Yao Shunyu, who opened the second half of the big model, also opened the second half of personal AI.

Recently, the updates of this celebrity OpenAI Chinese researcher have started to receive hot attention.

Previously, some people said he appeared on Xiao Zha’s must find list, while others recently revealed that he is about to join another giant - a Chinese tech giant - and there have been shocking rumors about the “transfer fee”. And there is another theory that Yao Shunyu chose to start her own business

Where is Yao Shunyu going? don’t know yet.

But Yao Shunyu resigned from OpenAI and received confirmation from different channels, only a personal official announcement was needed.

He seems to be about to embark on the second half of his personal AI.

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 points in the college entrance examination and entered the Yao class at Tsinghua University with Anhui Tanhua. Finally, he obtained a Ph.D. in computer science (language and reinforcement learning) from Princeton and joined OpenAI after graduation

And even earlier, Yao Shunyu had outstanding and well-known scientific research achievements, such as:

+Tree of Thoughts: Allowing LLM to think repeatedly, significantly improving reasoning ability

+SWE bench: A large model capability evaluation dataset.

+SWE agent: an open-source AI programmer.

• ReAct…

There are even philosophical ideas beyond age, such as a blog post on “The Second Half of AI”, which is highly regarded both inside and outside the AI industry.

So, what kind of young person is Yao Shunyu?

##Yao Shunyu’s Growth Path

Shortly after joining OpenAI in 2024, Yao Shunyu recommended a book in an interview called “G ö del, Escher, and Bach: The Great Accomplishment of Collecting Strange Materials”.

This classic work, written by artificial intelligence pioneer Hou Shida, cleverly blends G ö del’s incompleteness theorem, Escher’s illusion painting, and Bach’s polyphonic canon, demonstrating how these seemingly unrelated elements echo each other in the universal recursive system of computers, leaving behind a meaningful sentence: ‘Find and reap.’.

As inspired by this book, the interdisciplinary perspective and open attitude towards complex information seem to run through Yao Shunyu’s entire academic career, gradually forming a personal style trait.

Like all top students, Yao Shunyu attended one of the best schools in Hefei, Hefei No.45 Middle School (2009-2012), and after graduating from junior high school, she was promoted to Hefei No.1 Middle School.

In 2014, he won a silver medal with 495 points in the National Olympiad in Informatics (NOI); The following year’s college entrance examination ranked third in Anhui Province with a score of 704 in science, and I entered the Yao class of Tsinghua Cross Information Research Institute, majoring in computer science.

Behind this seemingly “standard academic genius” start, there is actually a different and even slightly rebellious temperament.

######△ From Qing Xiaohua

Yao Shunyu revealed in an interview that compared to other students in the Tsinghua Yao class who focus on one thing and constantly dig deeper, he loves to read a lot of mathematics, history, and even various miscellaneous things.

And his love for hip-hop music is no longer news.

Eminem, Egg Burger, Hot Dog, Jcole and other rap artists accompanied him in middle and high school, and at Tsinghua University, he was also a co-founder of the Tsinghua University Student Rap Club.

It is worth mentioning that at the opening ceremony of various types of independent selection interviews at Tsinghua University in 2019, Yao Shunyu gave future students a reason that can be called “freestyle”, explaining why he chose Tsinghua University:

But for me, choosing Tsinghua or Peking University is not a problem, because Peking University does not have a class named after my surname.

In addition to serving as a co-founder of the rap club, Yao Shunyu also serves as a recruitment volunteer for Tsinghua University and the chairman of the Yao Class Joint Committee. When talking about the influence of Yao Ban on him, he mentioned:

Yao Ban attaches great importance to the study of theoretical foundation courses, such as courses related to operating systems or circuit design, which may seem unrelated to scientific research, but now it seems to be somewhat helpful. It can give you a basic understanding of the full picture of computer science.

######△ (Group photo of Yao Ban, Tsinghua enrollment official account)

The cross-border temperament of being compatible with various types of information, exploring different disciplines, and finding fun from them is particularly evident in Yao Shunyu. In the words of Isaiah Berlin, he was more like a ‘fox’ than a ‘hedgehog’, and this was reflected in his subsequent research.

(Note: Berlin’s “hedgehog” metaphorically refers to those who focus on a single core idea, applying all their experiences to this central viewpoint; while “fox” refers to those who have a wide range of interests, are good at flexibly dealing with different problems, and rely more on diverse strategies and perspectives.)

Surprisingly, before the second semester of her junior year, Yao Shunyu had neither been exposed to AI nor conducted any scientific research.

An opportunity for exchange led him to go to MIT, after which he began to conduct research in computer, vision, and cognitive science with Wu Jiajun.

Yao Shunyu stated that while studying under Wu Jiajun and his senior Jun Yan Zhu, he not only mastered the basic skills of experimental and reporting research, but was also deeply influenced in his thinking. He realized the intersection of psychology and artificial intelligence and learned to think about problems from a higher dimension and holistic framework.

After completing his four-year studies, Yao Ban officially went to Princeton in 2019 to pursue a PhD, demonstrating his cross disciplinary spirit once again.

During his undergraduate studies, he mainly focused on computer vision, while during his doctoral studies, he was originally admitted to the field of computer vision, but due to a shift in interest, he contacted a mentor in the field of natural language processing (NLP) on his own. In the end, he joined Karthik Narasimhan’s team by chance and began researching natural language processing and reinforcement learning.

This is equivalent to shifting from vision to language, but ‘finding it leads to gaining something’. Yao Shunyu recalled afterwards that this crossover was also his lucky spot, as it coincided with the rise of GPT-2 at that time. Therefore, in his first year of doctoral studies, he had already started thinking about how to turn the language model into an agent

And the focus on (generic) agents runs through his research.

Dr. Yao Shunyu’s first job in his career was called CALM (2020), researching how to use language models as agents to play language games.

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

Yao Shunyu said that although this work is not as famous as Swebench, ReAct, or Mind Tree, it is of great significance to him.

In Zhang Xiaojun’s conversation, we found that Yao Shunyu’s attention to language had already sprouted in this paper five years ago.

+Language is a tool invented by humans to achieve generalization, which is more fundamental than anything else.

In other words, playing games with language is infinite. Agents can use language to reason and combine, find suitable actions in different contexts, and thus generalize.

In other words, agents also need to have the ability to cross boundaries, and language is an excellent medium.

However, Yao Shunyu also realized that without a good task or environment, even playing the “game” with high scores is meaningless.

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

Similarly, later classic works * * SWE Bench * * and * * SWE agent * * were also aimed at verifying agent capabilities for a meaningful task (real-world programming).

And by 2022, the emergence of GPT-3.5 changed everything.

As is well known, the efficiency of repeatedly trying and making mistakes on a blank sheet of paper is extremely low, and in traditional reinforcement learning, such inefficient attempts are not uncommon: agents are usually either limited to a single task, such as Go, or blindly explore in a vast action space.

GPT-3.5 makes 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 fine tuned, it can become an agent with cross-border temperament and generalization ability.

As Yao Shunyu said, if your Pre Training already includes everything, then RL (reinforcement learning) is just a skill that inspires 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 job (and currently his most cited job). Based on this, his research gradually shifted towards two core areas: how to do valuable tasks and environments that are more relevant to the real world; The second is how to do some simple yet universal methods.

However, in order to achieve universality and generalization, one must learn to reason. And the language model just provides a strong enough prior, which allows you to reason, and reasoning can generalize between different environments.

Therefore, based on the work of GPT, thought chain, etc., the Tree of Thoughts (ToT) enables language models to achieve more efficient complex problem-solving abilities than traditional left to right reasoning through multi-path exploration and self-evaluation.

Looking back at Yao Shunyu’s studies and research, it is not difficult to find that both his pursuit of general agents and his passion for language as a cognitive and decision-making medium are continuous explorations of cross-border thinking and generalization ability.

##28 year old in the second half of the ninth inning

Compared to Hot Dog’s 23-year-old, Yao Shunyu, who graduated with a PhD and entered OpenAI at the age of 28, truly ushered in her “second half of the ninth inning”.

On August 1, 2024, Yao Shunyu announced her joining OpenAI, which was undoubtedly highly anticipated.

PS: When it comes to the connection between Yao Shunyu and OpenAI, a not too cold knowledge is that Yao Shunyu’s mentor Karthik Narasimhan is the second author of GPT-1.

In just over a year since joining OpenAI, Yao Shunyu’s most famous achievement was her blog post “The Second Half” published in April this year, which announced that the game with AI main threads has entered the second half. **

(Yes, AI has also ushered in the second half of its ninth game)

Yao Shunyu stated that the development of AI can be divided into two stages: upper and lower. In the first half, the focus was on models and methods, with the core building of models and training methods; As AI technology matures, the focus in the second half shifts to how to define tasks with practical significance and how to effectively evaluate the performance of AI systems.

The new evaluation should be close to real-world tasks, optimize model formulas to solve tasks, and form a new positive cycle.

It is not difficult to see the transformation of Yao Shunyu’s attitude here: from a research worker who focuses on solving problems, gradually growing into a “product manager” who is good at discovering needs.

According to a just revealed interview with Zhang Xiaojun, Yao Shunyu also talked about the underlying logic of entrepreneurship:

Before graduation, all I thought about was how to create an Einstein? I was a relatively academic person at that time - you were at Princeton, and your idols were von Neumann and Einstein - so naturally, the most interesting task that came to mind was, can I discover the next theory of relativity? This undoubtedly marks the achievement of AGI or ASI (Artificial Intelligence). Later, when I went to Silicon Valley and California, and joined the company, I found that human organization is also an interesting thing. **If we could create a new company, a one trillion dollar, agent-based company, it would be very interesting. **

When it comes to whether they will go out to start a business, Yao Shunyu candidly stated that most people at OpenAI would consider starting a business. He needs to do more challenging things, and starting a business is natural. Of course, we should still find something good. He likes to think things through before doing them.

Of course, leaving OpenAI doesn’t necessarily mean starting a business. For example, Karthik Narasimhan, the mentor of Yao Shunyu, chose to leave after staying at OpenAI for a year and went to Princeton to continue his ivory tower research.

Now, it has been a whole year since Yao Shunyu joined OpenAI.

There is a rumor that “Yao Shunyu joined Tencent with a salary of over 100 million yuan”, but it has been denied by Tencent’s official statement.

There is also a brand new startup company that has just been registered, and the legal representative is “Yao Shunyu”——

But it is still difficult to determine whether this “Yao Shunyu” is the one we are talking about, after all, there are more than one good name in the AI field called “Yao Shunyu”

In short, congratulations to Yao Shunyu for starting the second half.

Reference link [1] https://openreview.net/profile?id= ~Shunyu_Yao1&utm_source=chatgpt.com
[2] https://ysymyth.github.io/?utm _source=chatgpt.com
[3] https://x.com/shunyuyao12?lang=en&utm _source=chatgpt.com