Methodology

AGI Watch grades the AI 2027 scenario against reality. The goal is practical: give founders a clear planning signal without pretending the future is more precise than it is.

1. The authors’ own figures (primary)

The AI Futures Project publicly grades its own predictions. Wherever they publish a number, we use it verbatim:

• Aggregate pace of progress — ~65% of predicted pace (Feb 2026 scorecard), revised to ~75% with July 2026 data.
• OpenAI revenue — ~$20B actual vs ~$18B predicted.
• METR 80% coding time horizon — 1.04× their central trajectory.
• Revised takeoff window — mid-2028 → mid-2030, from their AI Futures Model adjusted for compute and labor-growth slowdowns.

Source: Grading AI 2027's 2025 Predictions (Eli Lifland & Daniel Kokotajlo, Feb 12 2026; updated Jul 2026).

2. Our estimates (always marked "~")

Where the authors report predicted-vs-actual values but no multiplier, we derive one and flag it:

• SWE-Bench Verified ~0.2× — predicted 85% by mid-2025 from a 72% baseline (13 points of gain); actual best was 74.5% (2.5 points). 2.5 ÷ 13 ≈ 0.19.
• Valuation ~0.5× — $500B was predicted for Jun 2025 and reached Oct 2025. A rough elapsed-time ratio; the authors describe it as "well behind pace."

What we deliberately don't show

• AI R&D uplift — behind pace per the authors, but no defensible single multiplier exists.
• Largest training run — behind pace (no confirmed run substantially beyond GPT-4.5), but training compute is too opaque to quantify honestly.

A number we can't defend is a number we don't chart.

The re-projected timeline

The bottom track on the tracker re-projects the scenario's milestones at the observed ~0.7× pace, anchored to the scenario's publication (April 2025), and uses the authors' own revised takeoff window rather than a naive linear stretch. Their personal medians differ (2029 for full coding automation per Kokotajlo; early 2030s per Lifland) — the scenario's "2027" was their modal year, not their median.

Update cadence

The tracker updates when the underlying data moves: on each AI Futures scorecard, on METR releases, and on major benchmark results. Every change is logged on the changelog with the date and what moved. The raw data behind every chart is public at data/tracker.json.

Corrections

Spotted an error? Tell us — corrections ship with a changelog entry, never silently.