OpenAI Expands Custom Model Program to Tailor Generative AI Solutions for Enterprises

OpenAI is ramping up its Custom Model program to assist enterprise clients in crafting bespoke generative AI models tailored for specific use cases, domains, and applications.

Introduced last year at OpenAI’s inaugural developer conference, DevDay, Custom Model provided companies with the opportunity to collaborate with dedicated OpenAI researchers to train and optimize models for specific domains. Since its launch, “dozens” of customers have enlisted in the program. However, OpenAI recognizes the imperative to enhance the program to further optimize performance, leading to the introduction of assisted fine-tuning and custom-trained models.

Assisted fine-tuning, a novel component of the Custom Model initiative, incorporates advanced techniques beyond fine-tuning, including additional hyperparameters and parameter-efficient fine-tuning methods at a larger scale. This empowers organizations to establish data training pipelines, evaluation systems, and other supportive infrastructure to enhance model performance for particular tasks.

Custom-trained models, another facet of the program, are developed in collaboration with OpenAI, utilizing the company’s base models and tools (e.g., GPT-4). These models cater to customers requiring deeper fine-tuning or infusion of new, domain-specific knowledge.

OpenAI cites examples of successful collaborations, such as SK Telecom, which enhanced GPT-4’s performance in “telecom-related conversations” in Korean, and Harvey, a recipient of support from OpenAI’s AI-focused venture arm, which created a custom model for case law incorporating extensive legal text and feedback from expert attorneys.

In a blog post, OpenAI emphasizes its belief that personalized models tailored to industry, business, or use case will become the norm for organizations. The expansion of techniques available for custom model development enables organizations of all sizes to derive more meaningful, specific impact from their AI implementations.

While OpenAI is experiencing remarkable growth, with reported annualized revenue nearing $2 billion, there is internal pressure to sustain momentum, especially with ambitious projects like the purported $100 billion data center co-developed with Microsoft on the horizon. The company anticipates that services like custom model training will continue to fuel revenue growth while it navigates future endeavors.

Moreover, fine-tuned and custom models offer potential benefits for OpenAI’s model serving infrastructure. Tailored models are often more efficient and performant than general-purpose counterparts, presenting a compelling solution for addressing compute capacity challenges amid escalating demand for generative AI solutions.

In addition to the expanded Custom Model program, OpenAI has unveiled new model fine-tuning features for developers working with GPT-3.5, including enhanced dashboards for comparing model quality and performance, support for third-party integrations (initially with Weights & Biases), and tooling enhancements. However, details on fine-tuning for GPT-4, launched in early access during DevDay, remain undisclosed.