The case

Evidence

Every claim the Foundation makes about AI traces to a primary source. These are the figures we lean on most, and how the material behind them was gathered.

The figures

Months on the bottom, size on the side. A quantity that doubles every 4 months is 64 times larger after 24. METR measures this pace for the length of tasks AI can finish; the curve shows the arithmetic, not their data points.

4 modoubling time of AI task length since 2024 (METR)
4×growth in frontier training compute, every year (Epoch AI)
10,000×further scale-up physically reachable by 2030 (Epoch AI)
2047even-odds date for human-level AI in the largest researcher survey, 13 years sooner than the year before
2028Metaculus median for weakly general AI, 1,700 forecasters

What they add up to, and the objections we answer, is in the manifesto. Every document is listed under sources.

Method

The corpus was collected from the official arXiv API using 17 thematic queries and de-duplicated by normalising version numbers. 766 unique studies remain, spanning 2008–2026.

31 papers were selected for full-text reading by a keyword-relevance score combined with topic overlap and publication year, plus one manually added survey. Text was extracted with pdftotext.

Two sources initially blocked by bot protection were recovered through the Wayback CDX API: OpenAI's “Planning for AGI and beyond” and the Metaculus questions. Archive raw mode returns gzipped bytes, which required explicit decompression.

Two version numbers taken from search results proved wrong when checked against the primary documents and were corrected: Anthropic's RSP is v3.4 (not v3.0) and DeepMind's FSF is v3.1 (not v3.0).