Written by: Techub News Compilation

Introduction

On the day of the official release of GPT-4.5, OpenAI's Chief Research Scientist Mark Chen rarely accepted an interview with the Big Technology Podcast. This marks the first in-depth interview with OpenAI's senior executives in four years, coinciding with the launch of their latest flagship model. Mark Chen was recently promoted to this position months ago, and his remarks not only aim to introduce GPT-4.5 but also to elucidate OpenAI's key thoughts on model scaling, research paradigms, and future directions. In light of external eagerness for GPT-5 and widespread discussions about the "scaling wall," his insights provide a firsthand perspective from OpenAI's core research team.

Summary

• GPT-4.5 is the latest milestone under OpenAI's "predictable scaling paradigm," achieving a "quantum" leap in performance compared to its predecessor, comparable to the transition from GPT-3.5 to GPT-4.
• OpenAI is currently advancing the forefront of model capabilities along two parallel axes: "unsupervised learning" and "reasoning," with GPT-5 set to be the culmination of both pathways.
• The returns from model scaling have not diminished; GPT-4.5 demonstrates that performance can still be enhanced as expected by increasing computation, data, and optimizing algorithms. Additionally, model architecture optimization (like mixture of experts) and core capability development are two relatively independent dimensions.
• Beyond traditional benchmarking, GPT-4.5 shows higher "emotional intelligence" (EQ) and creative writing abilities, representing a new dimension of model capability exploration.
• OpenAI's strategy involves concurrently "advancing the frontier of intelligence" and "reducing the cost of capability acquisition," creating flagship models while offering cost-effective smaller models to form a complete product lineup.

GPT-4.5: The Latest Proof of Predictable Scaling

Mark Chen clearly positions GPT-4.5 as the latest achievement under OpenAI's "predictable scaling paradigm." This paradigm began with GPT-3 and has evolved through GPT-3.5 and GPT-4, continuing with GPT-4.5. He explains that based on the accumulated experience from previous model training, the team can accurately predict how model performance will improve when computation resources, data scale, and algorithm efficiency increase by an order of magnitude. GPT-4.5 is the next point on this predictive curve, with performance improvements similar in scale to the leap from GPT-3.5 to GPT-4.

Regarding why GPT-4.5 is being released instead of the highly anticipated GPT-5, Mark Chen states that the naming always aims to reflect the actual leap in model capabilities. The current enhancement in model abilities aligns with the "4.5" designation. He acknowledges that the interval between GPT-4 and GPT-4.5 seems longer, but attributes this, in part, to OpenAI investing considerable effort in another crucial track: the development of the "reasoning" paradigm. He emphasizes that the company’s research is exploratory and will progress simultaneously in all promising directions.

Two Expansion Axes: Unsurpassed Learning and Reasoning in Parallel

Mark Chen elaborates on OpenAI's concurrent two expansion axes. The first is the traditional "unsupervised learning" expansion, which enhances model capabilities through increased computation, data, and improved algorithms, with GPT-4.5 being the latest achievement along this route. The second is "reasoning" expansion, which aims to enable deeper thinking, planning, and self-correction capabilities in models, exemplified by OpenAI's o1 series models.

He clarifies that these two paradigms are not opposing but rather complementary. "Unsupervised learning" provides a broad foundation of world knowledge for models, while "reasoning" abilities need to build upon this knowledge. A knowledge-deprived model cannot learn deep reasoning out of thin air. Additionally, there is a feedback loop between the two paradigms, where advancements in reasoning capabilities may require or promote further expansion of foundational knowledge.

In application scenarios, the two types of models also exhibit different characteristics. Large models like GPT-4.5 respond quickly and are suitable for knowledge-based work and creative writing that require instant feedback; reasoning models like o1 may need longer "thinking" time and are better suited for solving complex, multi-step problems. Mark Chen particularly notes that in scenarios like creative writing, GPT-4.5 has already surpassed the current reasoning models.

Addressing Concerns of the "Scaling Wall": Returns Unabated, Optimization Ongoing

In response to industry concerns about the potential diminishing returns of scaling large models or hitting a "scaling wall," Mark Chen provided a clear denial. He stated that during the development of GPT-4.5, the performance returns from increasing computation and data resources have been consistent with previous model experiences and have not diminished. OpenAI's scaling mechanism remains effective and predictable.

Regarding earlier reports that training for GPT-4.5 faced multiple interruptions and challenging problems, Mark Chen responded that the "training-diagnosis-intervention-restart" model is standard in the development process of all large foundational models and is not unique to GPT-4.5. He emphasized that viewing these behemoths as large-scale experiments, with monitoring and adjustments throughout, is key to ensuring ultimate success.

When the conversation shifted to peers like DeepSeek significantly enhancing efficiency through model architecture optimizations (like mixture of experts models), Mark Chen differentiated between the "core model capability development" and "reasoning service optimization." He believes making models run more efficiently and cost-effectively (reasoning optimization) is independent of enhancing their core intelligence. OpenAI is continuously investing in both areas and has successfully reduced the reasoning costs of the GPT-4 series models by several orders of magnitude. He confirmed that architecture optimization techniques like mixture of experts are universal technologies applicable to both GPT-type foundational models and reasoning models, and OpenAI has conducted numerous explorations in this area.

Large Models vs. Small Models: Balancing Frontier Advancement and Inclusive Delivery

In response to the community view that "niche, specialized small models are the future," Mark Chen articulated OpenAI's balancing philosophy. On one hand, the company's fundamental mission is to "advance the frontier of intelligence," which requires continuous development of flagship models like GPT-4.5. He strongly believes that pushing model capabilities from "99.9 percentile" to "world-class" levels is fundamentally significant, as it may unlock entirely new, transformative applications, such as truly powerful AI agents.

On the other hand, OpenAI is also committed to making advanced capabilities more accessible. This is reflected in providing cost-effective "small" models like GPT-4o-mini, enabling more users to access near-frontier capabilities at an affordable price. Therefore, OpenAI's strategy is to create a complete suite of products from flagship models to lightweight models, rather than an either-or choice. Mark Chen believes that niche models will not disappear, but OpenAI will simultaneously build the strongest foundational models and strive to deliver these top-tier capabilities to everyone at lower costs over time.

Beyond Traditional Benchmarks: A New Dimension of Emotional Intelligence and Capability Exploration

Besides meeting expectations for improvements in traditional benchmarks such as mathematics, coding, and knowledge Q&A, Mark Chen highlighted GPT-4.5's significant advances in "emotional intelligence" (EQ) and creativity. He illustrated that when users share dilemmas with the model, GPT-4.5 can provide more concise, empathetic, and human-like responses instead of mechanically listing a long series of suggestions. In tasks like creative writing and generating ASCII art, the new model also shows more nuanced and less error-prone capabilities.

He anticipates that some may criticize this as "shifting the focus of evaluation," but he firmly denies this assertion. Mark Chen reiterated that GPT-4.5 has fully met the expected expansion law in traditional hard metrics. He emphasized that each new model release is a process of "discovering use cases" in collaboration with the community. GPT-4 is already highly intelligent, and GPT-4.5 reveals new possibilities in more subtle areas like emotional understanding and creative expression. OpenAI shares these early findings to invite users to jointly explore the boundaries of the model's new capabilities.

Team, Products, and Future: Agents as a Natural Extension of Model Abilities

In the closing of the interview, Mark Chen briefly discussed the state of OpenAI's team. He acknowledged the rapid changes in the AI field, stating that personnel movement is a natural phenomenon in the healthy development of the industry, providing opportunities for outstanding internal talents to shine. He firmly believes that OpenAI still possesses a world-class talent pool, significantly differing from the standards of other companies in the industry.

Regarding the classic debate of "better models" versus "better products," Mark Chen's stance is clear: more powerful models are the fundamental premise for higher-level products. He used the Deep Research functionality as an example, pointing out that this capability to conduct in-depth research, synthesize information, and generate complete reports is only achievable when the model has a sufficient level of reasoning and knowledge base. The advancement of model capabilities will naturally give rise to new product forms and application paradigms, and AI agents are the forefront embodiment of the current combination of model abilities and products. In the future, as models continue to advance along the reasoning and unsupervised learning axes, the capabilities and reliability of AI agents will undergo a qualitative leap.

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。