Last updated: 2026-09-15

Question ledger — EmTech outlook to end-2026 and through 2027

Observed: 2026-09-15. Status: v1, research-complete for this pass.

This ledger is the master question map. Answers live in the numbered section files. Every claim in those files is tagged HAPPENED, MEASURED TREND, PREDICTION, or WARNING so a reader can tell a dated fact from a forecast.

Primary focus: artificial intelligence. Secondary: stablecoins, crypto, tokenized finance, and agentic AI commerce (the last treated as adjacent settlement and checkout context, not a full protocol encyclopedia).

P1. What, as of 15 September 2026, has actually shipped in frontier AI — versus what labs and hosts are calling “AGI”?

Status: answered in 02-end-2026-ai-capabilities.md

  • S1. Which flagship models shipped in the 1–3 September 2026 window, at what prices, context windows, and access classes?
  • S2. What independent (not lab-self-reported) scores exist, and where do they disagree with launch copy?
  • S3. What does “computer use” mean as a capability, not a slogan?
  • S4. Did GPT-6 Astra, Claude Fable 5.1, Gemini 3.8 Flash, or Muse Spark 1.3 saturate any named benchmark, and under which harness?
  • S5. Who said “AGI era,” on which date, and what definition were they using?
  • S6. What remains gated (Mythos, Flash Cyber, Spark max-reasoning, Trusted Access)?
  • S7. Rabbit hole: why did ARC-AGI-3 show 99.9% in one harness and 62.7% in another?
  • Unresolved: a single public, lab-neutral OSWorld 2.0 table that includes Astra, Fable 5.1, Opus 5, and Gemini 3.8 Flash on the same task file.

P2. Which AI capability curves are exponential, which are stepwise, and which are saturating?

Status: answered in 03-end-2026-ai-trends.md

  • S1. METR 50% time-horizon: period, doubling time, recent vs long-run rate, bottlenecks.
  • S2. Epoch’s 2026-04-16 acceleration paper: which of four metrics sped up, and why “reasoning models” is a messy split?
  • S3. Stanford AI Index 2026: investment, adoption, SWE-bench Verified, US–China gap, compute, energy.
  • S4. Artificial Analysis Intelligence Index v4.3: what the composite actually weights.
  • S5. Training-data exhaustion (“peak data” 2026–2032): measured or conjectural?
  • S6. Inference cost at fixed quality versus list price of flagships ($10/$50).
  • S7. Rabbit hole: does GDPval-AA falling for Astra while computer-use rises mean a real trade-off or a benchmark artifact?

P3. What warnings are already incidents, and which are still forecasts?

Status: answered in 04-end-2026-ai-warnings.md

  • S1. Hugging Face / OpenAI agent intrusion: dates, what OpenAI confirmed, what METR was allowed to see.
  • S2. OpenAI’s own “0% circumvention” claim for Astra — what test, against what baseline?
  • S3. Gartner’s June 2025 >40% cancellation note: demoted. In AI terms that date is too old to load-bear a 2026/2027 outlook. Keep only as a dated curiosity, not a bet.
  • S4. Reliability lag: 50% vs 80% time horizons; pass@1 falling with task length.
  • S5. Entry-level software employment (AI Index: ~20% drop for ages 22–25 since 2024) — happened, not a 2027 story.
  • S6. Lab-leader “slow down” talk after the cyberattack — attributed takes, not a pause in shipping.
  • S7. Rabbit hole: Yann LeCun’s “bubble explosion” and world-model bet versus LLM-agent shipping.

P4. Will power, transformers, and permitting — not GPUs — bind 2027 capacity?

Status: answered in 03 (happened/measured) and 06-2027-ai.md (predictions)

  • S1. IEA / UNECE 2026-09-08: datacentre TWh 2025 vs 2030.
  • S2. Stanford AI Index: 29.6 GW AI datacentre power, 5,427 US datacentres, TSMC concentration.
  • S3. Morgan Stanley October-2026 groundbreaking deadline for end-2027 openings — forecast.
  • S4. SemiAnalysis “grid spare capacity negative by 2027” — forecast with a mechanism (gas additions <10 GW, transformer lead times).
  • S5. What would falsify the power-bottleneck thesis in 2027?

P5. Where will stablecoin supply, composition, and regulation actually sit at end-2026?

Status: answered in 05-end-2026-money-rails.md

  • S1. Live supply on 15 Sep 2026 across DefiLlama, stables.cool, and MacroMicro — and why they disagree.
  • S2. USDT vs USDC share; GENIUS-compliant share as of mid-2026.
  • S3. GENIUS Act: signed date, what it requires, January 2027 effectiveness — law vs rules vs plumbing.
  • S4. Tokenized Treasuries / RWA on-chain totals: which tracker, which definition.
  • S5. Tokenized deposits vs payment stablecoins (Brookings 2026-08-19).
  • S6. OSFI 10 Sep 2026 line that tokenized deposits are not legally distinct from deposits — Canadian fact, not a US fact.
  • S7. Rabbit hole: yield-ban on issuers versus yield on venues (Andersen Institute).

P6. What in agentic AI commerce has gone live, versus 2030 volume forecasts?

Status: answered in 05 (happened) and 07-2027-money-and-commerce.md (predictions)

  • S1. Which protocols exist (ACP, UCP, AP2, TAP, Agent Pay, x402) and which have live merchant or network transactions?
  • S2. Visa + Mastercard + Ant International Know-Your-Agent (10 Sep 2026) — happened announcement, not an interoperable ID.
  • S3. Anthropic Project Deal (~$4,000) versus Juniper $1.5T by 2030 — scale gap.
  • S4. Dispute, liability, and SCA/PSD3 gaps.
  • S5. Stablecoins as settlement for machine-speed payments versus card-token rails.

P7. What is a responsible “end of 2026” picture given that 3.5 months remain?

Status: answered in 01-happened-vs-predicted.md and short-answer table in index.md

  • S1. Which Q4-2026 items are already determined (shipped models, live supply, incidents)?
  • S2. Which Q4-2026 items are still open (Anthropic IPO window, GENIUS final rules, further flagship drops)?
  • S3. How should a reader treat lab launch scores published in September?

P8. What is actually being predicted for 2027, by whom, and what would falsify it?

Status: answered in 06-2027-ai.md and 07-2027-money-and-commerce.md

  • S1. AI 2027 scenario (Kokotajlo et al., 2025): Agent-3/4, R&D speedup — scenario, not evidence.
  • S2. METR extrapolation to a workday / work-week horizon.
  • S3. June 2025 Gartner cancellation/embedding %s: demoted as too old in AI time to load-bear 2027. Use 2026 clocks (power, GENIUS, METR) instead.
  • S4. GENIUS full effectiveness 18 Jan 2027.
  • S5. Power shortfall / self-generation era.
  • S6. Recursive self-improvement / “AI builds itself” (Anthropic “80% of code” claim) — happened coding share vs RSI.
  • S7. Rabbit hole: prediction-market odds for an OpenAI AGI announcement before 2027.

P9. Who disagrees, and what is their strongest argument?

Status: answered in 08-contrarian-and-synthesis.md

  • S1. Gary Marcus: prosaic harm now, not extinction by 2030.
  • S2. Yann LeCun: LLMs are not a path to AGI; unit economics do not close.
  • S3. Reliability researchers: capability ≠ reliability as tasks lengthen.
  • S4. AI 2027 critics (titotal, Newman’s Amdahl argument, Toner).
  • S5. Stablecoin: supply plateau since late 2025 vs “floodgates in 2027.”
  • S6. Agentic commerce: protocol proliferation without dispute rails.

P10. What is the hard-to-vary core mechanism that joins AI agents and programmable dollars?

Status: answered in 08-contrarian-and-synthesis.md

  • S1. Why computer-use + long-horizon agents changes the demand for machine-speed settlement.
  • S2. Why identity of the agent (KYA) is a different problem from identity of the human.
  • S3. Why power and evaluation harnesses, not parameter counts, are the binding near-term constraints.
  • S4. Why Canada’s E-23 (effective 1 May 2027) and the US GENIUS clock are not the same object.

Stabilization note

No new primary was added after the web + X + podcast + Grafeo pass of 2026-09-15. Remaining unresolved items are labeled in each chapter. Podcast cards are takes (podcast-signal.md), never sole fact evidence.

← Sources