Weng Jiayi's technical path, from a young age, was characterized by foresight and practical implementation. He possessed an extraordinary intuition for mathematics, enabling him to quickly find solutions. However, he perceived his own speed in understanding new knowledge as slower than average, requiring two to three times the effort to fully grasp. This self-awareness fostered a strategic approach to early learning: actively seeking out high school courses from his teachers in middle school, completing most of his high school mathematics by the second year, and delving into calculus in the third year, which he called "investing in his future." In China, for students outside Beijing, directly gaining admission to Tsinghua or Peking University was an almost insurmountable feat. To find a shortcut for college entrance, he turned to informatics competitions. Previously, he had achieved some success in mathematics competitions, but due to limited school resources and insufficient preparation, he decisively switched to informatics. Subsequently, he broke through the competitive landscape of the Fujian provincial team. In his second year of high school, he secured the highest score in the competition with a problem on minimum bipartite covering, earning him a sixty-point reduction in his Tsinghua admission score. This breakthrough was, in fact, a narrow victory, as the Fujian provincial team that year only secured a single bronze medal in that competition, awarded to Weng Jiayi himself. The competition experience not only served as a stepping stone for his education but also shaped his engineering mindset. During his third year of high school preparation, he would even code directly on an iPad without a compiler. This experience of training without an environment provided him with a more complete understanding of program structures and honed his ability to quickly locate bugs. He was particularly engrossed in constant optimization within algorithmic competitions; these engineering challenges of optimizing time and code length brought him joy and a sense of pure technical delight. This dedication to doing things thoroughly laid the groundwork for his later deep dive into infrastructure.

Upon entering Tsinghua University, Weng Jiayi became a campus legend, partly due to his breaking of an unwritten rule: open-sourcing all his assignments. At Tsinghua, information disparity was often treated as a survival resource, with past assignments and course materials largely held by a select few. Weng Jiayi, however, believed everyone should have equal access to this information. Consequently, he collected old assignments and course materials passed down from previous cohorts and uploaded them to GitHub after confirming there were no copyright issues. This act made him a household name in Tsinghua's Computer Science department, with classmates jokingly remarking that while juniors might not know who donated the building, they certainly knew Weng Jiayi, as they all relied on his assignments to survive. This influence was considered more potent than donating a building. Weng Jiayi's research explorations at Tsinghua were also marked by pragmatism and trade-offs. In his second year, he inadvertently joined Tsinghua's Computer Science department's academic innovation program. Initially, he had a vague interest in graphics, randomly selecting reinforcement learning. He discovered this field was essentially about playing games, training agents through reinforcement learning to complete tasks in video games, which was far from his initial expectations. However, this experience exposed him to the core logic of reinforcement learning for the first time. Concurrently, he harbored strong interests in computer graphics and cybersecurity. During his university years, he collaborated with a senior student to discover and fix vulnerabilities in Tsinghua's campus network, allowing for the free download of transcripts that previously cost ten yuan. In his computer graphics course, he invested significant time and effort to complete a demanding 16K high-definition image rendering assignment, almost no one else dared to attempt, and even invented a new algorithm for accelerating iterative convergence, ultimately earning one of the only two A+ grades in the class. However, one cannot pursue two paths on the research journey simultaneously. Weng Jiayi ultimately made a trade-off, abandoning his beloved computer graphics to focus on reinforcement learning. This decision was not impulsive but stemmed from his pursuit of technological implementation. Reinforcement learning had broader application scenarios and better allowed him to leverage his engineering strengths to create genuinely valuable outcomes. After confirming his focus on reinforcement learning, Weng Jiayi embarked on his first research project: training a neural network in a 1990s game to have an agent complete tasks like monster hunting, collecting health packs, avoiding obstacles, and clearing levels within a fixed map. Ultimately, he won this project using reinforcement learning methods. However, he candidly admitted he did not enjoy the entire process. In his view, the project's environment was too singular, and the training process was akin to "alchemy," requiring constant use of various tricks to prevent training crashes. Even if it didn't crash, he didn't know how to tune parameters to maintain performance, relying entirely on heuristic methods—a difficulty ten to a hundred times greater than tuning parameters in computer vision. This research model, relying more on luck than skill, made the pragmatic Weng Jiayi feel uncomfortable and led him to ponder whether the problems in reinforcement learning lay not in the algorithms themselves but in the lack of stable, user-friendly experimental tools.

Upon returning from his internship, Weng Jiayi entered the application season for graduate studies abroad. However, this period was not smooth sailing. Watching his peers enter prestigious university labs like Carnegie Mellon University and Stanford, publishing papers in top international conferences, while he had nothing, the immense gap led him into a brief period of despondency. He quickly adjusted, however, because he firmly believed that GPA was not the sole evaluation metric and that one must create their own evaluation system. He cited his Tsinghua advisor's criteria for the Computer Science department: three major indicators—papers, competitions, and GitHub repositories with over a hundred stars. Throughout his undergraduate studies, Weng Jiayi aimed to maintain just enough academic performance with minimal effort. Since his GPA was calculated before the end of the semester, he was satisfied with a B+, unwilling to waste any more time. This attitude of not blindly adhering to a single evaluation system allowed him to dedicate more energy to projects he was genuinely interested in. During the period from late 2019 to early 2020, overlapping with the pandemic and the application season, he launched two highly successful projects, which he termed "charity projects" because they were entirely non-profit yet genuinely solved people's needs. The first project, "Tian Shou" (天授), originated from his aversion to "alchemy-style" research. In late 2019, he discovered that RLlib, the most mainstream reinforcement learning framework at the time, had a severe problem: its extreme abstraction and hundreds of thousands of lines of code made it almost impossible to modify, failing to meet the practical needs of researchers. Consequently, during the 2020 Chinese New Year holiday, he began writing his own reinforcement learning experimental platform, completing the first version in just two weeks. Unlike RLlib's decay, Tian Shou pursued extreme consistency from its design inception, with its core advantages being simplicity, ease of modification, and stability—precisely hitting the pain points of research users. Tian Shou's success validated Weng Jiayi's judgment that the value of technology lies not in complexity but in its ability to solve practical problems. The second project, "Tui Xue Wang" (退学网), arose from his personal need. During the pandemic, he needed to check the status of his visa appointment but lacked suitable tools, so he manually coded a lightweight web scraper. Believing many others might have similar needs, he open-sourced it. The project gained rapid traction upon release, initially garnering over a million clicks and accumulating tens of millions to date. Although the scraper later became ineffective due to updates in the US consulate's system and Weng Jiayi did not maintain it, this short-lived project fulfilled its mission perfectly, helping countless individuals solve their visa inquiry challenges. These two projects brought Weng Jiayi unprecedented satisfaction. When asked when this pursuit of impact began to emerge, he recalled a flash of inspiration in his third year of high school: if life were a game, the final score would be measured by how many people remembered your name after you're gone. In his view, a person's value is not in the honors they receive but in the meaningful contributions they make to others and society. One cannot be good to everyone, but one can try to do good to those around them within their capabilities and contribute to things that are meaningful to everyone.

In 2020, Weng Jiayi received a Master's offer from Carnegie Mellon University. Due to the pandemic, he attended classes remotely from home. This period became a critical juncture for his career choices. He began applying for jobs, initially submitting applications to eighteen companies and receiving offers only from Google and AutoML. However, Weng Jiayi declined Google's offer because he did not want to be a cog in a large corporation doing work he wasn't particularly passionate about. Subsequently, he continued submitting his resume and received more high-profile offers, including one from Huifang Quant (幻方量化), which was establishing an AI lab at the time and later developed into DeepSeek. He also received offers from Nvidia, TikTok, and Meta's AI research. The offer from Huifang Quant once made him waver; if there were no better options, he might have joined them to focus on building reinforcement learning infrastructure. However, he ultimately chose OpenAI, which was not yet widely recognized and had not yet launched ChatGPT. This decision was not driven by salary or fame but by his pursuit of top-tier industrial-grade research. In his opinion, OpenAI and DeepMind at the time were the two best labs globally for reinforcement learning, and he wanted to experience how world-class research was conducted, rather than staying in a small, workshop-style research environment. What he truly wanted to learn was the methodology of industrial-grade research: how to quickly implement ideas, iterate efficiently, and build stable, reliable systems. Weng Jiayi eventually joined OpenAI's reinforcement learning group, personally interviewed and recruited by co-founder John Schulman. He still vividly recalls the interview; John Schulman gave the same engineering problem to only two people: him and Andrej Karpathy, a key member of the code-indexing project. It was an end-to-end open-ended problem. John Schulman gave him three hours, and Weng Jiayi completed the code in just two hours, even fixing a bug on the spot. He surmised that John Schulman valued his engineering capability, as Schulman had commented on his "beautiful GitHub profile," indicating an endorsement of his evaluation system. This recognition further solidified Weng Jiayi's career direction: to deeply cultivate the engineering field and be a "shovel seller." When asked if he had considered pursuing a Ph.D., Weng Jiayi stated directly that he had never seriously considered it. In his view, if the goal is to enter the industry, pursuing a Ph.D. is a waste of life. One can use a Master's degree as a stepping stone to meet the criteria for a Ph.D. entry, allowing companies to choose a Master's candidate instead of a Ph.D. The key is to clearly understand one's differentiated advantages. He further explained that a Ph.D. training core focuses on storytelling, paper writing, and diagramming, whereas the industry requires rapid iteration capabilities and system correctness. He quoted an OpenAI colleague, a former Ph.D. in reinforcement learning who later developed a well-known RL framework, saying: "Teaching a researcher how to do engineering is much harder than teaching an engineer how to do research." Weng Jiayi's core viewpoint is that the value of research lies in validation, and the key to validation is infrastructure. As long as the foundational system is correct, hyperparameters are reasonable, and iteration efficiency is high, effective ideas can be quickly screened. In the AI industry, every company's infrastructure has bugs to varying degrees; whoever fixes more bugs achieves better model performance. He is no longer interested in parameter-tuning research, preferring to lay solid foundations in infrastructure, allowing others to experiment and publish papers, perhaps even with his name attached. This is precisely why he calls himself the most customer-facing among shovel sellers, as the reinforcement learning module sits at the very top of the infrastructure stack, directly serving the core needs of model training.

Regarding the definition of AGI (Artificial General Intelligence), he revealed that there are various interpretations within OpenAI; his personal definition is that AI can complete 80%-90% of tasks he deems meaningful. Current technology is still a significant distance from this goal. He frankly admitted that the code he is responsible for daily cannot yet be safely handed over to models for modification, as the datasets for AI infrastructure are extremely low in coverage and prohibitively expensive, thus not yet touching this domain. He believes AI development involves the gradual replacement of repetitive labor, not a sudden industry-wide disruption. The core value of engineers—the ability to solve complex problems, innovative thinking, and engineering experience—is difficult for current AI to replace.

Regarding external criticism that OpenAI is "not open," Weng Jiayi expressed understanding of the trade-off between open-sourcing and company survival. He explained that resource-intensive model development requires securing capital infusion and commercial sustainability. OpenAI's "Open" strategy does not mean complete open-sourcing but rather making powerful tools accessible to the public through low-barrier entry, such as the free version of ChatGPT. He believes directly open-sourcing model weights might lead to misuse, and productized experiences are more beneficial for humanity. He candidly stated that while theoretically possible to achieve open-sourcing and community feedback, in reality, once open-sourced, competitors using closed-source models would exert pressure, affecting fundraising and experimentation.

The biggest challenge for OpenAI in achieving AGI is execution. He emphasized that stable organizational execution is paramount. Recalling the ouster and return of Sam Altman in late 2023, he pointed out that individuals with purely technical backgrounds may not be able to lead an AI company; people who can secure funding, computing power, and resources are needed. Sam Altman has become OpenAI's spiritual symbol, embodying the balance of technology, resources, and management. He believes that in a healthy organization, everyone is replaceable, provided the organization has the capacity for self-renewal and continuous cultivation of new talent. OpenAI's advantage lies in its infrastructure iteration speed being "not too bad."

He envisions an AI agent with infinite context memory as a potential future CEO, capable of overcoming limitations in human organizational management, such as inconsistent context sharing. Finally, when asked about AI solving world problems, he chose "predicting the future" but believes this would lead to the collapse of all value systems, rendering human effort, choices, and emotions meaningless. He proposed that time might not flow linearly, and one's future self might assist their past self in making certain decisions. He views AI development with caution, emphasizing that humans need to find ways to coexist with technology while preserving their own value and meaning.

Regarding entrepreneurship, he tends to focus on genuine needs rather than the technology itself. He believes OpenAI remains a worthwhile place to be, boasting top technical teams, ample resources, and the fastest iteration speeds, allowing him to continuously leverage his strengths and build more influential infrastructure. His goal is to maximize his mentions in OpenAI's blog posts, not for fame, but to increase the visibility of his work and inspire the industry. His advice to young people is: continue investing in engineering construction rather than pure academic research. He believes modern academia is often stuck in overfitting, researching repetitive tasks without translating to practical problem-solving. AI labs critically need infrastructure talent because it's a bottomless pit requiring continuous investment and solid capabilities. In closing, he left a thought-provoking statement: "I thought I figured out what I wanted, but I haven't quite figured it out. This question is worth a lifetime of contemplation."