Realistic Futures of AI
A living model of plausible AI development, 2025–2040. Every claim is wired to sources and constraints; the sheet unfolds only as far as physics, economics, and evidence let it.
REV 2026-11-01 Week of 2026-11-01Plausible scenarios, 2030–2040
FOUR FOLD STATES OF ONE SHEETModerately Accelerating Path
Most ProbableContinued model improvements, increasingly capable agents, AI-integrated teams. Partial AGI-like systems in narrow domains. No runaway recursive self-improvement.
High-Acceleration Path
OptimisticBreakthroughs in architecture or hardware. Rapid progress in tool-form agents. Semi-autonomous research assistants. Significant shift in scientific productivity.
Low-Acceleration / Regulated Path
ConservativeStrict compute caps, global licensing, slower innovation, strong safety constraints.
Wildcards
High UncertaintyEnergy breakthroughs, new algorithmic paradigms, brain-computer interfaces, digital deflation, catastrophic misuse, or governance shocks.
The creases: 17 constraints
All constraints2 mountain folds — hard
Physics and mathematics. These creases cannot be violated; every scenario deploys inside them.
15 valley folds — soft
Empirical patterns. These creases hold until evidence justifies re-folding — each override is recorded.
Core documents
Baseline model
Evidence-based trajectories for AI development: capability paths, alignment, hardware, socioeconomic impacts, and scenario analysis.
SHEET 02Timeline predictions
When Kurzweil, Amodei, Altman, LeCun, Metaculus, and others expect AGI, ASI, and the Singularity.
SHEET 03Background theories
The constraint catalog: foundational limits from physics, economics, and systems theory that bound every prediction.
Software development milestones
YEAR-BY-YEAR CAPABILITY TRACK · 2025–2030
Revision record
All revisions| Rev | Description | Tags |
|---|---|---|
| 2026-11-01 | Week of 2026-11-01 Anthropic filed to go public, US banks started shutting the door on third-party AI agents, and New York floated a bill that would regulate models able to automate AI research; a thin week for capability news. | policysafetyhardwareindustry |
| 2026-10-25 | Week of 2026-10-25 An always-on consumer agent entered research preview, Google's Gemini 4 Argon arrived behind Opus 5.5, and the data center debate reached a state legislature. | modelsagentspolicyhardware |
| 2026-10-18 | Week of 2026-10-18 A quiet window: Google's Gemini 4 Argon reached general availability, below Opus 5.5 at max effort in contemporary comparisons, with a lower cost to run the index. | modelshardware |
| 2026-10-11 | Week of 2026-10-11 OpenAI published 722 machine-written math manuscripts of which fewer than half are formally checked, and Anthropic's Haiku 5.5 pushed the falling-price pattern to the cheapest tier. | modelssciencesafety |
| 2026-10-04 | Week of 2026-10-04 The White House signed a voluntary audit accord with six frontier-lab leaders, California expanded worker protections alongside its existing audit frameworks, and OpenAI withheld a finished flagship over alignment results. | safetypolicymodels |