news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Author: Du Yu

Source: Hard AI

On Tuesday, August 20, OpenAI announced that it will launch a "personalized customization service" for corporate customers, which means that they can use their respective company data to "fine-tune" OpenAI's most powerful AI model GPT-4o.

Some analysts said that fine-tuning is a brand new feature for OpenAI's current flagship model, and neither GPT-4o nor its predecessor GPT-4 had previously provided fine-tuning functionality.The latest move illustrates the intensifying competition among major AI model development companies for corporate customers.

Currently, competition among startups for commercial AI products is becoming increasingly fierce, and corporate customers are facing increasing pressure to prove that they can obtain significant returns on their AI investments.

OpenAI's competitors, especially Google and Microsoft, are providing the ability to customize artificial intelligence models. OpenAI released a self-service fine-tuning API for the GPT-3 model in 2023, and launched a cheaper GPT-4o mini model fine-tuning service in July this year. In addition, AI model fine-tuning has also spawned "intermediaries" such as Cloudera that help companies train general models based on data sets in specific fields.

Olivier Godement, head of API products at OpenAI, told the media that OpenAI hopes to help enterprise users more easily adjust their most powerful models by working directly with them, rather than using external services or less powerful model products: "We have always paid great attention to lowering the threshold, reducing friction, and reducing the amount of work required to get started."

To fine-tune the model, customers must upload a specific dataset to OpenAI’s servers.OpenAI"Data privacy and security" was emphasized in the official website announcement：

“The fine-tuned models are fully controlled by you and you fully own your business data, including all inputs and outputs. This ensures that your data can never be shared or used to train other models.
We also implement layered security mitigations for our fine-tuned models to ensure they cannot be misused. For example, we continuously perform automated security assessments on our fine-tuned models and monitor usage to ensure applications comply with our usage policies.”

John Allard, an OpenAI software engineer in charge of model customization, said that fine-tuning training takes an average of one to two hours. Similar to GPT-4o mini, initially users can only use text-based data to fine-tune the model, but not images or other content.

OpenAI's official announcement stated that customized versions of GPT-4o after fine-tuning can improve the performance and accuracy of applications, and the GPT-4o fine-tuning function is now open to all developers at all paid usage tiers.

In terms of costs,GPT-4o fine-tuning training costs $25 per million tokens, inference costs $3.75 per million input tokens, and $15 per million output tokens, but 1 million training tokens will be provided free of charge per organization per day before September 23. GPT-4o mini fine-tuning will provide 2 million training tokens for free per day before September 23.

Model fine-tuning refers to adjusting and optimizing the parameters of a pre-trained general large model for a specific task or data set to improve the model's performance in a specific type of task or subject area. According to Baidu Smart Cloud:

Although pre-trained large models have powerful general capabilities, they often seem to be unable to cope with complex problems in specific fields. Fine-tuning is similar to supplementing and strengthening professional knowledge based on general knowledge, making the model more "down-to-earth" and better able to solve practical problems.
Model fine-tuning enables the model to quickly adapt to new tasks and improve task completion and accuracy through targeted optimization while maintaining the original knowledge. This process not only accelerates the implementation of AI technology, but also reduces the cost and threshold of model development.

OpenAI said that the GPT-4o fine-tuning function is one of the most anticipated features by developers. From coding to creative writing, fine-tuning will have a huge impact on model performance in various fields. And OpenAI will continue to invest to expand the customization options of its models for developers:

Developers can now fine-tune GPT-4o using custom datasets to achieve higher performance at a lower cost and meet the needs of specific use cases. Fine-tuning enables the model to customize the structure and tone of responses or follow complex domain-specific instructions. Developers can produce strong results for their applications using only a few dozen examples from the training dataset.