Keep pulling the thread on François Chollet.
The top score in the ARC Prize 2025 Kaggle competition reached 24% on the ARC-AGI-2 private evaluation set.
Anthropic, Google DeepMind, OpenAI, and xAI reported ARC-AGI performance in their public model cards in 2025.
The ARC-AGI benchmark was established as an industry standard for AI reasoning in 2025.
Current frontier AI reasoning performance is fundamentally constrained by knowledge coverage, which is leading to new forms of benchmark contamination.
The upcoming ARC-AGI-3 benchmark will introduce interactive reasoning challenges that require exploration, planning, memory, goal acquisition, and alignment capabilities.
The ARC Prize 2025 global competition targeted the newly released ARC-AGI-2 dataset.
The ARC-AGI-2 dataset features greater task complexity compared to its predecessor.
The ARC Prize 2025 Kaggle competition attracted 1,455 teams.
The ARC Prize 2025 Kaggle competition received 15,154 entries.
Paper submissions for the ARC Prize 2025 nearly doubled year-over-year to 90 entries.
Zero-pretraining deep learning methods are achieving competitive performance with networks as small as 7 million parameters.