Core 17-Year-Old Kimi Architect Resigns, Set to Move to the Bay Area, Sparking Speculation About the Next Destination.
17-year-old Kimi Core Architect Leaves, Confirms Move to the Bay Area, Next Destination Sparks Speculation
Summary
Author｜Dongmei 1 17-year-old Kimi core architect leaves, denies going to Anthropic. On October 15, Chen Guangyu (nathan chen) announced on X that he recently
On October 15, Chen Guangyu (Nathan Chen) announced on X that he has recently left Moon’s Dark Side. He described his past year at Kimi as the best time of his life so far, and thanked the team for giving him an opportunity when he was 16.
Anthropic researcher Sholto Douglas invited Chen Guangyu in the comment section to come to San Francisco, writing: “Congrats, brother! Great to see you last week. Is the Carnegie Mellon plan still the same? You should come to San Francisco and work with Prime/Base10!”
Chen Guangyu later replied, confirming that he will head to the San Francisco Bay Area in the United States.
This interaction sparked speculation about his joining Anthropic and participating in Claude’s development. Someone asked in the comments if he was joining Anthropic, and Chen Guangyu quickly gave a very direct answer: No.
As of publication, we can confirm from Chen Guangyu’s own statements that he has indeed left Kimi and will move to the Bay Area, but his next employer, position, and start date remain unknown.
Regardless of where he ultimately chooses to go, some of the technical contributions this young researcher has left behind are worth noting: co‑first author of Attention Residuals, author of the MoMoE project, and a series of research and technical shares focused on optimizers, GPU kernels, and model training efficiency.
Chen Guangyu’s previous personal homepage indicated that he worked at Kimi on large language model scaling and pre‑training, involving Kimi K3, with research focus on model architecture and optimizers, while also paying attention to efficient, hardware‑adapted algorithms and compute kernels. Before joining Kimi, he conducted research on application interpretability at Tilde Research. It should be noted that this homepage still retained the description of his work at Kimi at the time of retrieval, and his employment status should be based on the latest resignation statement.
According to a previous report by Xinhua Finance, Chen Guangyu is from Shenzhen and graduated from Huizhou China Resources Xiaojingwan Besis International School’s high school in 2026. His identity as a “17‑year‑old researcher” attracted attention after the release of the Attention Residuals paper in March of this year.
The Shenzhen Special Zone Daily recorded his research starting point at the time: by reading classic papers and following GitHub open‑source projects to fill in the basics, then sharing technical thoughts on social platforms, which caught the attention of a leader at a Silicon Valley AI startup. After a timed experiment, he earned an opportunity to intern in the United States and worked there for seven weeks.
In November 2025, he joined the Kimi team as an intern.
From his personal blog, it is clear that he actually contributed to the research of these projects, not just a name on a paper.
In his 2025 technical articles, he discussed using Triton to implement matrix multiplication and L2 cache optimization, using TileLang to implement Flash Attention, and the forward computation of Triton Flash Attention. These topics focus on a concrete problem: how to turn mathematically described algorithms into programs that run efficiently on GPUs.
These works are consistent with his later research directions. Model architecture determines how computation is organized, optimizers affect how training updates parameters, and compute kernels determine how fast these operations run on hardware. From his public projects, his interests span these interrelated layers.
Guangyu Chen’s most notable achievement is the publicly released Attention Residuals (《Attention Residuals》) on March 16, 2026. He, along with Yu Zhang and Jianlin Su, are listed as co‑first authors with equal contributions.
This study focuses on how different layers within a large model transmit and utilize information.
In common PreNorm Transformers, residual connections add each layer’s output with a fixed weight. The paper points out that as the network deepens, this accumulation can cause the magnitude of hidden states to keep growing, gradually diluting the contribution of individual layers. The idea of Attention Residuals is to replace the fixed addition with cross‑layer softmax attention, allowing the model to selectively aggregate representations from previous layers using learnable, input‑dependent weights.
You can think of it as: after processing through many layers, the model no longer only receives the continuously summed result of prior information, but gains a mechanism to re‑select and recombine early representations. Here, “attention” operates along the depth of the model, which is different from the usual attention between tokens.
However, allowing each layer to access the outputs of all previous layers also increases storage and communication overhead. To address this, the team further proposed Block AttnRes: dividing the network layers into multiple blocks, retaining ordinary residual accumulation within each block, and performing attention aggregation between blocks to reduce extra cost.
When citing Su Jianlin’s research review, China News Weekly reported that the initial idea and small‑scale experiments were proposed by Su Jianlin, while Zhang Yu and Chen Guangyu later participated in larger‑scale validation and jointly introduced the Block Attention Residuals design presented in the paper. This division of labor indicates that Chen Guangyu’s significant contribution includes advancing the preliminary scheme into a form more suitable for large‑scale training.
The paper validated the approach on the Kimi Linear architecture with a total of 48 billion parameters and 3 billion activation parameters, using a pre‑training dataset of 1.4 trillion tokens. Official project documentation states that in their scaling experiments, Block AttnRes can achieve the loss level of a baseline model that would otherwise require 25% more compute. This conclusion applies to specific experimental conditions and should not be generalized to claim that all model training will obtain the same benefit.
In fact, before joining Attention Residuals, Chen Guangyu had already participated in Tilde Research’s MoMoE project.
The paper “MoMoE: Memory‑optimized Mixture of Experts,” released by Tilde in July 2025, lists Nathan Chen as one of the authors. His personal blog also features this work.
MoMoE addresses the engineering efficiency challenges of mixture-of-experts models. MoE reduces active computation by invoking only a subset of experts for each token, but practical implementations can still be limited by suboptimal data layout, fragmented computation, and intermediate tensor storage. MoMoE fuses Triton kernels, optimizes memory layout, and offers configurable back‑propagation schemes, trading off between compute and memory usage, with a focus on model training and fine‑tuning.
The value of this work lies in turning the theoretical advantages of the model architecture into actual throughput and VRAM savings through low‑level implementation. It is continuous with Chen Guangyu’s later focus on hardware‑adapted algorithms and compute‑kernel directions.
He also participated in optimizer‑related research. The paper “Gram‑Space Manifold Muon,” released by Tilde in October 2025, lists Ben Keigwin, Dhruv Pai, and Nathan Chen as authors; the work is also listed on his personal website. Compared with a vague claim of “model development,” these traceable projects more concretely illustrate his research scope.
In addition, China News Weekly reported that one of the earliest tasks he took on after joining Kimi was accelerating KDA, the computation kernel of Kimi Delta Attention.
During the spread of the resignation news, discussions also emerged about million‑dollar job offers, dropping out, and specific positions.
A few days ago, investor Deedy Das shared on X the story of a 17‑year‑old who received multiple lab invitations. Deedy wrote on X: “Today I met a kid who, at 14, wasn’t very good at English but taught himself AI with large models, interned for a year at a top AI company, then flew across half the globe to attend a top‑10 U.S. university for undergrad; dropped out two weeks later; and at 17 received five job offers from a leading new lab in San Francisco, each with an annual salary over $1 million.”
Deedy also lamented that traditional education has completely disappeared, and today’s kids are fundamentally different from those of the past.
Online discussions later linked it to Chen Guangyu, believing that the prodigy mentioned by the investor was Chen Guangyu.
But these are just speculation; Deedy later did not reveal the real name of the teenager he met.
At 17, Chen Guangyu’s research career is just beginning. From reading papers and writing compute kernels to participating in large‑model architecture research, he has already turned curiosity into concrete results. After moving to the Bay Area, he will encounter new problems, partners, and choices.
More exciting than which company he’ll join next is what new surprises this young man will bring to the industry. For him, the future offers plenty of time to explore, and enough space for his work to stay ahead of the “genius” label.
Reference links:
https://www.chinanews.com/sh/2026/08-10/10675225.shtml
https://arxiv.org/abs/2603.15031
https://x.com/nathancgy4/status/2110522061953327565?utm
Disclaimer: Original content from InfoQ, does not represent the platform’s views, does not constitute investment advice, and unauthorized reproduction is prohibited.
The QCon Global Software Development Conference 2026 (Shanghai) will be held from October 22–24. This year’s conference focuses on engineering practices in the Harness AI era, covering everything from “building AI” to “harnessing AI.” Topics include AI‑Native architecture, Agent Runtime, AI Infra, Agent security and observability, Loop Engineering, embodied intelligence, and world models. We invite the global tech community and frontline industry practitioners to share the most valuable engineering experiences of the AI‑Native era. For more details, scan the QR code or contact ticket manager at 18514549229 for inquiries.
Get MirrorSpeed