OpenAI Chief Scientist Jakub Pachocki and Chief Research Officer Mark Chen discuss the transition from "atmosphere encoding" to "atmosphere research."

Author: Techub News Compilation

Introduction

In September 2025, OpenAI's Chief Scientist Jakub Pachocki and Chief Researcher Mark Chen rare jointly accepted an in-depth interview with renowned venture capital firm a16z. As two core leaders of the OpenAI research team, they oversee one of the most high-profile and cutting-edge research teams in today's AI field. This dialogue delved into the GPT-5 released in September 2025, the underlying concept of "reasoning," the future directions of AI research, and how to build a research culture and organization capable of continuously generating disruptive innovations.

This is not only an interpretation of OpenAI's latest technological achievements but also a rare glimpse into the internal operational model, future vision, and technological philosophy of this world-leading AI company. In today's fiercely competitive AI technology landscape, where product iterations are accelerating, the thoughts of these two tech leaders provide crucial clues for understanding the next phase of AI development.

Summary

The core goal of GPT-5 is to make "reasoning" the default behavior of AI, aiming to bridge the gap between rapid-response models and deep-thinking models.
OpenAI's ultimate research goal is to create "automated researchers" that allow AI to autonomously discover new knowledge, particularly in the fields of mathematics, programming, and fundamental sciences.
The potential of reinforcement learning (RL) has not yet been fully explored, and its combination with pre-trained models is key to achieving continuous breakthroughs, evolving towards a learning method that is closer to human learning.
The key to building a successful research culture lies in "protecting basic research," giving teams the space to think about long-term, fundamental questions, and avoiding falling into the mindset of short-term product competition.
The vision from "atmosphere coding" to "atmosphere research": when AI tools become powerful enough, creation (coding, research) will focus more on intuition, taste, and high-level ideas instead of mechanical execution.

GPT-5: Bringing Reasoning Mainstream

The interview began with the GPT-5 released in September 2025. Jakub Pachocki explained that GPT-5 represents OpenAI's attempt to mainstream "reasoning" abilities. Prior to this, OpenAI's models were roughly divided into two series: the "rapid response" models represented by GPT-2/3/4, and the "deep thinking" models represented by the O series (like o3). The former excels at providing immediate answers, while the latter requires long periods of thought to deliver the best results.

"Tactically, we do not want users to be confused about which mode to use," Pachocki said. Therefore, one of the research focuses of GPT-5 is to identify and determine the "optimal amount of thinking" required for different prompts, removing this choice burden from the users. They believe that the future direction is to revolve more around reasoning and agents, and GPT-5 is a significant step towards providing reasoning capabilities and smarter behaviors by default.

Mark Chen added that GPT-5 has been improved in many areas compared to o3, but the core focus of this release is undoubtedly to make the reasoning mode accessible to more users.

Evaluation Criteria: From Saturation Metrics to Discovering New Knowledge

When asked how to assess model progress, Pachocki pointed out that many evaluation metrics (Evals) used in recent years have approached saturation; for instance, improving accuracy from 96% to 98% is no longer significant. More importantly, the paradigm of AI research has changed.

During the era of GPT-2/3/4, the basic paradigm was to pre-train on massive datasets and then use evaluation metrics as a measure of the model's generalization ability. Now, techniques such as reinforcement learning allow for in-depth training in specific domains (like serious reasoning), enabling models to become experts in those areas. Although this could achieve excellent performance in specific evaluations, it does not necessarily indicate the same level of generalization capability.

Therefore, OpenAI is currently more focused on evaluations that reflect the model's ability to "discover new things." Chen mentioned that one of the most exciting advancements this year is the model's performance in math and programming competitions (like AtCoder). However, these competitions are themselves being "conquered" by the models.

"The next set of evaluations and milestones we are preparing will involve achieving real progress and discoveries in economically relevant fields." Pachocki emphasized. This means the evaluation criteria will shift from solving known problems to creating new and valuable knowledge in open environments.

Automated Research: Ultimate Goals and Long-term Reasoning

When asked about the research roadmap for the next 1-5 years, Pachocki clearly stated: "The core goal of our research efforts is to create an automated researcher, that is, automation in the discovery of new ideas." Within this, automating machine learning research itself is a specific direction, but it may seem overly self-referential. Therefore, they are also committed to advancing the automation process in other scientific fields.

A good way to measure this progress is to observe the "time span" in which the model can perform effective reasoning and make progress. Currently, the model is already able to perform coherent reasoning for about 1 to 5 hours on problems of similar difficulty to high school competitions. OpenAI's research focus is to extend this time span, including the model's ability to make long-term plans and maintain memory.

This also echoes the evaluation question: "How long can the model run autonomously?" Evaluations of this kind are especially important to them.

Chen further elaborated on the core role of reasoning in long-term operations. Just like humans solving mathematical problems, it requires trying different methods, learning from mistakes, and iterating repeatedly. The ability to maintain robustness over a long time is precisely what deep reasoning grants to agents.

Reinforcement Learning: The Source of Continuous Breakthroughs

Since the release of o1, reinforcement learning (RL) has become a powerful tool for OpenAI to achieve continuous breakthroughs. Although there are often predictions that RL's benefits will plateau or face issues like model collapse, OpenAI's model performance has consistently shattered such expectations.

Pachocki explained why RL is so effective. He believes RL is a very flexible approach. OpenAI recognized the strength of RL early in deep learning but faced long-standing challenges in providing a suitable "environment" for RL models to interact with the real world or simulated environments. The emergence of language models solved this "anchor point" problem.

"When you pre-train in the extremely rich and robust environment of natural language, you gain the ability to pursue different goals and ideas based on that." Combining the general learning ability of deep learning with the goal-directed training of RL has resulted in a tremendous chemical reaction. Pachocki calls this the most exciting phase in OpenAI's research over the past few years, where they have discovered many new directions and promising ideas, all of which continue to yield results.

For businesses or researchers looking to leverage RL, Chen's advice is: "The most important thing is not to think the status quo will last forever." Just like two years ago when everyone was focused on how to build fine-tuning datasets, he believes that reward modeling and other methods will also evolve rapidly, ultimately developing into simpler and more human-like learning approaches.

From "Atmosphere Coding" to "Atmosphere Research"

As former competitive programmers, Pachocki and Chen are deeply impressed by AI's progress in the programming field. Chen shared a story: last weekend he talked with some high school students, and they stated, "The default coding method now is 'atmosphere coding.' For them, hand-coding everything from scratch has become a strange concept. Why not use AI to assist?

This inspired Chen: "I hope the future will be 'atmosphere research.' When AI tools are powerful enough, the research process will also focus more on intuition, taste, and high-level concepts, leaving tedious execution and verification to AI.

So, what makes a great researcher? Pachocki believes that "perseverance" is key. Research is about exploring the unknown, and most attempts are likely to fail. Therefore, researchers need to have a mindset of "being ready to fail and learn from it," while maintaining absolute honesty about their hypotheses. They need to have confidence in their ideas and persist, while also being able to objectively assess progress and adjust in a timely manner.

Chen added that experience is crucial, helping you judge whether the difficulty of problems is appropriate and manage your emotions over long-term research. Cultivating a nose for "interesting problems" through discussions with colleagues and reading excellent papers is also part of the research process.

Building and Protecting a Top-tier Research Culture

As leaders of the OpenAI research team, how do Pachocki and Chen attract and retain top talent? Pachocki pointed out that the fundamental motivation is OpenAI's commitment to basic research. They are not keen on focusing on what models competitors have released but rather on frontier innovations. "People are inspired by this mission." In addition, building a good culture, cultivating talent pipelines, and having a deep talent reserve ("deep bench") are also crucial.

In recruitment, they are not just looking for the most active people on social media, but rather value the ability to "solve difficult problems in any field." Many of the most successful researchers had previously worked in other fields such as physics, computer science, or finance, possessing a solid technical foundation and a willingness to tackle ambitious problems.

"Protecting basic research" is key to creating a winning culture. Chen explained that it is essential to ensure researchers have the space to think about what truly matters in the next year or two, rather than being distracted by short-term product demands or falling into a competition of releases with other labs. "We need to ensure that people have that comfort and space to think: What will things look like in a year or two?"

Balancing research with products is another challenge. Their approach is to clearly distinguish researchers who truly care about the product and are responsible for its success, allowing them to work closely with product teams. At the same time, the company's leadership fully recognizes and supports the long-term vision of research, understanding that current products are not the endpoint but rather part of a collaborative vision for the future with R&D.

Resources, Computation, and Constant Challenges

In resource allocation, managing computational resources among different projects is an important task for the two leaders. They candidly admitted that historically, more computational resources have flowed into core algorithm advancement rather than productized research, but this allocation is dynamic and flexible.

When asked where an additional 10% of resources would be allocated, Pachocki replied: "Computation." He believes the statement that "we will soon enter a data-constrained phase" is inaccurate; computational resources will remain a decisive factor in the foreseeable future. "Anyone who says that just needs to work in my position for a week, and they'll realize that no one ever says, 'I have all the computational resources I need.'" Chen agreed with a smile.

Finally, when asked what principles should remain unchanged in the rapidly evolving wave of AI, Pachocki mentioned broader physical constraints, such as energy, as well as new constraints brought by future robotics technologies. However, at the forefront of intelligence, "I won't make too many assumptions." Maintaining an open and learning mindset is crucial. Chen shared the "secret" to keeping the team running at high speed: OpenAI's research culture has made him never feel stagnant in learning, with new breakthroughs and achievements constantly emerging, requiring full effort to keep up. This sense of ongoing challenge and growth is a driving force.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。