Chris Ré

Chris Ré (Christopher Ré) is a professor of computer science at Stanford whose collaborative AI work includes Snorkel, which turns heuristic labeling rules into training supervision, and FlashAttention, which computes exact attention with fewer transfers between levels of GPU memory. He belongs to the Stanford AI Lab and is affiliated with the Machine Learning Group and the Center for Research on Foundation Models.[2][3][5][4]

In August 2026, Ré and Stuart Sul described a concrete change in their GPU programming practice. The previous year, they had built an abstraction layer to manage megakernels’ complex synchronization and control flow, then spent months fighting race conditions and deadlocks. For a 2026 mixture-of-experts megakernel, they removed that layer and used agents to produce code optimized for the target hardware. As they put it in “Retire the Abstractions”, “But agents let us put the abstraction in the prompt, in an incomplete, messy form, rather than in carefully designed C++ templates.”[6]

Their argument concerns abstraction’s role in managing cognitive load, with important limits. They could not generate a megakernel in one shot, and removing the framework meant retaining reference implementations, numerical tolerances and expectations about profiled execution. “We can only retire a layer once we have an oracle that outlives it.” They also acknowledged that their experience came from a domain they knew deeply: abstractions still provide shared surfaces for reuse, review and teaching newcomers. Their prediction that CUDA domain-specific languages might become unnecessary included their own ThunderKittens, which they said they would continue maintaining.[6]

Ré and Sul’s February 2026 ThunderKittens 2.0 release emphasized subtraction alongside new features: removing unnecessary memory instructions, assembler inefficiencies and build complexity. One investigation turned on a typo in Nvidia’s PTX documentation that obscured implicit pipelining between scale-copy and matrix-multiplication instructions. Once they recognized it, they combined those operations in the same thread and removed unnecessary barrier waits. They reported roughly a 10% improvement for their NVFP4 matrix-multiplication kernel. They separately credited Nash Brown with identifying assembler behavior that inserted avoidable serialization loops.[7]

The same account illustrates their concern with how performance is measured. They reported that the CUTLASS profiler’s integer-like inputs reduced transistor bit-flipping and power consumption, affecting clock throttling and making competing kernels appear faster. Their public benchmark convention consequently used bitwise-identical random inputs, input groups to produce cold-cache conditions where needed, 500 warmup iterations to reach a steady power state, and 100 profiling iterations without intermediate synchronization. The lesson was grounded in their experiments: apparently minor measurement choices could change reported throughput.[7]

Earlier, Ré coauthored the 2022 FlashAttention paper with Tri Dao, Daniel Y. Fu, Stefano Ermon and Atri Rudra. Its central design choice was to account for GPU memory traffic: tiling reduces reads and writes between high-bandwidth memory and on-chip SRAM while preserving exact attention. The authors reported a 15% end-to-end wall-clock speedup for BERT-large at sequence length 512 against the MLPerf 1.1 training record, and a 3× speedup for GPT-2 at length 1,000. These were results for the specified workloads and baselines.[4]

Ré also participated in Latent Space’s December 2023 NeurIPS recap, in its closing segment on state space models.[1]

Before this foundation-model and GPU work, Ré worked on probabilistic databases, information extraction and training-data creation. That earlier work supplies context for his contributions to systems that combine data processing with machine learning.[8][5]

Ré earned a B.S. from Cornell in 2001 and a Ph.D. from the University of Washington in 2009. He served as an assistant professor at Wisconsin–Madison from 2009 to 2013 before joining Stanford’s faculty. He became a MacArthur Fellow in 2015.[8]

His earlier collaborative work included DeepDive, an inference engine for turning “dark data” buried in text, tables and images into structured knowledge bases. It extracted relationships among entities and inferred assertions that could be integrated into existing databases. Rather than assuming every input was correct, it treated data, tool outputs and user input as observations from which to infer answers. The challenge was considering many possible interpretations efficiently. The MacArthur Foundation’s 2015 account described applications to biomedical literature and human-trafficking investigations; Ré’s homepage lists DeepDive publications with collaborators including Feng Niu, Ce Zhang and Jude Shavlik.[8][3]

In the 2017 Snorkel paper, Ré collaborated with Alexander Ratner, Stephen H. Bach, Henry Ehrenberg, Jason Fries and Sen Wu on another bottleneck: creating training labels. Users wrote labeling functions expressing heuristics whose accuracies and correlations could be unknown. Snorkel then denoised their outputs without access to ground truth during that step, implementing the team’s data-programming approach. In a user study, the authors reported that subject-matter experts built models 2.8× faster and improved predictive performance by an average 45.5% compared with seven hours of hand labeling.[5]

Ré describes student development as the lab’s central purpose: “the lab's real goal is to help students become professors, entrepreneurs, and researchers.” On his personal homepage, he connects that goal to former group members entering professorships and to companies and investment firms he has cofounded with students and collaborators.[3]