news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

TencentTechnology News, August 21, local time in the United States on Tuesday,OpenAIA new feature was released that will allow enterprise customers to customize the company'sAI (AI) The most powerful model for startupsGPT-4o。

OpenAI plans to launch the customization feature on Tuesday, commonly known in the AI ​​industry as fine-tuning. Fine-tuning allows existing AI models to be trained with additional information about a specific task or subject area.

For example, a company that manufactures skateboards might fine-tune an AI model so that it can be used as a customer service chatbot that can answer questions about wheels and skateboard maintenance details.

Fine-tuning, the latest feature of OpenAI’s flagship model, is particularly important at a time when startups are competing fiercely in the field of artificial intelligence products and companies are eager to demonstrate a return on their investments in artificial intelligence.

It is worth noting that this feature is introduced for the first time in GPT-4o and its predecessors, and OpenAI has previously opened up fine-tuning rights for multiple types of models including GPT-4o mini to users, providing users with more cost-effective options.

Faced with model customization services from many technology companies on the market, Olivier Godement, product director of OpenAI API, emphasized that OpenAI is committed to establishing direct partnerships with companies to simplify and accelerate the process of customers adjusting their top models, thereby avoiding users turning to external services or less powerful alternatives.

He also noted: “We are always committed to lowering the technical threshold and reducing operational barriers, thereby reducing the workload for users to start and adjust models.”

When customers fine-tune the model, they need to transfer the data to the OpenAI server. According to John Allard, a software engineer for custom work at OpenAI, this process takes about one or two hours on average. Initially, fine-tuning is limited to text data and does not support images or other media formats.

As OpenAI distributes free tokens, it is facingGoogle、AnthropicProprietary model suppliers such as Nous Research Hermes 3,Meta Fierce price war for open source models like Llama 3.1.

However, the advantage of using OpenAI and similar closed-source/proprietary models is that developers do not need to undertake the server hosting task of model inference or training by themselves. They can either use OpenAI server resources or seamlessly connect to their preferred server through the API.

However, research shows that fine-tuning models can bring risks, including deviations from the original safety guardrails and performance guarantees, which in turn affects their overall effectiveness. Whether companies are willing to take this risk needs to be weighed by themselves. But OpenAI clearly believes that it is worth it and encourages users to consider fine-tuning as an optimization option.

Separately, OpenAI said Tuesday it will feature content from brands including Vogue, The New Yorker and Wired in its products. The agreement also allows OpenAI to use Condé Nast content to help train its artificial intelligence models, which require large amounts of data to learn.

The announcement marks a step toward OpenAI’s efforts to strike deals with media companies, rather than fight them over how to use news articles and other content in its AI tools. The two sides did not disclose the amount of the agreement.

The following is the full text of the official news released by OpenAI:

Today, we’re launching fine-tuning for GPT-4o, one of the most requested features from developers, and we’re also providing 1 million training tokens per day for free to every business until September 23rd.

Developers can now fine-tune GPT-4o using its unique dataset to achieve higher performance at a lower cost for specific use cases. Fine-tuning techniques give the model the ability to flexibly adjust the structure and tone of responses, and even follow complex and highly specialized domain instructions, with only a small number of training samples (such as dozens of examples) to bring significant results to applications.

From coding to creative writing, fine-tuning capabilities cover a wide range of areas, profoundly affecting and improving the overall performance of the model. This is just the beginning, and we will continue to invest in expanding our model customization options for developers.

Starting today, GPT-4o fine-tuning is now fully available to all paying developers. Please visit the fine-tuning dashboard directly, click the "create" button, and then select "GPT-4o -2024-08-06" from the base model drop-down list to start the fine-tuning process. Regarding the cost, the GPT-4o fine-tuning training cost is set at $25 per million tokens, while the inference cost is $3.75 per million input tokens and $15 per million output tokens.

In addition, the GPT-4o mini fine-tuning function is also open to all paying developers. You only need to select "GPT-4o-mini-2024-07-18" as the base model in the fine-tuning dashboard. Special offer: To celebrate the launch, we provide up to 2 million training tokens for free every day for GPT-4o mini users. This offer is valid until September 23.

Fine-tuning success examples

Over the past few months, we’ve worked with many trusted partners to fine-tune GPT-4o and understand their use cases. Here are a few successful examples:

1. Cosine performs amazingly in the SWE-bench benchmark

Startup CosineGenieIt is an artificial intelligence software engineering assistant that can autonomously identify and fix vulnerabilities, build features, and efficiently collaborate with users to refactor code. It can also reason about complex technical issues and make changes to the code with higher accuracy and less token requirements.

Genie is powered by a fine-tuned GPT-4o model that incorporates the real-world experience of real software engineers, allowing the model to learn to respond in specific ways. In addition, the model has learned to format the output into specific formats such as patches that are easy to integrate into the code base, further improving work efficiency.

The SWE-bench validation benchmark results released last Tuesday showed that Genie achieved 43.8%SOTAThe scores are the best in the world, especially in the Full test, where its SOTA score is as high as 30.08%, a significant leap from the previous best score of 19.27%, marking a major breakthrough in the history of the benchmark. SWE-Bench is a test for evaluating artificial intelligence software engineering capabilities.

2. Distyl ranks first in BIRD-SQL benchmark

Distyl, the AI ​​solutions partner for Fortune 500 companies, recently ranked first on the BIRD-SQL benchmark, the leading text-to-SQL benchmark. Distyl’s fine-tuned GPT-4o achieved 71.83% accuracy on the leaderboard, performing well on tasks such as query reformulation, intent classification, thought chaining, and self-correction, with a particularly strong performance on SQL generation.

Data Privacy and Security

Fine-tuning the model is completely under the control of the user, who has absolute ownership of the business data, including all inputs and outputs. This ensures that the user's data will never be shared or used to train other models.

In addition, we deploy multiple layers of security protection for fine-tuned models to prevent abuse. For example, we continuously run automated security assessments on fine-tuned models and monitor usage to ensure that applications comply with our usage policies.

We are eager to see what you can create by fine-tuning GPT-4o. If you are eager to explore more model customization possibilities, please feel free to contact our team, we will be happy to provide you with support and assistance! (Compiled by Jinlu)