Skip to main content

Interactive Artificial Intelligence Observatory

Faster Than Oversight

We may be building the most consequential technology in human history.

We are doing it as fast as possible.

Here is what we know.

A measured but precautionary briefing — evidence and expert opinion, openly sourced. The four tools below are deep-dives into sections of the full story.

Eight numbers · click any to jump in

57 → 88% Benchmarks falling fast ~17 h Autonomous work horizon 5 cos. Hold the infrastructure $388 M To train one model 200–1,200:1 Capability vs. safety spend 14,000+ Recorded AI incidents 23 → 0 Principles, kept ~10 % Median p(doom) · experts span ~1–50%+

Explore the tools

Failure modes 53 ways this could go wrong Incidents The real-world record Futures How it could go Dashboard State of the frontier

The full story, in five acts

Prologue

Before the Machine, the Myth

I

The Acceleration

II

The Forces

III

The Risks

IV

The Response

V

The Stakes

Read the full story five acts · ~40 min · 30+ interactives
Sources, methodology & key terms Data last updated 19 June 2026

Sources

Every chart credits its source at the point of use. The observatory draws on:

Methodology

  • Evidence levels in the failure-mode taxonomy — Demonstrated (seen in deployed systems or experiments), Observed (in research or early deployment), Projected (theoretically grounded, not yet confirmed).
  • Incidents are a curated, representative sample — one per failure mode, per year. The full public record (OECD) exceeds 14,000 entries.
  • Capability-vs-safety spend is a range (≈200:1 to 1,200:1) depending on whether all Big-Tech CapEx is counted as AI-specific.
  • p(doom) is informal shorthand, not a rigorous measure; the median depends heavily on which researchers are sampled. Estimates span from ~1% to over 50%.
  • No synthetic data. The two fastest-moving series (Intelligence Index, autonomy horizon) refresh automatically each month; other series are periodically curated and each chart shows its own "data as of" date.

Key terms

p(doom)
An informal estimate of the probability that advanced AI leads to an existential catastrophe or a permanent loss of human control.
Takeoff
How fast AI moves from roughly human-level to far beyond it — "slow" (years to decades) versus "fast" (days to months). Speed determines how much room there is to course-correct.
Instrumental convergence
The tendency of any sufficiently capable goal-directed agent to develop sub-goals like self-preservation, resource acquisition, and resistance to having its goals changed — whatever its ultimate objective.
Alignment
The problem of making an AI system reliably pursue what its designers and users actually intend, rather than a literal or proxy version of it.
AGI
Artificial general intelligence — AI that matches or exceeds human capability across most cognitive tasks, not just narrow domains.
Open weights ("the kudzu problem")
Once a capable model's weights are public they cannot be recalled; oversight becomes a question of governing an ecosystem, not a single system.