17-Year-Old Researcher Chen Guangyu Leaves Kimi and Receives Invitation from Anthropic
17-year-old AI researcher Nathan Chen (Chen Guangyu) announced he has left the Kimi team at Moon's Dark Side, having joined at the age of 16 in November 2025 and worked for less than a year.
He is a co-author of the Attention Residuals paper alongside Zhang Yu and Su Jianlin, which replaced traditional fixed residual connections with deep attention mechanisms and developed Block AttnRes to reduce memory and communication overhead, later applied to the Kimi K3 architecture. Musk publicly praised it as "Impressive work from Kimi."
During his time at Kimi, he also participated in optimizing the GPU kernel for Kimi Delta Attention, researched training optimizers, and assisted in creating demonstration cases for the K3 release. In his resignation post, he stated that the past year was one of the best experiences of his life and thanked the team for the opportunity given to him at 16.
Anthropic researcher Sholto Douglas congratulated him in a comment and invited him to San Francisco to work with the team, asking if he still plans to go to CMU; discussions indicate he is heading to the Bay Area.
The paper achieved about 1.25 times computational efficiency advantage after pre-training 14 trillion tokens on the Kimi Linear 48B model and showed significant improvements on inference benchmarks like GPQA-Diamond, becoming part of the underlying technology for K3.
In market dynamics, buyers are leading labs like Anthropic competing for young architectural talent, while sellers are the training opportunities from Chinese open-source teams; the event is driven by public research output and talent mobility, with resources and attention shifting from the Kimi team to Bay Area labs, benefiting cutting-edge institutions that gain early contributors, while the original team faces pressure from losing core young researchers.
His journey began in 2024 with exposure to machine learning, leading to co-authoring a rewrite of the decade-old residual structure, reflecting the accelerating effect of open-source weights and papers on researchers from non-traditional backgrounds.
Source: Public Information
ABAB AI Insight
Chen Guangyu was brought into the core team of Kimi at the age of 16, participating in the Block design of Attention Residuals and optimizing the KDA kernel, transitioning the architecture from theory to practical training in K3; previously, Su Jianlin proposed the initial idea, and he and Zhang Yu refined the large-scale trainable version, forming a short-cycle path from concept to product realization.
The capital path is reflected in quickly accumulating visibility through public papers and open-source code, attracting direct interaction from labs like Anthropic; the shift from Kimi's open-source model contributions to Bay Area cutting-edge labs makes personal research output a signal for cross-institutional talent mobility, rather than traditional academic credentials or competition paths.
This is similar to historical cases where young researchers entered top labs through public work, akin to paths taken by RoPE inventor Su Jianlin and others who gained industry recognition through architectural innovation; currently, we are in a phase where open-source models coexist with cutting-edge closed labs, with talent flowing from Chinese open-source teams to American cutting-edge labs.
Essentially, this represents a technological replacement: public architectural innovation and open-source implementations are beginning to replace traditional academic qualifications, allowing a 17-year-old researcher to directly enter core areas like Scaling RL through concrete contributions, shifting talent selection from educational background to verifiable model improvement output.
ABAB News · Cognitive Laws
Public contributions open doors faster than degrees
Open-source models are accelerators for young researchers
One architectural change equals crossing a decade-long threshold