GPT-4.5: A New Milestone in the Scaling Paradigm, Says OpenAIs Chief Research Scientist

OpenAI has unveiled its largest language model to date, GPT-4.5. According to Marc Chen, OpenAI’s Chief Research Scientist, this model demonstrates that the scaling of AI models has not yet reached its limits.

On Thursday, OpenAI introduced its latest language model, GPT-4.5, boasting it as their most extensive and powerful chat model thus far. The company plans to first deploy it to Pro users, followed by Plus, Enterprise, Team, and Edu users in the coming weeks.

For Chen, GPT-4.5 addresses concerns from critics who doubt that research labs can continue to innovate and develop larger models.

“GPT-4.5 is indeed proof that we can continue the scaling paradigm,” Chen explained in an interview. “This marks a significant step up in capability.”

When asked why the new model wasn’t labeled GPT-5, Chen clarified that OpenAI aims to adhere to recognizable naming conventions. With a predictable scaling approach—such as moving from GPT-3 to GPT-3.5—they can anticipate the advancements that will arise from significantly greater computational power and efficiency. The new model aligns with what one would expect from GPT-4.5.

Chen emphasized that OpenAI can now scale along two distinct axes. “GPT-4.5 is our latest experiment in scaling through unsupervised learning, but there is also reasoning involved,” he stated.

He attributed the longer development time between GPT-4 and GPT-4.5 to the company’s strong focus on enhancing reasoning paradigms.

These two approaches can complement each other: “You need knowledge to build reasoning on top of it. A model can’t blindly learn reasoning from scratch,” Chen explained. The two paradigms reinforce one another, creating feedback loops.

Chen pointed out that GPT-4.5 is “intelligent” in a different way compared to reasoning models. It has a substantially broader knowledge base about the world. Compared to GPT-4, users preferred the new model for everyday applications by 60%, and for more complex cognitive tasks, this preference rose to nearly 70%.

When it comes to potential limitations on scaling, Chen was clear: “We are seeing the same outcomes. GPT-4.5 is the next step in this paradigm of unsupervised learning.” He clarified that OpenAI is very rigorous in its methodology, making predictions based on all previously trained models to gauge expected performance.

In addition to traditional benchmarks, where GPT-4.5 shows similar improvements as the transition from GPT-3.5 to GPT-4, Chen noted that the model possesses new capabilities. He cited its ability to generate ASCII art—something that previous models typically struggled with.

Chen also dismissed claims that developing GPT-4.5 was particularly challenging. “The development of all our foundational models is experimental. This often entails pausing at various stages to analyze what’s happening and restarting runs,” he explained. This was not particularly characteristic of the GPT-4.5 development process, but OpenAI has also employed this strategy for GPT-4 and the O series.

However, it is noteworthy that a significantly smaller model, Claude 3.7 Sonnet, outperforms GPT-4.5 from Anthropic in several areas, which seems relatively outdated since its data was collected only in October 2023. One likely reason for this is advancements in synthetic data since 2023.

[Source](https://the-decoder.com/gpt-4-5-is-proof-that-we-can-continue-the-scaling-paradigm-says-openais-chief-research-officer/)