Sean Lie

Sean Lie is CTO and co-founder of Cerebras. He helped launch its effort to bring wafer-scale computing to market alongside Andrew Feldman, Gary Lauterbach, Michael James and Jean-Philippe Fricker. His AI work includes helping enable unstructured sparse training on the Cerebras CS-2 and a public argument that faster inference can give reasoning models more computation within an application's response-time limit.[2][3][4]

In the June 11, 2025 article The Cerebras Scaling Law: Faster Inference Is Smarter AI, Lie and James Wang argued that inference speed matters because reasoning models generate additional tokens before delivering an answer. Applications impose latency limits, which constrain how many of those tokens a model can use. Faster generation, they argued, lets a model spend more computation on reasoning while still responding quickly enough for real-time use.[4]

Lie and Wang described a change in their stated direction: “We started the company to build the fastest chip.” Their newer goal was to use speed to make more capable reasoning models practical in interactive applications, calling that direction their “new North Star.” This was a joint argument about how to use inference-time computation, rather than a claim that speed alone improves every model or workload.[4]

Before that 2025 argument, Lie contributed to the hardware support behind another approach to model efficiency. The July 17, 2024 version of Sparse-IFT: Sparse Iso-FLOP Transformations for Maximizing Training Efficiency, coauthored with Vithursan Thangarasa, Shreyas Saxena and Abhay Gupta, credits him with helping bring up CS-2 sparsity support that was crucial for training and benchmarking Sparse-IFT variants of GPT models. He also provided feedback on the manuscript's structure and presentation.[3]

Sparse-IFT expands model capacity within a fixed theoretical FLOP budget. For example, widening a layer while making its weights sparse gives dynamic sparse training a larger space of possible connections to explore without increasing the dense comparison's arithmetic budget. Saxena conceived the key idea of matching the Sparse Wide transformation's FLOPs to a compact dense model; Thangarasa contributed methods, implementation, spectral analysis and experiments, while Gupta contributed optimizer validation and vision and GPT experiments. Lie's specifically credited contribution helped make the GPT experiments possible on CS-2.[3]

The work also illustrates why an arithmetic budget and elapsed training time are different. The authors reported minimal training overhead through 75% sparsity in their CS-2 benchmark, but a 2.82× increase in training time at 90% sparsity even with sparse acceleration. They concluded that realizing sparsity's benefits requires software and hardware co-design: fewer active weights do not automatically translate into proportionate wall-clock savings.[3]