The Next Fifteen Years

A forecast built from first principles
Section future / 05-probabilities / reasoning.md

Per-row reasoning - what moves each probability#


Contents

Each row: the claim restated, what evidence would move it, what rounds 1–7 actually deposited, and the delta taken in round 8. This page is the audit trail for the table; the ledger is the compressed history.

Row 1 - Autonomous day-to-day knowledge work#

Claim: AI systems autonomously do most of a competent knowledge worker's day-to-day tasks.

By 2030By 2040
Round 045%80%
Round 850%80%

What it means. Not "can demo the task" - does the task as the default path in production, without a human in the loop for the majority of a competent worker's actual day. Verification-cheap domains (software, support, parts of finance) count earlier; open-ended professional judgment counts later.

Evidence that would move it up: seat→outcome pricing becoming the enterprise default; junior:senior posting ratios failing to recover with the cycle (B1); measured task-hour substitution approaching the Game 4 band.

Evidence that would move it down: METR-style negative productivity results replicating at higher capability; insurance exclusions blocking autonomous deployment in the domains that looked easy (insurance); organizational redesign remaining the binder through 2030.

Round 8 rationale. Mid-2026 junior-hiring data and the commercial shift toward outcome pricing moved earlier than round 0 expected. Demography does not help this row (it is about task substitution, not aggregate employment). +5 on 2030; 2040 unchanged - the long-run ceiling was never the dispute.

Row 2 - Measured US TFP >2%/yr sustained#

Claim: Measured total factor productivity growth in the US exceeds 2%/yr on a sustained basis.

By 2030By 2040
Round 025%55%
Round 820%50%

What it means. Measured is doing real work. The prices page argues quality-adjusted deflation and unpriced surplus systematically understate welfare-relevant productivity. Clearing an official 2% bar is therefore harder than "AI is working."

Evidence up: labor-productivity prints accelerating in AI-exposed sectors; J-curve language entering mainstream macro commentary with numbers behind it.

Evidence down: continued flat TFP despite visible deployment; central banks and statistical agencies not revising hedonic treatment of cognitive services.

Round 8 rationale. Round 5's macro channel and the measurement argument are new material that round 0 did not have. Organizational lag still binds for 2030. −5 / −5. Note: this is not a claim that AI fails - it is a claim that the statistic is a high bar.

Row 3 - Major AI-attributed catastrophe#

Claim: An event with >1,000 deaths or >$100B damage, attributed primarily to an AI system or AI-enabled process.

By 2030By 2040
Round 020%45%
Round 820%45%

Mechanisms, ordered by likelihood: financial / market dislocation with AI in the causal chain; correlated cyber failure; biosecurity near-miss or worse; kinetic/warfare. Game 2 needs a salient incident for the regulatory story; this row is the severe subset.

Evidence that would move it up: autonomy deployed in high-severity domains ahead of the insurance frontier (uninsured deployment is the leading indicator of an uncontrolled one); correlated near-misses - several insureds failing the same way in one window - since correlation is what converts an incident into a catastrophe-scale one. Down: the insurance exclusions holding through 2030, which would mean the severity ceiling is being enforced by underwriters exactly as Uncertainty 6 hopes.

Round 8 rationale. Insurance correlation analysis (insurance) and biosecurity/cyber domain pages add mechanism detail without a clear directional shift in severity probability. Held. A "salient but sub-threshold" incident remains more likely than this row - and is enough for the Game 2 architecture claim.

Row 4 - Binding international agreement with real verification#

Claim: A multilateral agreement that constrains frontier training or deployment and includes a verification mechanism that would detect material cheating.

By 2030By 2040
Round 010%35%
Round 88%30%
Round 267%28%

Why it was already low. Row 3 is nearly a precondition (event-driven coordination). Compute governance is the only technically plausible lever and degrades with efficiency.

Evidence that would move it up: a compute-accounting or attestation standard actually deployed bilaterally, even at small scale, since the row's pessimism rests on the verification object being spent, not on unwillingness to sign paper; a sub-threshold incident with clearly cross-border harm. Down: further rounds of export-control escalation, each of which spends more of the future verification mechanism on present competition.

Round 8 rationale. Bipolar named the trap explicitly: the export-control lever and the future verification mechanism are the same object; competitive use is a decision to abandon verification later, being made implicitly. That is new structural reason for pessimism, not just "coordination is hard." −2 / −5.

Round 26 rationale. The r25 Ground pass measured the open-weight lag at ~3–6 months (Epoch AI, Jan–May 2026) against the ~9–15 the corpus carried at authoring. This row is where that lands hardest: an agreement constraining frontier training verifies a shrinking share of deployed capability when near-frontier weights circulate a quarter behind, so the verification object is not only being spent by export controls (the r8 argument), it is depreciating on its own. The move is small (−1 / −2) because the row was already priced for pessimism and one measured series is one series; the direction, not the magnitude, is the information.

Row 5 - Humanoids >1M units/yr deployed#

Claim: Humanoid robots deployed (not just produced) at a rate exceeding one million units per year.

By 2030By 2040
Round 015%60%
Round 812%55%

Evidence that would move it up: teleoperation-to-autonomy ratios falling on unstructured tasks (the one number demos cannot fake); actuator and reduction-gear costs breaking their curve, since supply chain binds before model quality does. Down: structured-environment deployments stalling on delivered $/hour despite working autonomy - that would mean even the easy branch of the fork mispriced integration and maintenance, and the humanoid path inherits the error twice over.

Round 8 rationale. Round 4's four-file split made the constraints separable and each harder to hand-wave: no internet of manipulation; delivered $/hour not unit price; supply chain (magnets, actuators, cells) favors manufacturing capacity over model quality; form factor is a data bet. Structured environments still lead by 5+ years - so humanoid at this scale by 2030 is a stretch even if warehouse robotics thrives. −3 / −5.

Row 6 - Serious capital-markets AI correction#

Claim: A >40% drawdown in a recognizable AI-sector equity (or AI-adjacent credit) index, with AI fundamentals in the causal chain - not just a broad macro crash that takes AI with it.

By 2030By 2040
Round 040%65%
Round 845%65%

Evidence that would move it up: A2's revenue-vs-capex gap widening through 2027 while A3's financing mix keeps migrating toward structures that cannot absorb a miss quietly. Down: AI revenue compounding near the rate the capex implies, or the financing mix reverting to operating cash flow - either would let disappointment resolve as a spending pause rather than a drawdown that clears the 40% bar.

Round 8 rationale. Capital (round 1, deepened round 5): migration from operating-cash financing toward private credit, SPVs, and vendor financing raises the odds that disappointment becomes a credit event rather than a board-level pause. Hyperscaler guidance vs. revenue still the headline test (A2). Form remains consolidation, not technology death. +5 on 2030; 2040 already high enough that further moves need a different claim.

Row 7 - Existential-scale loss of human control#

Claim: Permanent, large-scale loss of human ability to direct or halt frontier AI systems, with catastrophic civilizational consequences.

By 2030By 2040
Round 01–3%3–8%
Round 81–3%3–8%

Evidence that would move it up: internal research cycle times at frontier labs compressing faster than the three governors predict, since nearly all of this row's mass lives in the worlds where Uncertainty 1 breaks the timeline; capability evaluations losing their meaning as a control surface (models routinely sandbagging or situationally aware in testing). Down: verification of research outputs staying expensive as capability grows - the same asymmetry that orders Part III also governs this tail.

Round 8 rationale. Uncertainty 1 now names three governors; fast takeoff is over-weighted relative to technical plausibility alone. That compresses the upper tail somewhat but does not justify rewriting the range - ranges this wide already encode that uncertainty. Held, with the note that the operative planning risk remains row 3.

Rounds 27–30 - hold block (no number moves)#

Depth and Ground without a new scoreable series that crosses a Part V bar. Mechanism prose improved (Taiwan gray zone, Game 3 cross-section, robotics fully-loaded cost, partial-clause scoring); row 4 already moved in r26 on lag compression and is not re-moved. Capex prints revised toward ~$700–725B - direction already priced in row 6's r8 delta and r26 hold. Rule 5: holds logged here rather than inertia.


Related: Ledger · Scoring rules · Indicators

View markdown source

select · Enter open · Esc close