OpenAI held its first DevDay, where numerous amazing new features were announced.
The OpenAI DevDay conference was hosted by the company’s CEO Sam Altman. He highlighted that ChatGPT has become one of the fastest growing consumer products and reached 100 million users per month just a few months after its launch. Today, more than 100 million people use ChatGPT every week. And more than 2 million developers create their solutions using the ChatGPT and Whisper APIs by OpenAI.
Here’s a summary of OpenAI’s major announcements in case you missed the event.
Launching GPT-4 Turbo
Users will soon gain access to a new model – GPT-4 Turbo. This model provides a 128K context window, has a lower cost, and contains the equivalent of more than 300 pages of text in a single prompt. GPT-4 Turbo encompasses knowledge of world events until April 2023. The pricing for input and output tokens has been reduced by 3 and 2 times respectively compared to GPT-4.
Assistants API
Developers can now create specialized AI assistants that implement specific instructions, use additional knowledge from the user base, and can call upon models and tools to perform their tasks (for example, a Python code interpreter).
In fact, this is an opportunity to personalize the OpenAI model for your tasks and your knowledge base.
Multimodal capabilities in the API
In GPT-4 Turbo, you will be able to transmit not only text but also images through the API and create prompts about them. It will also be possible to generate images using an API request and convert text to human-like speech using the speech synthesis API.
Reproducible Model Calls
API calls now have a seed parameter – a beta function that allows you to get reproducible results, which is useful for cases of reproducing requests for debugging, writing more diverse unit tests, and generally monitoring model behavior.
Log probabilities
In a few weeks, OpenAI plans to launch a function that returns the log probabilities of each next token in GPT-4 Turbo and GPT-3.5 Turbo. This will be useful for developing features such as autocomplete in search engines, as well as for detecting model hallucinations.
More tokens per minute
Token limits per minute for paid GPT-4 users have been doubled. The new limits are available according to the usage tier.
Custom Models
A new program is available for organizations that require more customization than the usual model setup. The program allows you to work with OpenAI researchers to create custom GPT-4 models for specific domains.