Edited by Panda
Today, a tweet from a couple of days ago started gaining traction, containing just five words: Scott Gray has left OpenAI.
The renowned engineer, often hailed as the "God of CUDA Kernels" and the "world's strongest GPU programmer," has become another heavyweight talent to depart from OpenAI this year. Although he hasn't posted a tweet to confirm the news himself, and his LinkedIn page has yet to be updated, his bios on both 𝕏 and Bluesky have changed from "GPU Geek at @OpenAI" to "Former GPU geek at @OpenAI." He also added a note at the beginning, stating that he is currently independent and exploring some neuroscience-inspired AI methods.
An engineer who spent a decade at OpenAI announced his departure with a single line in his bio.
As fate would have it, exactly ten years ago on August 16, OpenAI published a blog post titled "Team update," introducing five full-time members who joined that month. The list included Dario Amodei, Filip Wolski, Jack Clark, Scott Gray, and Zain Shah.
https://openai.com/index/team-update-august/
Ten years later today, Dario is the CEO of Anthropic, Jack Clark heads policy at Anthropic, and Scott Gray has marked OpenAI as his "former" employer.
As of now, OpenAI has not issued any statement regarding Scott Gray's departure, and Gray himself has not provided specific details. The common source for all current reports is the change in his social media bios, followed by a viral leak post on 𝕏. There is no public information available regarding his exact departure date, future destination, or whether he has joined a new company.
It is worth noting another detail on his account. Gray's last original 𝕏 tweet was posted on November 20, 2023, containing a single sentence: "OpenAI is nothing without its people." That was the third day after Sam Altman was ousted by the board, and over 700 employees had signed a joint letter demanding the board's resignation. This phrase was repeatedly copied and pasted within OpenAI at the time, serving as a collective pledge of loyalty. After that, although he replied to a few tweets (replies that also stopped after February 2025), he never posted an original tweet again.
Scott Gray Profile
We previously published an article specifically introducing Scott Gray; see The God of CUDA Kernels, the World's Strongest GPU Programmer? Who is this Behind-the-Scenes Mastermind at OpenAI. Here is a brief recap.
His reputation rose during his time at Nervana Systems. Back then, the vast majority of developers relied on NVIDIA's CUDA C/C++ and official libraries like cuBLAS and cuDNN. Multiple layers of software abstraction shielded hardware details but also created a performance ceiling. Gray's judgment was that these abstraction layers had to be bypassed. Therefore, he wrote maxas, an assembler for the Maxwell architecture that allowed developers to write SASS machine code directly, manually allocate registers, manage memory latency, and control instruction pipelines.
To prove this approach worked, he hand-wrote an SGEMM kernel using maxas. On a GM204 GPU, the kernel achieved 98% of the hardware's theoretical peak performance and was 4.8% faster than NVIDIA's official, closed-source, and similarly hand-written cuBLAS. The subsequent maxDNN applied the same methodology to convolutions: achieving a stable compute efficiency of 93% to 95% across all convolutional layers of AlexNet, while the contemporary cuDNN's efficiency fluctuated wildly between 32% and 57%.
This became the methodological foundation for all his subsequent work: refusing to accept the performance limits imposed by abstraction layers.
In that August 2016 OpenAI team update, the official introduction provided only two technical assessments, noting that he previously focused on optimizing deep network performance on GPUs at Nervana, and that his assembly-level optimizations for dense linear algebra and convolution were "still the fastest to date." The same paragraph also mentioned that when he wasn't coding, he was usually reading the latest research in neuroscience and related fields.
It seems that, ten years later, his new direction of "neuroscience-inspired AI methods" is a return to an interest that never actually changed.
After joining OpenAI, his role shifted from "optimizer" to "enabler." In 2017, along with Alec Radford and Durk Kingma, he released block-sparse GPU kernels. Unlike unstructured sparsity which removes individual weights, block sparsity slices the weight matrix into fixed-size blocks and zeros out entire blocks. Specialized kernels completely skip the zero-value blocks during computation, making them potentially orders of magnitude faster than cuBLAS for dense matrices or cuSPARSE for general sparse matrices. This set of kernels was open-sourced at the time and directly spawned a subsequent series of sparse attention work.
OpenAI block-sparse kernel diagram, https://cdn.openai.com/blocksparse/blocksparsepaper.pdf
Following this thread, the 2019 paper "Generating Long Sequences with Sparse Transformers" by Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever reduced the time and memory footprint of attention from O(n²) to O(n√n). The paper explicitly listed "fast attention training kernels" as one of its three main contributions.
Later on, his name appeared on the author lists for GPT-3 and "Scaling Laws for Neural Language Models," as well as the DALL·E paper and the OpenAI Five technical report.
During this period, he also earned the title of "God of CUDA Kernels." This occurred in September 2025 when former OpenAI employee Rohan Pandey posted on 𝕏 that only about one person at the company was responsible for inference-side CUDA kernels. Colleagues called his attention kernel "the Bob kernel," which executed trillions of times a day across hundreds of thousands of GPUs. After the post went viral, commenters widely speculated that "Bob" was Scott Gray.
The related post also mentioned: globally, there may be fewer than a hundred people capable of writing high-performance CUDA kernels for the training process (especially for backpropagation). This is because the skill requires simultaneous mastery of parallel computing theory, GPU hardware architecture, and deep learning algorithms, whereas most practitioners remain at the application or inference optimization layer.
And now, Scott Gray has left.
OpenAI's Brain Drain in 2026
OpenAI has lost a considerable amount of talent this year. According to a tally by 𝕏 blogger Chubby, 12 senior leaders have departed OpenAI this year, spanning operations, commercial, product, research, safety, ethics, and hardware.
A few days ago, Brad Lightcap, who spent eight years at OpenAI serving as CFO and COO, announced his departure on 𝕏, stating he wanted to "start something new." Two days later, Denise Dresser, who only took over as Chief Revenue Officer in December of last year, announced she was leaving, with Dali Rajic, former President and COO of Wiz, stepping in as her replacement.
On the commercial and product lines, Fidji Simo, CEO of the Applications business, took a medical leave of absence in April for health reasons and transitioned to a part-time advisory role in July. Former CMO Kate Rouch stepped down in April for medical treatment. Srinivas Narayanan, CTO of the enterprise business, also announced his departure in April. On the research and product side, Kevin Weil, who led OpenAI for Science, and Bill Peebles, head of Sora, announced their departures on the same day in April; Weil is currently building an AI science startup. Barret Zoph returned to OpenAI in January from Thinking Machines Lab to lead enterprise AI sales but left again in June, with the company providing no reason.
Three members of the safety and ethics team departed concurrently in July: Johannes Heidecke, head of safety systems, left amidst a reorganization of the safety team, with his duties folded into the research department led by Mia Glaese. Chief Futurist Joshua Achiam departed after nearly nine years; the mission alignment team he previously led had been dissolved in February. Chloé Bakalar, the ethics lead, left after less than a year. Earlier in March, Caitlin Kalinowski, head of robotics and consumer hardware, resigned because she disagreed with the company's partnership with the Pentagon. She stated on LinkedIn that such issues should undergo more thorough deliberation, later adding that her dissent was primarily aimed at the governance process.
This list still omits one person: in January of this year, Research Vice President for inference models Jerry Tworek resigned to start his own venture after nearly seven years, citing the desire to pursue research directions that were "very hard to do inside OpenAI."
Conclusion
Scott Gray's departure might not immediately alter OpenAI's model release cadence. A decade of accumulated engineering infrastructure won't grind to a halt because one person leaves. However, for a company increasingly reliant on scale, efficiency, and cost control, losing an engineer who could push hardware performance to its absolute limits while being deeply involved in multiple generations of core models is not a trivial change.
The timing of this departure is also noteworthy: OpenAI is preparing to go public, transforming into a corporate giant striving for revenue and participating in a global infrastructure race. Now, the very people who shaped it earliest are re-evaluating what they want to do next.
Scott Gray's choice wasn't to join another frontier lab or announce a startup; he merely stated in his bio that he is independently exploring "neuroscience-inspired AI methods." From the engineer in the 2016 official introduction who read neuroscience papers in his spare time, to rewriting neuroscience into his bio ten years later, this feels more like a return after a long detour: having squeezed the existing computing paradigm to its absolute extreme, he is now searching for an alternative possibility.
In 2023, Gray wrote, "OpenAI is nothing without its people." Looking back today, that statement was both a response to that corporate crisis and a belated footnote. OpenAI won't lose everything because one person leaves, but a company's direction is ultimately shaped by those who stay, those who leave, and the problems they each choose to explore.
© THE END
For reprints, please contact this official account for authorization.
Submissions or press inquiries: liyazhou@jiqizhixin.com