While waiting for GPT-4, OpenAI is developing the GPT-3.5 language model

More than two years after the launch of the GPT-3 language model, OpenAI plans to launch its successor GPT-4 in the medium term. Meanwhile, the artificial intelligence company has developed a series of AI applications, including a chatbot, based on GPT-3.5, an improved version of GPT-3.

text-davinci-003, ChatGPT: these tools are based on OpenAI’s GPT-3.5

At the end of November, OpenAI announced the release of a new version of the GPT-3 language model dubbed text-davinci-003. This tool makes it possible to manage more complex instructions and generate much more precise renderings than before. On December 1, 2022, as part of of a public demonstrationOpenAI presented the features offered by its latest tool, the ChatGPT chatbot.

This has the particularity of addressing a huge list of subjects, including more technical themes such as complex scientific concepts. Unlike GPT-3 which can predict which text follows a string of words provided by a person, ChatGPT tries to respond to user queries in such a way that its response approximates that which a human might have formulated. The chatbot answers questions, and can even admit mistakes if a user proves them wrong. Thanks to reinforcement learningthis one will remember that he made a mistake and will avoid repeating the error.

What do text-davinci-003 and ChatGPT have in common? Both exploit the same language model recently developed by OpenAI, intended to be superior to GPT-3.

GPT-3.5 before the arrival of a more efficient GPT-4 in 2023?

GPT-3.5, as its name suggests, acts as an intermediary between GPT-3 and the future GPT-4. It was trained using code that was developed over a year ago. The model has learned the various meanings of the words, so that they can link them together and form coherent sentences while avoiding discriminatory biases. Like GPT-3, it has been trained using hundreds of thousands of web pages: Wikipedia articles, social media posts, press articles, blog posts, etc.

For OpenAI, GPT-3.5 is a gateway for creating GPT-4. The firm does not necessarily seek to build a model with a large number of parameters in order to surpass neural networks such as those used for Gopher (280 billion parameters) or the Chinese model PanGu-Alpha (200 billion).

Its objective will be to propose in 2023 a language model capable of carrying out more precise searches thanks to a more advanced data analysis process and to ensure that the generation of text is just as credible as that proposed by the best writers. . Finally, OpenAI will work on a better contextual understanding in order to maintain consistency throughout the dialogue with a chatbot for example.

We want to give thanks to the author of this write-up for this remarkable material

While waiting for GPT-4, OpenAI is developing the GPT-3.5 language model


Explore our social media accounts along with other pages related to themhttps://www.ai-magazine.com/related-pages/