Research
Featured releases, papers, and explainers.
doubleAI’s WarpSpeed: Surpassing Expert-Written Kernels At Scale
WarpSpeed is doubleAI’s AI system for GPU performance engineering, demonstrated by independently generating and verifying a hyper-optimized drop-in version of NVIDIA cuGraph that delivers broad speedups across graph algorithms and GPU architectures
WarpSpeed and the Need for Artificial Expert Intelligence
WarpSpeed showcases Artificial Expert Intelligence as a solution to expertise bottleneck. Machines can surpass human experts even in data-scarce, hard-to-validate, deeply technical domains.
From Reasoning to Super-Intelligence: A Search-Theoretic Perspective
A search theoretic perspective on reasoning gives rise to a new learning paradigm that efficiently infers the underlying thought process behind scientific results.
FormulaOne: A Benchmark for deep Algorithmic Reasoning
A reasoning benchmark at the level of real world mathematically complex problems, stumping all frontier AI models.
FormulaOne: Measuring the Depth of Algorithmic Reasoning Beyond Competitive Programming
FormulaOne is a 120-problem MSO-based dynamic programming benchmark on tree-like graphs where frontier models solve under 1%.
Artificial Expert Intelligence through PAC-reasoning
A theoretical paradigm that provides robust guarantees for reliably decomposing complex problems, with a practical mechanism for controlling reasoning precision.
Video Explainers
How our founders breakdown the next frontier.
Can AI think like a Scientist? Intelligence, Autonomy, and the Limits of Alignment
Prof. Amnon Shashua examines the next frontier beyond generative AI, focusing on whether AI systems can move from pattern generation to scientific reasoning, discovery, and autonomous problem-solving.
Learning to Reason: How can models learn to actually reason, not just imitate reasoning patterns?
Prof. Shai Shalev-Shwartz, doubleAI’s Co-founder, introduces a search theoretic perspective on Chain of Thought (CoT) learning and explains why many of today’s approaches often fall short: they drift off distribution, lack structured search, and can lead to escalating inference costs.
