Written by: Techub News Edited

Introduction

As GPT-5 is officially released and sparks global discussions, Matthew Berman, the host of a well-known technology channel on YouTube, conducts an in-depth interview with Mark Chen, the research director of OpenAI. As the head of OpenAI’s research team, Mark Chen has been deeply involved in the core development of the Codex to the GPT series of models. This conversation comes at a time when OpenAI has released its most powerful model to date, GPT-5, along with two high-performance open-source models. Mark Chen's insights unveil the thoughts of this leading AI laboratory on model architecture, data strategy, research philosophy, and reflections on future forms of intelligence.

Summary

• GPT-5 represents the first deep integration of the pre-training paradigm and reasoning paradigm, marking a fundamental expansion of model capabilities.
• Synthetic data plays a key role in GPT-5 training, aimed at specifically enhancing the model's performance in certain fields (such as programming), and its use is on the rise.
• OpenAI has a steadfast and long-term research roadmap, and progress from external sources (including outstanding results from Chinese laboratories) has not shaken its core research direction.
• The recently released open-source models by OpenAI aim to set a new benchmark for safety and performance for the community and have undergone rigorous risk assessment framework testing.
• The future development of AI will focus more on improving the model's reasoning efficiency, long-term memory capabilities, and exploring organizational forms of agent collaboration, with the ultimate goal of creating AI that can accelerate scientific research.

GPT-5: Integration of Pre-training and Reasoning Paradigms

Mark Chen regards GPT-5 as an important milestone in OpenAI’s research journey. He points out that if GPT-4 represented the extreme achievement of the pre-training paradigm, then GPT-5 is the first successful product to deeply integrate the pre-training paradigm with the reasoning paradigm. This means that GPT-5 not only inherits the broad knowledge brought by large-scale pre-training but also deeply integrates the deep thinking abilities similar to O series models.

This integration brings significant efficiency improvements. Mark Chen uses multi-modal tasks as an example: when the model processes complex visual inputs and needs to extract information from them, the “thinking time” required by GPT-5 has been reduced several times compared to previous models. This enhancement in reasoning efficiency is global, enabling the model to provide users with the needed deep analysis more rapidly. The goal is to allow users to not manually select “whether to activate reasoning mode”, as the model can automatically switch seamlessly between quick responses and deep thinking based on task complexity.

In programming capabilities, GPT-5's progress is particularly evident. Mark Chen reveals that internal evaluations show GPT-5 achieving a win rate of over 70% on code generation tasks compared to previous models (like GPT-4o and o1 series). The model not only generates prettier and more robust front-end code but also significantly increases the amount of code it can coherently output in a single conversation, easily surpassing a thousand lines and handling vast codebases in the real world. All this stems from the team's positioning of GPT-5 as a model “tailored for developers”, investing considerable effort in practical scenarios like long contexts and code reliability.

Data Strategy: The Rise and Challenges of Synthetic Data

In response to questions about whether “high-quality new data has run out”, Mark Chen admits that data acquisition is always a challenge, but not insurmountable. In addition to continuously seeking and licensing more public data sources, a key innovation of GPT-5 lies in the large-scale and strategic application of synthetic data.

The synthetic data project led by researcher Sebastian Bubbeck has “borne fruit” in the development of GPT-5. Mark Chen emphasizes that OpenAI believes synthetic data has the potential to provide higher quality content than human data and can substantially improve models in a way that goes beyond merely expanding the knowledge base. Synthetic data is not used to indiscriminately increase the amount of data across all fields but is selectively used in “specific areas where exceptional model performance is desired”. Programming is one of the key areas emphasized in the release of GPT-5.

In response to industry skepticism that “synthetic data generated from previous models can only provide marginal improvements to later models,” Mark Chen states that their research demonstrates that synthetic data can bring meaningful advancements. While he is unable to disclose the exact ratio of synthetic data to human data in GPT-5, he clearly states: “Over time, the proportion of synthetic data is becoming increasingly significant.” This indicates that the strategic importance of synthetic data in future model training will be greater.

Research Philosophy: A Steadfast Roadmap and “Vibe Check”

When asked whether he draws inspiration from the recent excellent open-source models and efficiency innovations from Chinese laboratories (such as DeepSearch), Mark Chen gave an unexpected answer: OpenAI's research roadmap has changed very little over the past few years, even in the face of significant external breakthroughs.

He explained that this roadmap outlines the path to AGI (Artificial General Intelligence), clarifying short-term, medium-term, and long-term goals. OpenAI’s advantage lies in its steadfast belief in the path it has chosen, focusing on executing the established plan rather than making “knee-jerk” adjustments based on external progress. For example, the development of reasoning models is a core part of the roadmap that has already been determined, pursued independently by OpenAI to achieve general intelligence. Of course, he also highly appreciates the excellent work of Chinese laboratories in architecture research and efficient kernel development, which he sees as worthy of learning.

When deciding when a model is “ready” for release, Mark Chen openly admits that it “has a bit of artistry.” The team must balance the pursuit of perfection with timely deployment. The model must pass a rigorous “vibe check”—meaning it feels good in actual use without concerning small flaws or anomalies. He shared his personal testing methods: including some cutting-edge mathematical problems (like designing an even random number generator for modulo 42), generating physical simulations or user interfaces to test code robustness and physical intuition, and serving as a creative writing “thought partner” to assess its style, persuasiveness, and understanding. He believes that humor remains a “gold standard” challenge for models at present.

Open-source, Safety, and Community Benchmarks

When discussing the recently released 20 billion and 120 billion parameter open-source models, Mark Chen appeared very excited. He believes that providing powerful models that can run on laptops or even smartphones will greatly lower the entry barriers for enthusiasts, scholars, and developers with special needs, promoting broader innovation.

More importantly, this open-source initiative aims to “set a new benchmark” for the industry. OpenAI conducted extensive safety assessments for this purpose. They used an internal “Preparedness Framework” to test the model from multiple risk dimensions, including biological, chemical, and cybersecurity. Assessment scenarios included: if malicious actors fine-tuned these models in an attempt to maximize their danger in aspects like cyber-attacks, where would their “risk limit” be? OpenAI ensures that the potential risks of these models before release are below the “high-risk” threshold set by the company. Mark Chen states that this is an important practice to establish norms for “what constitutes a safe, responsible open-source release.”

Future Outlook: Intelligent Agents, Memory, and Self-improving AI

Regarding the future forms of AI, Mark Chen responded to colleague Noam Brown’s views on the debate between “omnimodels” and various specialized smaller models. He believes that having a “super brain” capable of handling all tasks is attractive, as it can somehow internalize various sub-modules. However, OpenAI is also committed to building “organizational AI,” which is a framework where multiple AI agents collaborate to achieve advanced complex goals. This is similar to contemplating “which is more efficient: organization or individual,” and is an active area of research.

The substantial amount of “scaffolding” (i.e., complex prompt engineering and context management) present in current AI applications, in Mark Chen's view, compensates for the model's capability deficiencies. As models become more general and robust, the demand for complex scaffolding will decrease. One of the key breakthroughs needed to achieve this lies in “memory.” Models will need to handle extremely long contexts, accommodating the entire codebase, documents, and even daily visual information of users, enabling them to act more autonomously on behalf of users and reduce frequent manual intervention.

Looking forward to the next 6 months, Mark Chen is most excited about continuing to advance the “reasoning extension paradigm,” exploring more methods that leverage test-time computations, and innovating in reinforcement learning (RL) optimization and goal design. Over the next two years, he hopes to see models reaching performance levels in AI research that match his own, ultimately creating a “self-improving” system capable of driving AI innovation autonomously.

Finally, for developers and knowledge workers feeling anxious about the rapid advancements in AI coding abilities, Mark Chen offers an optimistic and pragmatic suggestion: embrace tools and accelerate oneself. By learning to collaborate with AI, transform yourself into a super individual that enhances efficiency by 2x or 3x while delving into the understanding of technological principles and contributing to them. He strongly believes that while AI automates certain superficial tasks, it will invariably create new layers that require human adaptation and value addition. He is confident that this technology will ultimately enhance the quality of life for the majority through driving scientific progress and economic prosperity.

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。