So who exactly is Scott Gray? And why he matters Gray is a fairly old-school OpenAI technical figure. OpenAI announced his arrival all the way back in 2016, when the company was barely a year old. Before OpenAI, he worked at Nervana Systems, where he specialized in optimizing deep neural networks on GPUs. OpenAI described his assembly-level dense-linear-algebra and convolution optimizations at the time as among the fastest available. So he wasn’t simply someone working on “GPUs” as he humbly put it. His specialty sits very close to the compute efficiency / kernels / inference layer underneath modern AI systems. For example, in 2017 - Gray, Alec Radford and Diederik Kingma published OpenAI’s work on GPU kernels for block-sparse weights. They showed that specialized kernels could, depending on sparsity, operate orders of magnitude faster than conventional alternatives such as cuBLAS. He subsequently appears in the official credits for some of OpenAI’s biggest systems. For GPT-4, OpenAI specifically lists Scott Gray among only five people under “Inference research”. He is also credited as a core contributor to GPT-4o. That makes his departure worth watching. The really intriguing part is what he’s doing next. Hints from himself - “Exploring some neuro-inspired AI approaches” And that actually connects to a long-standing interest of his. His older public bios explicitly described an interest in neuroscience and its application to machine learning. His Bluesky profile similarly says he has a background in computer/systems/cognitive neuroscience and is interested in incorporating more insights from neuroscience into AI/ML. So this may represent a fairly meaningful intellectual shift: from making today’s neural networks run extraordinarily efficiently → toward questioning whether we should build neural networks differently in the first place. Definitely keeping my eyes on him - best of luck to Scott!