OpenAI CEO Sam Altman gave a presentation at the DevDay conference to talk about upcoming updates to the ChatGPT chatbot and new tools for developers.
Turbo mode
First of all, the team introduced an improved version of GPT-4 Turbo with an expanded context window capacity of 128,000 tokens. This value is equivalent to 300 pages of text in a single query.
The artificial intelligence is more functional and has knowledge of world events until April 2023.
The chatbot has received an updated function call system that allows you to create a request for two actions at the same time, such as "open the car window and turn off the air conditioner." An improved neural network API remembers and is "more likely" to reproduce the correct function parameters.
In addition, GPT-4 Turbo has been taught to carefully follow the requested format when a special parameter is specified (e.g., "always respond in XML"). The neural network also supports JSON mode to compose the obtained results.
A new parameter, seed, ensures reproducibility of the output by forcing the neural network to repeat consistent results. A beta feature that provides more control over model behavior is useful for creating multiple debugging requests and complex unit tests.
In addition OpenAI has released a turbo version of GPT-3.5 with a context window for 16,000 tokens. The neural network supports similar functionality to GPT-4 Turbo, but in a slower mode.
Useful Assistant
API Assistants are purpose-built AI that has specific instructions, uses additional knowledge, and can call models and tools to perform tasks.
The assistant interface provides interpreter and code extraction capabilities in Python. The tool is also able to execute some functions that previously had to be manually prescribed, and allows for the creation of "high-quality artificial intelligence applications."
"The API is designed with flexibility in mind: use cases range from a natural language data analysis application, programming assistant, AI-powered vacation planner, voice-controlled DJ, intelligent visual canvas - the list goes on for a long time," OpenAI emphasized.
Additional features
The advanced functionality that the turbo versions of GPT have received makes it possible to implement additional solutions. For example, the Chat Completions API allows the AI to perceive images as input data, so that the neural network can create captions for drawings, make detailed analysis of photos, or read and then retell documents.
ChatGPT's "vision" has given rise to the BeMyEyes app, which uses the extension to help blind and visually impaired people with everyday tasks like navigating indoors.
Developers can now integrate the DALL-E 3 generative neural network into their products directly through the interface. The tool has built-in moderation of output content to combat copyright infringement.
In addition, OpenAI products have gained full support for text-to-speech functionality and have six preset voices. The extension has various modes, for example for real-time conversation or creating a high quality audio track.
For those who are not satisfied with the usual ChatGPT functionality, an experimental "fine-tuning" feature and a Custom Models tool have been added. This will allow you to change the language model code at any stage, starting with training.
Lastly, OpenAI introduced custom, narrowly focused versions of the neural network - GPTs. They are something similar to web browser extensions.
Some can perform specific tasks to search for information on the Internet or serve as a virtual assistant in work processes.
According to a company statement, GPTs don't require you to write code to create them. The tool can be made for personal use, corporate use, or shared.
A dedicated extension store will be available later in November. Third-party users will be able to add their developments to the platform, the best of them will be chosen by the OpenAI team.
Top comments (0)