OpenAI recently announced a new feature allowing enterprise customers to customize their most powerful model, GPT-4o, using their own company data. This customization process, known as fine-tuning, enables existing AI models to receive additional training on specific tasks or subject areas.
For example, a skateboard manufacturer could fine-tune an AI model to serve as a customer service chatbot capable of answering detailed questions about wheels and skateboard maintenance.
This feature is particularly significant as it's the first time fine-tuning has been introduced for GPT-4o and its predecessors. OpenAI had previously allowed users to fine-tune other models, including GPT-4o mini, offering more cost-effective options.
Olivier Godement, OpenAI's API product lead, emphasized their commitment to simplifying and accelerating the customization process for their top models through direct collaboration with enterprises. This approach aims to prevent users from turning to external services or less capable alternatives.
The fine-tuning process requires customers to transfer data to OpenAI's servers, typically taking about one to two hours. Initially, fine-tuning will be limited to text data, with no support for images or other media formats.
While fine-tuning offers potential benefits, research suggests it may also carry risks, including deviations from original safety guardrails and performance guarantees. Companies will need to weigh these risks against the potential advantages.
OpenAI has also announced partnerships to display content from brands like Vogue, The New Yorker, and Wired in its products. This agreement allows OpenAI to use Condé Nast's content for training its AI models.
The company is offering 1 million free training tokens per day for each enterprise until September 23rd. GPT-4o fine-tuning is now available to all paid developers, with training costs set at $25 per million tokens and inference costs at $3.75 per million input tokens and $15 per million output tokens.
OpenAI has highlighted successful examples of GPT-4o fine-tuning, including Cosine's Genie AI software engineering assistant and Distyl's performance in the BIRD-SQL benchmark test.
Regarding data privacy and security, OpenAI emphasizes that users retain full ownership of their business data, including all inputs and outputs. The company has implemented multi-layered security measures to prevent misuse of fine-tuned models.