions (labor vs synthesis/natsec). Both oppose training caps (playbook: pause <5%).
Evidence: node1 modal; node2 BMIA §; (internal note) §Executive summary
p(doom) channel: ↓ P(coordinated pause); bio tail trimmed if modal Node 2 (screening) — conditional not automatic.
Conf: M
Phase: ~2028 (Node 3 at C5+)
Claim: Theft/natsec crisis; export controls; race accelerates.
Why: Modal Node 3: P(export expansion 0.70) × P(security spend 0.75); Anthropic RSI conditional pause needs verification not incident. Theft attempt base rate ~78% experts by 2030 (AI 2027 Security Forecast) but timing at C5+.
Evidence: node3; https://ai-2027.com/research/security-forecast; Fable export ban Jun 2026
Conf: M
Phase: ~2028 Q2–Q3 (Node 4 at C10)
Claim: Whistleblower → oversight not halt; P(extinction | E, modal) ~12–22%.
Why: Product of Node 4 branch (modal 0.58 given E) × conditional chain in node4 §34–37. Prior nodes raised race speed and lowered pause probability.
Evidence: node4; Saunders 2024, FLI 2023 analogues
Conf: L–M (wide conditional chain)
Shared method claims
P(method) — FLI pause letter → zero binding policy
Claim: Costless open letters do not move US law.
Why: 1000+ signers Mar 2023; no moratorium, no licensing; playbook documents zero policy effect; Cruz moratorium stripped 99–1 Jul 2025 shows anti-regulation floor also extreme but opposite direction.
Evidence: (internal note) §2.1; (internal note) §2023.03 FLI
Conf: H
P(method) — SB 1047 veto → SB 53 transparency survives
Claim: Industry kills ceilings; reporting survives with lab buy-in.
Why: Newsom veto Sep 2024; SB 53 signed Sep 2025 with Anthropic/Encode support; Wiener/Encode path documented in (internal note).
Evidence: Playbook §1.2; primer §2.4; node1 SB53 0.85
Conf: H
P(method) — 12–18 mo regulatory delay invariant
Claim: Even salient AI risks face quarter-to-year institutional lag unless Trigger E (incident/leak).
Why: COVID Jan–Mar 2020; OSTP 50bp paused; BMIA referred committee Jan 2026; RAISE effective 2027-01-01 for Dec 2025 sign — ~13 mo state example.
Evidence: Playbook; Node 2 TAIL-B 0.15; Node 4 media fatigue
Conf: M
Related
- README.md — full index
- main forecast — consumes branch outputs