Ben Goertzel doesn't hedge much. The Mindplex editor-in-chief, SingularityNET CEO, and chair of the AGI Society told IBM Think that artificial general intelligence could materialize within nine to twelve months—a sharp compression from the three-to-four-year window he'd been giving until recently.
What changed his mind? AGI-26, the 19th annual Conference on Artificial General Intelligence, held in San Francisco July 27–30. Goertzel said the single most striking thing wasn't a paper or a benchmark but watching AI researchers themselves use "hives and swarms of agents" to do their actual work. Plan experiments, write code, chase down citations, find connections between ideas. The humans were stepping back. The agent clusters were doing the thinking.
The data backs the intuition. The 2026 Stanford AI Index shows AI agent accuracy on OSWorld, a test of whether an AI can complete real tasks across a computer operating system, jumped from roughly 12% to 66% in a single year, brushing against the human baseline of 72%. Google DeepMind's Gemini Deep Think, meanwhile, earned a gold-medal score on International Mathematical Olympiad problems.
But here's the wrinkle, and Goertzel acknowledges it: the same systems that ace Olympiad math score just 50.6% on reading analog clocks. Humans hit 90.1%. On ARC-AGI-3, where agents are dropped into unfamiliar digital environments and must figure out the rules from scratch, the gap remains vast. The intelligence is deeply uneven.
Still, Goertzel left the conference more optimistic than he arrived. "There are more virtuous cycles like that going on than I had realized," he said. For the person who has spent decades building toward AGI that's a notable shift in tone.