Dive Brief:
- OpenAI API customers can now fine-tune GPT-3.5 Turbo to customize models for specific use cases at scale, the company announced Tuesday. Customers will be able to fine-tune GPT-4 models this fall.
- Fine-tuning allows customers to improve steerability, reliable output formatting and tone, OpenAI said. Early users were able to shorten their prompts — and in turn costs — while keeping performance up, the company said.
- Fine-tuned data will pass through OpenAI’s Moderation API and a GPT-4 powered moderation system to detect training data that conflicts with safety standards and to preserve the default model’s safety features, the company said in the blog post. Data sent in and out of the API belongs to the customer and OpenAI or any other organization will not use it to train models, the company said.
Dive Insight:
Most organizations have moved past the initial hype of OpenAI’s technology and are looking for business-specific generative AI and LLM use cases to bring value to the enterprise.
Fine-tuning is critical for organizations looking to advance beyond a generalized chatbot. In 2021, OpenAI began allowing customers to fine-tune the GPT-3 model, and early wins indicated the process could boost model accuracy.
Large enterprises, such as PwC, JLL and EY, have spoken with CIO Dive about the value customized models can bring, even in a conversational assistant function. Fine-tuned models can be trained to be more knowledgeable on a specific subject, whether that be commercial real estate, laws and regulation or something else.
CIOs and their organizations are most interested in adopting industry-specific generative AI applications, according to Info-Tech Research Group data.
The biggest drawbacks of adopting OpenAI’s technology for enterprises is the whirlpool of risks, emerging regulations and privacy concerns. The majority of CIOs are bypassing adoption of tools, such as ChatGPT or Midjourney, and instead, waiting for professional-grade generative AI tools with oversight, Info-Tech Research Group found.
