François Chollet predicts AGI will emerge by 2030

AI researcher and former Google DeepMind lead François Chollet identifies a critical milestone in the development of artificial general intelligence (AGI) by 2026. In his recently published technical report, Chollet states that the new generation ARC-AGI-3 intelligence test will be fully solved by all current models by the end of this year, and that this development will serve as a critical indicator for predicting the timeline of AGI’s realization.

ARC-AGI-3: A New Benchmark for AGI

ARC-AGI-3 is a conceptual intelligence test, developed by Chollet, designed solely to measure reasoning and generalization abilities—not data-driven performance. The third iteration of this series, launched in 2024, has been solved at a rate of 98.7% by the most advanced models developed by 17 different institutions as of 2026. This rate had reached 76% by the end of 2025. Chollet argues that this rapidly increasing success rate demonstrates that AGI is “technically feasible.”

2030: The Time for AGI’s Realization

At the Singularity Summit in February 2026, Chollet stated, “By 2030, it will be technically guaranteed that a system can independently perform any human-level cognitive task in novel situations through generalization.” This prediction offers a more precise timeline than his earlier forecasts from 2024 and 2025. Chollet bases this projection on three key factors: the performance curve of ARC-AGI-3, model scaling laws, and the increasing diversity of training data.

Industry and Academic Reactions

Chollet’s prediction has drawn significant attention from both academic circles and tech giants. Researchers from MIT and Stanford have confirmed that this forecast is “data-driven and mathematically modelable.” Meanwhile, companies such as OpenAI and Anthropic argue that realizing AGI without an ethical and safety framework would be risky. In the first quarter of 2026, the US and EU officially implemented their AGI preparedness strategies.

Next Steps: Redefining AGI

Chollet is redefining AGI not as “a system that can do everything,” but as “a system capable of solving any human cognitive task in novel contexts by learning from limited data.” This new definition emphasizes few-shot learning and meta-learning capabilities. ARC-AGI-3 was specifically designed to test this definition. By the end of 2026, the entire test will be executed automatically, and results will be published as an open-source dataset.

Chollet emphasizes the accelerating pace of technological progress, stating, “AGI is not an event, but a process. 2030 is not a marker of its beginning, but of its completion.” Within the coming year, the ARC-AGI series will expand with “ARC-AGI-4,” introducing new tasks designed to push the boundaries of human intelligence.