OpenAI's GPT-3 natural language model is faster and cheaper to use, thanks to the new API function that integrates fine-tuning, so users can customize their models to produce better results from their workloads .

By automating fine-tuning of the API, developers can now create GPT-3 versions tailored to their enterprise applications. According to independent AI research and deployment companies, developers only need to use a single command in the OpenAI command line tool to start using customized APIs. The customized version will start training and then immediately available to the API.

"You can use existing datasets of almost any shape and size, or gradually add data based on user feedback," OpenAI said in a blog post about the announcement. "Through fine-tuning, one API customer was able to increase the correct output from 83% to 95%. By adding new data from their product every week, another reduced the error rate by 50%."

OpenAI stated that the custom GPT-3 API function has been Alpha tested with customers for workloads such as content generation, classification and text summarization, and has received positive reviews from users. They said that the calculation cost is lower, the training speed is faster, and the results are different. change.

Rachel Lim, the principal engineer and application engineer of the API fine-tuning project, said: “People have been calling for fine-tuning because it basically unlocks new features that the basic GPT-3 model can’t achieve.” The OpenAI team told EnterpriseAI . "If you look at the pre-GPT-3 models, many models have to be fine-tuned to work. So now with GPT-3, this is the first of many large language models, and it’s the first to be used without fine-tuning This kind of model. But imagine if you add a layer of fine-tuning to it, it will only make it more capable and able to solve more use cases."

Lim said that fine-tuning will lead to better and more reliable results. "There is a huge difference between the cool demos you can show on Twitter and the production-quality apps in which your paying users expect the results to be of a certain quality and robustness. Therefore, we found that fine-tuning usually Push the product to the end, so that it can provide more reliable services for customer-specific use cases."

Lim says that fine-tuning uses training data to "guide" the model, or adjust it to give "weights" inferred from information. The model uses these weights to make predictions.

"This allows you to control the model even before you give it [instructions, for example]'English, German, English, German, English, German'," Lim said. "Now, it already knows that it is a model that should be made in English, German, English, German, English, German. Because of this, you don't have to give it a few shots (a few) examples. For example, you only need to say'English' , It will know that it should be'German' next."

Lim said that because many users require more accurate results than what is possible with the standard GPT-3 model, fine-tuning of the function is required.

"A lot of people have used [the original GPT-3] to build production-quality applications, but for some users, 85% accuracy and 95% accuracy are completely irrelevant to whether you can get people to pay. Different from your product," she said. "The increase in reliability is indeed the main value people get from fine-tuning."

Other important benefits are lower cost of use, because it needs to process less data to achieve its results, and because the data is provided to the customized model in advance, it needs less data next time because it has identified the customized mode , Lin said.

In November, OpenAI announced the cancellation of its waiting list for running workloads on GPT-3, making its AI modeling capabilities immediately available to developers and enterprises to solve the most challenging language problems.

OpenAI first launched its powerful GPT-3 natural language model in June 2020. Its test capacity is limited, and there is also a waiting list where developers can register to use its infrastructure and features in the future.

The general version adds conditions that prevent GPT-3 from being used to harm humans, and conditions that allow it to be used only in certain countries in the world. This means that developers in some countries, including Cuba, Iran, and Russia, cannot currently access it.

GPT-3 is a large-scale natural language model that runs exclusively on Microsoft Azure. GPT-3 stands for Generative Pre-trained Transformer 3, which is an autoregressive language model with 175 billion parameters. OpenAI claims that it is ten times that of any previous non-sparse language model.

The first version, GPT-1, was released in 2018, and the second version, GPT-2, debuted in 2019. With the release of GPT-3 in 2020, natural language processing (NLP) is in the enterprise more than ever.

The latest OpenAI API (including GPT-3 and now easily available) contains many security improvements, including the Instruct Series model to better comply with human instructions, dedicated endpoints for more realistic question answering, and free content filters to aid development Reduce accidental abuse by personnel.

So far, GPT-3 is mainly used for English language modeling.

In the past few years, GPT-3 has been gaining more and more attention in the field of enterprise IT.

Microsoft obtained an exclusive GPT-3 license from OpenAI in September 2020, expanding the existing relationship between the two companies. The license agreement covers GPT-3's use of all Microsoft products and services. In May, Microsoft announced the first integration of GPT-3 into one of its products, Microsoft Power Apps software, which is designed to make it easier for corporate employees to build no-code and low-code applications.

Microsoft's license agreement is not an exclusive arrangement. According to OpenAI, others can still use the GPT-3 model through OpenAI's API.

In May of this year, OpenAI launched a $100 million AI venture fund to provide investment funds for startups that are promoting interesting AI technologies. The fund will help some early-stage startups in areas where artificial intelligence can have a transformative impact, including healthcare, climate change, education, and areas where artificial intelligence tools can help people increase productivity through the use of personal assistants and semantics. Search, the company said.

According to OpenAI, many companies are now using GPT-3 for development, including Disney, IBM, Twitter, Salesforce, Cisco, and Intel.