A small team of NVIDIA experts has stunned the artificial intelligence community by winning a prestigious AGI competition. Ivan Sorokin and Jean-Francois Puget, both Kaggle Grandmasters representing NVIDIA, claimed first place on the Kaggle ARC Prize 2025 public leaderboard with a remarkable score of 27.64% on the ARC-AGI-2 benchmark.
NVIDIA’s Kaggle Grandmasters shock the AI world with their surprising victory in a groundbreaking AGI competition.
The competition tests artificial general intelligence (AGI) through challenging reasoning tasks that require machines to learn efficiently and think abstractly. What makes their victory particularly impressive is that they achieved this with a relatively modest 4-billion parameter model, outperforming teams from major AI companies including Google.
“The ARC-AGI-2 benchmark is considered a minimal test for early forms of artificial general intelligence,” explained competition organizers. The prize structure includes a grand prize of $700,000 for reaching the ambitious score of 85%, showing how challenging these tasks are for current AI systems.
The NVIDIA team’s success came from leveraging Kaggle’s powerful L4x4 GPU platform with 96GB memory, which allowed them to train and evaluate their complex model efficiently. They didn’t just rely on raw computing power, though. They incorporated innovative algorithms that helped their smaller model outperform larger competitors.
Unlike typical AI contests that focus on narrow abilities, the ARC Prize measures general intelligence by requiring AI systems to handle diverse, novel tasks without pre-programmed domain knowledge. The achievement, noted on December 5, 2025, marks a significant milestone in artificial intelligence development. The team later improved their score to 29.72% using a method costing only twenty cents per task. Their approach aligns with emerging trends in small language models that achieve impressive results with significantly less memory and computational requirements. Solutions are judged not just on performance but on conceptual progress, encouraging creative approaches to AGI.
The winning team’s approach emphasized model efficiency rather than simply increasing parameter count. Their victory demonstrates that thoughtful design can sometimes trump brute computing force in the race toward artificial general intelligence.
This achievement represents a significant milestone in AGI development and highlights how smaller, focused teams can make breakthrough contributions in a field often dominated by tech giants with vastly greater resources.
The competition continues to push the boundaries of what’s possible in machine intelligence.
References
- https://www.youtube.com/watch?v=MAtfV-reA7c
- https://hyper.ai/en/headlines/2d7042e2c964d16272d81a76e45411f8
- https://forums.developer.nvidia.com/t/nvidia-kaggle-grandmasters-win-artificial-general-intelligence-competition/353697
- https://manifold.markets/Bayesian/arcagi2-top-score-in-2025
- https://arcprize.org/guide
- https://www.kaggle.com/competitions/arc-prize-2025
