The Young Man Who Went to Baidu to Build Models

@Yunxiao Guo

Yesterday, news broke that Tianxiang Sun has joined Baidu as head of the Basic Model Unit (BMU).

Elsewhere first learned of Sun through investor Rui Han. About a year and a half ago, Han told us he had backed a "young talent" — Tianxiang Sun, born in 1997. Sun had just left Luyu Yang's company at the time.

The company he founded is called Rixingji (日行迹). It's an AI4S startup. As we understand it, the company has now been acquired by Baidu.

Rixingji takes its name from the astronomical term "analemma" — the figure-eight pattern traced by recording the sun's position at the same time and place each day over the course of a year. The company's slogan: "In a world of infinite problems, we need to build infinite minds."

What struck me most was how long it had been since I'd heard someone say "AGI" with such density.

For instance, around the time he started his company, he passed on opportunities in video models because "he felt it had nothing to do with AGI." He wanted to use AI to research AI — work that still belongs to today's most sought-after AI researchers. He believed the simplest methods are most effective, and that all of AGI might essentially reduce to a language model capable of studying AI itself. And he kept returning to his belief in non-consensus thinking and his hunger for change.

During those conversations, we also discussed Rixingji's ongoing fundraising, his observations on domestic AI labs, and his argument for why what he wanted to do would be difficult for top model vendors to pull off.

Sun is from Shandong. He once joked to us that Shandong is an "entrepreneurial black hole" — a place where consensus is too strong. And consensus means probable failure, especially for those who arrive late.

For a top AI talent to join Baidu today is, indeed, rather non-consensus. I haven't asked Sun why. But for someone who believes in non-consensus, he must see something the rest of us don't.

AI for AI

Rixingji's original vision: use AI to research AI.

How to understand this?

Currently, foundation models still rely heavily on top AI researchers "hand-crafting" them — which means model development remains bottlenecked by humans. Sun's view is that once AI's ability to research AI reaches human level, the rate of AI capability improvement will steepen significantly.

Analogous to human researchers, AI doing research also has two components: an Innovator — a hallucination generator that can produce research hypotheses at human level and scale; and a Verifier — an experimental agent that takes any hypothesis, writes code, runs experiments on GPU clusters, and tells you whether the result becomes knowledge or remains hallucination.

This past February, Rixingji held a livestream lasting over 200 hours, during which FARS (Fully Automated Research System), developed by the company, wrote 100 short papers on the spot.

Simply put, Rixingji wanted to build the "builder of AI." More ambitiously, it aimed to build the organizational capacity of a DeepMind, DeepSeek, or OpenAI itself.

NeoLab

We initially wondered: was Rixingji something of a sci-fi-tinged NeoLab?

Sun's answer was measured: "It still counts as a lab-style company." He believed a lab's atmosphere should be pure — no top-down organizational authority, but rather bottom-up expert authority. All meetings visible to everyone. Freedom to propose ideas and form research teams. This, he felt, was part of what gave DeepSeek its lab character.

We also discussed DeepSeek at the time. He said Wenfeng Liang's rare qualities were: first principles, long-termism, not worshipping the outside (neither Silicon Valley nor overseas talent), and good technical intuition. For example, in 2023, when many were binding multimodality, video, and AGI together, Liang's judgment was: none of this matters.

Building on this, his definition of NeoLab: a new technological paradigm for disruptive innovation. One key characteristic: path advantage outweighs resource advantage.

Of course, based on all this, Sun believed there are actually very few true labs in China, and even fewer NeoLabs.

Organization and People

Somewhat paradoxically, in our previous conversations, he had repeatedly expressed that a paradigm shift like AI for AI would struggle to emerge naturally within foundation model companies.

In recent years, virtually all foundation model teams have placed organization and talent at the highest priority, with top researchers courted aggressively by major domestic and international tech companies. The fundamental reason: model training demands enormous compute and time costs. Researchers' capabilities, their research taste, and whether the organization can unlock their potential — these can create order-of-magnitude differences in compute utilization, training efficiency, and model capability.

Some have even suggested that models are essentially a distillation of the smartest humans — AI researchers themselves.

But an organization struggles to fight itself. People are hired at great expense; organizational form and culture are iteratively tuned. If AI can truly do AI research, what it challenges extends from the toolchain to the very research organizations composed of researchers today. So the faster an organization's current progress, the greater its resistance to entering the next paradigm.

And now, Sun has chosen Baidu.

Baidu

As everyone knows, how hungry Baidu is for AI.

The prehistory needs no elaboration. Just one point: few may have noticed that Baidu's latest-generation foundation model is Wenxin / ERNIE 5.0, a model emphasizing multimodal capabilities released nearly nine months ago — in November 2025. It has since received updates to context length and video capabilities.

Many believe that today's Baidu lacks innovation capacity and technical brand appeal.

For Sun, might this instead represent a greater opportunity: a person who sees reshaping organization as a technical path, arriving at an organization with everything to rebuild.

Lose Interest Once Proven

As a child, Sun changed schools several times due to family job transfers. After transferring, he could usually rise to top of his class quickly, but not necessarily maintain it. He reflected on this and concluded: once he's proven himself, he loses motivation.

The typical dilemma was the gaokao. The day-in-day-out persistence and grinding — not his preferred rhythm. So when he reached university, he finally felt in his element: play when he wanted, until sick of it; rediscover joy in learning, spend time reading in the library. Frequent state-switching actually made him enjoy things more.

During his PhD, he spent much of his time playing — perhaps an entire summer chasing anime, grinding games. In rare concentrated bursts, he'd pursue things with passion and curiosity. This approach remained viable in Fudan's free atmosphere. In February 2023, he led the development of the MOSS large model, which entered open beta — China's first publicly beta-tested ChatGPT-like model.

This perhaps explains many of his choices: he won't sustain competition on consensus tracks for long, nor is he suited to extension work; he needs to keep proving new judgments, entering problems with greater change.

There's a phrase he's repeated many times: people should go where change is greater.

Cover image: Joseph Wright of Derby, Two Boys by Candlelight, Blowing on a Firebrand, 1766, Derby Museum and Art Gallery