A respected AI research foundation released a new benchmarking test which, it said, will warn when true artificial general intelligence arrives. Frontier AIs pass existing benchmarks easily — but they test knowledge more than reasoning, a comparatively easy task for AIs trained on vast amounts of text, according to researcher François Chollet. The ARC-AGI-3 test presents game-like puzzles the AI must work out on the fly, and Chollet said even top AI models score below 1%. He told Fast Company that any rapid improvement in scores would represent real progress towards AGI — which could be a useful canary in the coalmine, since the physicist who invented the term and the CEO of Nvidia both recently said AGI was already here.