17-Year-Old Researcher Chen Guangyu Leaves Moonshot AI, Rumored Destination Points to Discovery Loop

By Letters (Zimu Bang)

These past few days, the whole internet has been looking for one person: Chen Guangyu.

The reason: rumors that he received a job offer with an annual salary of $1 million. That figure isn't the ceiling in Silicon Valley—for a senior engineer with stock, it's not unusual to hit that number. But he is only 17 this year.

Even setting aside the salary, the fact that a 17-year-old is being fought over by Silicon Valley companies is itself surreal enough.

Then came rumor after rumor about where he was going. Some said he went to Anthropic, but Chen Guangyu himself later denied that.

Then others said he went to Carnegie Mellon to continue his studies, and still others said he joined Discovery Loop, founded by Google veteran Jeff Dean, working on using AI to automate scientific research.

01

On October 15, Chen Guangyu posted on social media announcing his departure from Moonshot AI. He said the year he spent there was "the best time of my life so far."

Under the post, Anthropic researcher Sholto Douglas left a comment, first congratulating him, then asking whether "the Carnegie Mellon plan is still the same," and inviting him to come work together in San Francisco.

Chen Guangyu's reply was: "Thanks Sholto! I'm heading to the Bay Area."

Since an Anthropic researcher had extended an invitation, outsiders naturally guessed he would join Anthropic. For a while, social media was buzzing with "Chen Guangyu joined Anthropic within 24 hours of leaving Moonshot AI."

But on October 16, someone asked him directly in the comments: Is it true you're joining Anthropic? His answer was a single word: "No." And just like that, he personally denied the rumor.

On October 20, an anonymous account called Arfur Grok, which specifically tracks AI talent flows in Silicon Valley, posted that Chen Guangyu had joined Discovery Loop.

Discovery Loop was founded in August this year by four former core Google researchers: Jeff, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals.

Among them, Jeff was formerly Google's chief scientist and helped lead the development of Gemini; Vinyals was also a key technical lead on Gemini.

What this company wants to do, in plain terms, is let AI do scientific research itself. The AI first proposes hypotheses, then designs and runs experiments, analyzes the results, and then carries out the next round of research based on the feedback, looping around and around. Discovery Loop—the name means exactly that.

The company's first funding round was led by Radical Ventures and Khosla Ventures, with Alphabet, Google's parent company, also participating. According to foreign media reports, Discovery Loop has not yet released any product but is already seeking a valuation of about $50 billion.

Discovery Loop said it does not comment on personnel matters, and Chen Guangyu did not respond.

At the same time, Chen Guangyu had originally also planned to return to Carnegie Mellon to continue his studies—that was what Sholto's question about whether "the Carnegie Mellon plan is still the same" referred to.

Chen Guangyu's resume shows that in August and September this year, he studied at Carnegie Mellon for about a month.

Menlo Ventures investor Deedy Das said that a 17-year-old had received five million-dollar-level offers, though he did not name the person. He also said that this teenager flew halfway around the world to attend an undergraduate program at a top-ten US university, then dropped out two weeks later.

Netizens therefore speculated that he might be referring to Chen Guangyu.

But one point needs to be made clear here: Carnegie Mellon University's main campus is in Pittsburgh, Pennsylvania, in the eastern United States, about 4,000 kilometers from the Bay Area—across the entire country.

There is a small branch campus in the Bay Area, in Mountain View in Silicon Valley (within the NASA Ames Research Park). However, this branch campus is very small and mainly handles master's programs and some industry-academia collaboration. Carnegie Mellon's core School of Computer Science (SCS) is in Pittsburgh.

So Chen Guangyu's trip to the Bay Area is most likely not for school.

Actually, there is another thread: the Bay Area's intersection with Chen Guangyu is not limited to Discovery Loop—there is also Tilde Research.

Tilde Research is a mechanistic interpretability lab in Palo Alto, founded in July 2024 and publicly launched in November, with the slogan "understanding intelligence."

Tilde Research focuses on model internals, new architectures, and pretraining science. In mid-2025 it completed an $8 million seed round.

Between Chen Guangyu and Tilde Research, you could say it was "fate, wonderfully inexplicable."

According to Chen Guangyu's own account, he once published an article on social media when he had only 20 followers, and it was replied to by Yang Songlin—a key contributor to DeltaNet, a PhD student at MIT, and the founder of the linear attention open-source community FLA—and the post spread from there. Ben Keigwin of Tilde Research also noticed Chen Guangyu because of this and took the initiative to arrange a call with him to introduce the project.

In July 2025, Chen Guangyu traveled alone to Palo Alto, California, to begin a seven-week internship at Tilde Research.

So some in academia have speculated that Chen Guangyu wants to return to Tilde Research to continue his previous research work.

Having said all that, there is still one important question left unaddressed: how is it that a 17-year-old boy can get so many big companies and investors chasing after him?

02

On March 16, the technical report "Attention Residuals," co-first-authored by Chen Guangyu, was liked on social media by Musk himself.

Musk commented on the report: "What Moonshot AI is doing is really impressive!"

In fact, this is just one of Chen Guangyu's highlights.

Chen Guangyu occupies the position where the entire industry is most short of people.

For example, a large model is like a big restaurant. Architecture is equivalent to how the menu is designed; the optimizer is equivalent to how the chef continuously improves his craft based on feedback; and the GPU kernel is the specific operation at the kitchen stove, such as how high the flame is, how the wok is tossed, and how the ingredients are arranged.

For the same dish, as long as the kitchen puts in the effort, even with the same equipment, several times more dishes can be produced in one night.

What makes Chen Guangyu special is that he has touched all three layers.

In May 2025, he wrote about how to implement matrix multiplication with Triton and do L2 cache optimization; in the same month, he also wrote about how to derive and implement an optimizer under a nuclear norm constraint.

In July, he implemented Flash Attention with TileLang; the related paper claimed it was 1.3 times faster than FlashAttention-2, and he subsequently participated in MoMoE.

In September, he wrote two Triton explainers in a row.

In October, he wrote Gram-Space Manifold Muon, also an article about optimizers.

From low-level GPU code to the middle-layer optimizer and then to the top-level model architecture, Chen Guangyu, who was only 16 at the time, single-handedly walked a road that others may not be able to walk in a lifetime.

He also put it very bluntly on his own homepage: what he cares about is "efficient, hardware-aligned algorithms and kernels," and he even specifically added the line "co-design them!"

This is a track that has only just emerged, and there is no "experience" to speak of.

An analysis by BigGo Finance mentioned that in the era of large models, in frontier fields such as low-level operator optimization and the mathematical derivation of attention mechanisms, there are no ready-made textbooks to copy from, and the smartest people in the world are actually standing on the same starting line.

The most crucial point is that Chen Guangyu truly "understands."

Chen Guangyu has an article titled "Triton Flash Attention Kernel Walkthrough: The Forward Pass."

At the beginning, Chen Guangyu says, I am not going to talk about basic syntax—the official documentation already explains that very well. This article is about understanding the why.

There is a detail in the article: Chen Guangyu's code splits the loop into two parts.

His explanation for this is that because the masking operation tl.where is very slow, using it on every block would slow things down. Splitting it into two parts—"outside the diagonal" and "on the diagonal"—means it is only used where necessary.

Using the kitchen example again: ingredients need to be checked before they go on the cutting board. The old rule was that no matter what the ingredients were, they all had to be checked once. But in reality, 99% of ingredients do not need to be checked at all, because they were already checked before being put in; only those ingredients that are prone to rotting and damage need to be checked.

Chen Guangyu's approach is equivalent to checking only the 1% that spoils easily, while the other 99% pass directly. This indirectly increases efficiency by 100 times.

In addition, Chen Guangyu is also someone who is very good at learning, with an astonishing rate of growth. Chen Guangyu once said in September 2025 that he had not been exposed to Transformer for long. When he first seriously encountered scientific research, Chen Guangyu said he understood nothing—he did not know what Transformer was, and he could not understand machine learning papers.

Today, Chen Guangyu has already become a leading figure in the industry.

The main reason for this is that his learning method is also very special. Chen Guangyu said that he started with a classic paper and spent several hours a day conversing with Gemini. When he encountered something he did not understand, he would send a screenshot to it, ask it to explain, and then ask it questions in turn.

At the same time, Chen Guangyu also reads GitHub, blogs, and researchers' personal homepages to study their "research taste." He does not follow the traditional learning route of studying one subject after another; Chen Guangyu thinks that is too slow.

Of course, AI is not omnipotent either. When facing major decisions, Chen Guangyu would consult DeepSeek researcher Yuan Jingyang. Chen Guangyu himself summed it up by saying that networks are important. Even if you are confident, you can still learn a lot of experience from your predecessors.

Chen Guangyu has a rating of over 2,000 on Codeforces (a programming competition platform), placing him in roughly the top 3%. This shows that he has a solid programming foundation.

03

After Chen Guangyu announced his departure, Du Yulun, the Moonshot AI researcher who made the call to hire him back then, commented under his departure post, saying that he and Chen Guangyu had "several long talks" about this decision.

He also said that seeing Chen Guangyu leave made him "sad and heartbroken," but at this age, exploring as much as possible and following new curiosities may be the better choice.

There is also a small detail here: Du Yulun also graduated from Carnegie Mellon.

Departures inevitably come with some reluctance, but they also leave something behind for the company. For example, DeepSeek's Guo Daya: after he left at the beginning of this year, the GRPO algorithm he had published continued to be used by DeepSeek to this day. DeepSeek's V4.1 Flash technical report mentioned that the first step of post-training was supervised fine-tuning (SFT) and GRPO-based reinforcement learning, respectively training expert models in various domains.

So what exactly did Chen Guangyu leave behind for Moonshot AI?

Chen Guangyu stayed at Moonshot AI for less than a year, joining in November 2025 as a 16-year-old intern and leaving in October 2026. But what he left behind has already penetrated the most core position of Kimi.

In July 2026, Kimi K3 was released, with 2.8 trillion total parameters. Moonshot AI called it the world's first open-weight 3-trillion-level model.

On the day K3 was released, on Artificial Analysis's intelligence index, K3 scored 57, ranking third, behind only Claude Fable 5 and GPT-5.6 Sol, and surpassing Claude Opus 4.8.

The Attention Residuals mentioned earlier is one of K3's underlying technologies.

Chen Guangyu also participated in the computational kernel acceleration of Moonshot AI's KDA (Kimi Delta Attention). In his departure post, he wrote that he and Zhang Yu together optimized its kernel and explored various possible designs.

However, around the time Chen Guangyu left Moonshot AI, news about K3.1 began to appear intensively.

On October 10, a preview page appeared on the open platform: 1 million token context, Low/High/Max three-tier reasoning, Agent, and Swarm multi-agent capabilities. Nomura Securities said that "K3.1 will be released before mid-October."

Around October 17, "k3d1-agent" appeared again on the official website's front end, and a new bot module was discovered; the official website's model fingerprint was also reported to have changed.

Judging from the current clues, K3.1 looks more like an enhanced version of K3, leaning toward Agent and long tasks, and has not changed its architecture.

So even though Chen Guangyu has left, his contribution to Moonshot AI is still there for all to see.

People leave, but the technology remains.