news

Cailianshe News, August 21 (Editor: Niu Zhanlin)Eastern Time Tuesday,OpenAIReleased a new feature that allows enterprise customers to customize its most powerful models using their own company dataGPT-4o, which will greatly improve the performance and accuracy of the application.

The move comes as start-upsArtificial Intelligence (AI)Faced with increasingly fierce competition in the product field, enterprises are also facing increasing pressure to demonstrateAIReturn on investment.

OpenAI’s latest customization feature is often referred to in the AI ​​industry as fine-tuning, which allows existing AI models to be trained on additional information for a specific task or subject area.

The feature is new for OpenAI’s flagship model, and the company is also allowing users to fine-tune many of its other models, including GPT-4o mini, a cheaper, leaner version of GPT-4o.

Over the past few months, OpenAI has been testing the fine-tuning feature with partners and has achieved a range of results.

For example, the AI ​​assistant from startup CosineGenie, designed to assist software developers. It is reported that Genie has outperformed its competitors in benchmark tests and demonstrated its outstanding ability to collaborate with users to autonomously identify and resolve errors, build features, and refactor code.

By fine-tuning the GPT-4o model, GenieSWE-BenchThe AI ​​model achieved a 30% score in the test, which is the highest score ever achieved by an AI model in this field. SWE-Bench is a test for evaluating AI software engineering capabilities.

Olivier Godement, OpenAI API product director, said the company hopes to make it easier for customers to adjust its most powerful models by working directly with companies, rather than using external services or less powerful products. "We have always been very concerned about lowering the threshold, reducing the inconvenience of use, and making it easier to get started with products or services."

Data Securityquestion

In order to fine-tune the model, customers must upload their data to OpenAI's servers, so data security has become a major concern for customers.

Recently, JPMorgan Chase has partnered with OpenAI to launch an artificial intelligence assistant to tens of thousands of employees to help them write emails, generate ideas, summarize documents, etc. This is also the initial application of artificial intelligence technology in the financial field.

However, due to concerns about data security, thisLLM The Suite's assistant is designed as a "portal" that allows users to access external large language models through the LLM Suite without exposing the data to the outside world.

In this regard, OpenAI emphasized that the fine-tuning model is completely under the control of corporate customers, and all input and output data will not be used to train other artificial intelligence models.

OpenAI has also implemented layered security mitigations for fine-tuned models to ensure that these models cannot be abused. For example, they continuously perform automated security assessments on fine-tuned models and monitor usage to ensure that applications comply with their usage policies.

Training takes an hour or two on average, said John Allard, a software engineer at OpenAI who works on customizations. Allard added that initially users will be able to fine-tune the model only with text-based data, not images or other content.

(Niu Zhanlin, Cailianshe)