Nathan Chen Reportedly Heads to Le Viet Quoc's Discovery Loop as Startup Seeks $50 Billion Valuation

A teenage Kimi researcher is linked to a new US lab

Nathan Chen Guangyu, a 17-year-old researcher who helped develop Moonshot AI's Kimi models, is reportedly joining Discovery Loop, the US startup founded by former Google researchers Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Le Viet Quoc. The hiring claim appeared on October 20, five days after Chen announced his departure from Moonshot to move to the United States.

The proposed move would bring a researcher with experience improving the cost efficiency of Chinese AI models into a company seeking to automate scientific research. Discovery Loop has yet to release a public product, but it has attracted investment from Alphabet and is reportedly seeking new financing at a valuation of about $50 billion.

Chen's appointment has not been publicly confirmed. He did not respond to a request for comment, while Discovery Loop said it does not comment on personnel matters. The hiring claim came from Arfur Grok, an anonymous social media account that tracks movements among AI researchers.

The account previously identified moves including Luke Metz's departure from OpenAI for Meta Superintelligence Labs in August and Andrew Tulloch's departure from Meta for Anthropic in September. Its latest claim places Chen at the intersection of two closely connected contests: the competition to recruit exceptional researchers and the effort to build AI that can carry out research itself.

For Moonshot, Chen's departure means losing a contributor to its flagship model family after less than a year. For Discovery Loop, his reported arrival would add practical experience in model architecture and efficiency to a team pursuing a much broader research ambition.

What Chen worked on at Moonshot

Chen joined Beijing based Moonshot AI in 2025 after an internship at a Silicon Valley startup. Born in 2009, he was also selected for the Rhodes Trust's Rise fellowship, a global talent programme for teenagers aged 15 to 17 funded by former Google chief executive Eric Schmidt and his wife, Wendy Schmidt.

At Moonshot, Chen reportedly worked on the Attention Residuals architecture and the Kimi K3 model. His contributions included technical changes that helped reduce costs across the Kimi model family, which competes with systems such as OpenAI's GPT and Anthropic's Claude.

Model architecture is the underlying design that determines how an AI system processes information. Improvements to that design can help a developer get more useful work from its computing resources. For a company serving many users, efficiency affects how much it costs to train and operate its models.

Kimi has also attracted attention as a family of systems with open weights. Weights are the numerical settings a model learns during training. Making them available allows other developers to download and adapt a model, subject to its licence, rather than access it only through the original company's service.

A departure discussed with a senior colleague

Chen announced his departure on social media, describing his time at Moonshot as the best year of his life so far. His exit followed discussions with Du Yulun, a senior researcher at the company.

Du wrote that he and Chen had spoken at length about the decision:

several long conversations about his decision

Du said the best course for the young researcher was to have room to explore. He described it as:

best for [Chen] to explore and follow his new curiosity as much as possible at his age

Those comments give a personal dimension to a move attracting attention because of competition between Chinese technology hubs and Silicon Valley. Chen had already gained experience in both countries before his reported recruitment by Discovery Loop.

Four founders and an ambition beyond chatbots

Discovery Loop launched in August with four founders drawn from Google's research leadership. Dean was Google's chief scientist. Ghemawat helped design major computing systems including MapReduce and Bigtable, Le was a founding member of Google Brain, and Vinyals was a leading researcher at Google DeepMind.

The company was established as a Delaware public benefit corporation, a legal structure that allows a business to pursue a stated public benefit alongside financial goals. Its research mission is to automate scientific discovery, beginning with AI development and potentially expanding into healthcare, materials and energy.

Discovery Loop has also recruited researchers from large US technology companies, including former Google engineers. Chen would likely be among its youngest employees if his appointment goes ahead.

The attraction for someone with his background is a research programme that connects directly with model development. Rather than concentrating solely on a chatbot or another application for customers, Discovery Loop wants to build systems that can help invent and test the next generation of AI.

How AI could run a research loop

Discovery Loop's proposed system would generate research ideas, design and run experiments, examine the results and decide what to try next. The goal is to perform much of that sequence without a person directing every step.

A hypothesis is a proposed explanation or idea that can be tested. In AI research, a system might propose a change to a model's design, test whether the change improves performance, and use the result to choose another experiment. Repeating that process creates a research loop.

Dean discussed this approach at the Agentic AI Summit at the University of California, Berkeley, in early August. He described a future in which AI could run thousands of scientific experiments in parallel, meaning at the same time, and use their results to identify the most promising next steps.

Discovery Loop plans to begin with machine learning algorithms and systems. That gives its proposed research tools an initial subject closely matched to the founding team's experience, before any expansion into other sciences.

What recursive self-improvement means

Dean connected the company's work with recursive self-improvement, often shortened to RSI. In plain language, this means using AI to improve the methods and systems used to build AI, then applying the improved system to further development.

Leading AI companies in both China and the United States are pursuing versions of that goal. Discovery Loop's approach links it to a broader vision of automated research: a system that learns from experiments and uses that knowledge to plan better ones.

Dean, explaining the scientific knowledge such a system would need, said:

By building models that are really good at understanding many, many different domains of science and engineering, you can get PhD-level expertise in a model across many different domains,

His vision includes shortening some research processes from years to hours or even minutes by testing many possibilities simultaneously. That remains an ambition for the technology. Discovery Loop has not released a public product demonstrating that research cycle.

Chen's reported experience improving model efficiency fits the company's initial focus. Automated AI research would involve running and comparing many experiments, making the computing cost of each attempt a practical concern.

The $50 billion figure is a fundraising target

Discovery Loop's financial backing includes Radical Ventures and Khosla Ventures, which led its initial financing, and Alphabet, Google's parent company. Other named participants include Lightspeed, Kleiner Perkins and Doerr Capital. Alphabet has also been described as a cloud partner, connecting the startup with computing infrastructure as well as investment.

September fundraising discussions reportedly sought a valuation of about $50 billion. Some descriptions placed that value after the first funding round; the more detailed financing accounts describe it as a target for another round, with terms still under negotiation and no completed transaction at that price announced.

The target followed discussions weeks earlier about raising approximately $1 billion at a valuation near $10 billion. Moving from $10 billion to $50 billion would mean a fivefold increase, or a difference of $40 billion, in the proposed value of the company.

A valuation is the price investors assign to the whole business when negotiating an ownership stake. The amount raised is the money the company receives. Discovery Loop's $50 billion target and the earlier proposed $1 billion investment describe different parts of a funding negotiation.

The company was already pursuing that larger price before Chen's reported recruitment, following its September funding discussions. Its founders, research plans and ability to recruit staff are central to the investment story while it works toward a public product.

How the developments unfolded

Discovery Loop's launch, financing discussions and reported recruitment of Chen took place over a short period. The sequence shows that the company was assembling its team while seeking a sharply higher valuation.

• 2025: Chen joined Moonshot AI after an internship at a Silicon Valley startup.
• August 5, 2026: Discovery Loop announced its launch and initial financing, with Dean, Ghemawat, Vinyals and Le as founders.
• Early August: Dean discussed automated research and recursive self-improvement at the Agentic AI Summit in Berkeley.
• August: Fundraising discussions reportedly involved approximately $1 billion at a valuation near $10 billion.
• September: New financing discussions reportedly targeted a valuation of about $50 billion.
• October 15: Chen reportedly announced his departure from Moonshot and move to the United States.
• October 20: Arfur Grok reported that Chen would join Discovery Loop.

The next concrete developments would be an announcement about Chen's role, a financing agreement or a public release from the company. No dates have been announced for those steps.

Key Points

• Nathan Chen Guangyu, 17, is reportedly joining Discovery Loop after leaving Moonshot AI.
• Chen worked on Kimi K3, Attention Residuals and improvements to model cost efficiency.
• Discovery Loop was founded by Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Le Viet Quoc.
• The startup wants AI to propose experiments, run them and use the results to guide further research.
• Its reported $50 billion valuation is a fundraising target, with no completed deal announced at that price.
• Radical Ventures, Khosla Ventures and Alphabet are among its initial investors.