Who is betting on what in AI, and against whom.
The benchmark camp 1 keeps failing to close.
The nonprofit that runs ARC-AGI, the benchmark built to resist memorization. ARC-AGI-2's best verified 2025 result was 54% from Poetiq's refinement loop over Gemini 3; ARC-AGI-3, launched in 2026, is interactive and the first milestone winner scored about 1.2%.
No capital figures recorded yet.