Yao Shunyu left OpenAI and started the second half.

Yao Shunyu, who initiated the "second half" for large models, has also embarked on the second half of his personal AI journey.

Recently, the activities of this star Chinese researcher at OpenAI have started to attract intense attention.

Previously, some said he was on Mark Zuckerberg's must - recruit list. Recently, others have reported that he is about to join another tech giant, specifically a Chinese tech giant, and there are even astonishing rumors about the "transfer fee." There is also a claim that Yao Shunyu has chosen to start his own business...

Where is Yao Shunyu going? It's still unknown.

However, Yao Shunyu's departure from OpenAI has been confirmed through various channels, pending only his personal official announcement.

It seems that he is about to start the second half of his personal AI journey.

This young man, who is just 29 years old, graduated from Hefei No.1 High School, won a silver medal in the NOI Olympiad, scored 704 in the college entrance examination, and entered the Yao Class at Tsinghua University as the third - ranked student in Anhui Province. Finally, he obtained a Ph.D. in computer science (in the fields of language and reinforcement learning) from Princeton University and joined OpenAI upon graduation...

Even earlier, Yao Shunyu had prominent and well - known scientific research achievements, such as:

Tree of Thoughts: Allows large language models (LLMs) to think repeatedly, significantly improving reasoning ability.

SWE - bench: A dataset for evaluating the capabilities of large models.

SWE - agent: An open - source AI programmer.

ReAct...

He even has philosophical insights beyond his age. A blog post titled "The Second Half of AI" has become extremely popular both inside and outside the AI circle.

So, what kind of young man is Yao Shunyu exactly?

Yao Shunyu's Growth Path

Shortly after joining OpenAI in 2024, Yao Shunyu recommended a book in an interview - "Gödel, Escher, Bach: An Eternal Golden Braid".

This classic work written by the pioneer of artificial intelligence, Douglas Hofstadter, skillfully integrates Gödel's incompleteness theorem, Escher's illusion paintings, and Bach's polyphonic canons, demonstrating how these seemingly unrelated elements echo each other in the general recursive system of computers, and leaving a profound message: Seek, and you shall find.

Just as inspired by this book, an interdisciplinary perspective and an open attitude towards complex information seem to have permeated Yao Shunyu's entire academic career, gradually forming a personal style and characteristic.

Like all top students, Yao Shunyu attended Hefei No.45 Middle School (2009 - 2012), one of the best schools in Hefei, during his junior high school years. After graduating from junior high school, he was promoted to Hefei No.1 High School.

In 2014, he won a silver medal with a score of 495 in the National Olympiad in Informatics (NOI). In the following year's college entrance examination, he scored 704 in science, ranking third in Anhui Province, and entered the Yao Class at the Institute for Interdisciplinary Information Sciences at Tsinghua University, majoring in computer science.

Behind this seemingly "standard top - student" start, there actually lies a somewhat different, even slightly rebellious temperament.

△ From Qing Xiaohua

Yao Shunyu revealed in an interview that compared with other students in the Yao Class at Tsinghua University who focus on a single area and dig deeper continuously, he prefers to read a lot of materials in mathematics, history, and all kinds of miscellaneous fields.

His love for hip - hop music is no longer news.

Rappers such as Eminem, Eggplant Egg, MC HotDog, and J. Cole accompanied him through his junior and senior high school years. At Tsinghua University, he was also one of the co - founders of the Tsinghua University Student Rap Club.

It is worth mentioning that at the opening ceremony of the 2019 independent selection re - examination for various types at Tsinghua University, Yao Shunyu gave a so - called "freestyle" reason to explain why he chose Tsinghua University:

But for me, choosing between Tsinghua and Peking University is not a problem because Peking University doesn't have a class named after my surname.

In addition to being the co - founder of the rap club, Yao Shunyu also served as a recruitment volunteer at Tsinghua University and the chairman of the Yao Class Alumni Association. When talking about the influence of the Yao Class on him, he mentioned:

The Yao Class attaches great importance to the study of theoretical foundation courses, such as courses related to operating systems or circuit design. These courses seem to have nothing to do with scientific research at first glance, but in hindsight, they are still helpful. They can give you a basic understanding of the overall picture of computer science.

△ (Group photo of the Yao Class, Tsinghua University Recruitment WeChat official account)

This cross - boundary temperament of being able to accommodate various types of information, engage in different disciplines, and find pleasure in them is particularly evident in Yao Shunyu. In Isaiah Berlin's words, he is more like a "fox" rather than a "hedgehog", and this is also reflected in his subsequent research.

(Note: Berlin's "hedgehog" metaphor refers to those who focus on a single core concept and apply all experiences to this central view; while the "fox" refers to those who have a wide range of knowledge and are good at flexibly dealing with different problems, relying more on diverse strategies and perspectives)

Surprisingly, before the second semester of his junior year, Yao Shunyu had neither been exposed to AI nor engaged in scientific research.

An exchange opportunity took him to MIT. After that, he began to conduct some research in computer science, computer vision, and cognitive science under the guidance of Wu Jiajun.

Yao Shunyu said that when studying under Wu Jiajun and his senior Jun - Yan Zhu, he not only mastered the basic skills of research such as experiments and presentations but also was deeply influenced ideologically. He thus recognized the intersection of psychology and artificial intelligence and learned to think about problems from a higher - dimensional and overall framework.

After four years of study in the Yao Class, in 2019, he officially went to Princeton to pursue a Ph.D., and his cross - boundary temperament was manifested again.

During his undergraduate years, he mainly studied computer vision. However, in his Ph.D. program, although he was initially admitted to the computer vision direction, he changed his interest and contacted a tutor in the field of natural language processing (NLP) on his own. Eventually, by chance, he joined Karthik Narasimhan's team and started researching natural language processing and reinforcement learning.

This was equivalent to shifting from vision to language, but "Seek, and you shall find". Yao Shunyu later recalled that this cross - boundary move was also his lucky break because it coincided with the rise of GPT - 2. Therefore, in his first year of the Ph.D. program, he was already thinking about how to turn language models into agents.

His focus on (general) agents has run through his research.

Yao Shunyu's first work during his Ph.D. was called CALM (2020), which studied how to use language models as agents to play language games.

In CALM, language acts like a medium: it transforms human experiences and semantic patterns into actionable action candidates while carrying context information, enabling agents to make efficient decisions in a vast action space.

Yao Shunyu said that although this work is not as well - known as swebench, ReAct, or the Tree of Thoughts, it is of great significance to him.

In a conversation with Zhang Xiaojun, we found that Yao Shunyu's attention to language had already germinated in this paper five years ago.

Language is a tool invented by humans for generalization, which is more fundamental than other things.

In other words, playing games with language is infinite. Agents can reason and combine based on language, find appropriate actions in different contexts, and generalization comes from this.

In other words, agents also need to have "cross - boundary" capabilities, and language is an excellent medium.

However, Yao Shunyu also realized that without a good task or environment, even if an agent gets a high score in the "game", it is meaningless.

Based on this thinking, his second work, WebShop, constructed a large - scale simulated e - commerce environment, enabling agents to navigate and operate on web pages by understanding complex text instructions, thus promoting the application and verification of language understanding and decision - making abilities in real - world tasks.

Similarly, the later classic works SWE - Bench and SWE - agent also verified the capabilities of agents for a meaningful task (real - world programming).

When it came to 2022, the emergence of GPT - 3.5 changed everything.

As is well - known, the efficiency of repeated trial - and - error by a blank - slate agent is extremely low, and this inefficient attempt is common in traditional reinforcement learning: agents are usually either restricted to performing a single task, such as playing Go, or blindly exploring in a vast action space.

GPT - 3.5 made people realize that what was previously missing was prior knowledge: through powerful language pre - training, common sense and language knowledge are integrated into the model, and then through fine - tuning, it can become an agent with cross - boundary temperament and generalization ability.

As Yao Shunyu said: If your pre - training already includes everything, then RL (reinforcement learning) is just a skill to stimulate these abilities.

Inspired by GPT - 3.5, Yao Shunyu developed ReAct (ReAct: Synergizing Reasoning and Acting in Language Models), which allows large language models to simultaneously "reason" and "act" when interacting with the external environment.

Yao Shunyu later commented that this was his favorite work (and also the work with the highest citation count so far). Based on this, his research gradually shifted to two cores: one is how to define valuable tasks and environments more relevant to the real world; the other is how to develop simple but general methods.

However, to achieve generalization and universality, one must learn to reason. And language models just provide a strong enough prior, which allows you to reason, and reasoning can generalize across different environments.

Therefore, based on the work of GPT and chain - of - thought, the Tree of Thoughts (ToT) enables language models to solve complex problems more efficiently than traditional left - to - right reasoning through multi - path exploration and self - evaluation.

Looking back at Yao Shunyu's academic and research journey, it is not difficult to find that whether it is his pursuit of general intelligent agents or his enthusiasm for language as a medium for cognition and decision - making, it is a continuous exploration of cross - boundary thinking and generalization ability.

The Bottom of the Ninth Inning at 28

Compared with MC HotDog at 23, Yao Shunyu, who graduated with a Ph.D. at 28 and joined OpenAI, has truly entered his "bottom of the ninth inning".

On August 1, 2024