OpenAI Announces New Capabilities, Developer Products
At its first-ever Devay conference for developers, OpenAI this week announced a slew of features that include a significant upgrade to its ChatGPT service, an improved version of GPT-4, and new capabilities designed with developers in mind.
According to OpenAI CEO Sam Altman, ChatGPT currently has 100 million weekly active users and two million developers currently building on its APIs. This was achieved in slightly less than a year since the launch of ChatGPT on November 30 last year.
ChatGPT
According to OpenAI, GPT-4's information source has been updated to April 2023, a huge step past the previous date of September 2021. In addition, the company unveiled “GPTs”, a feature that OpenAI bills as the new way for anyone to easily create custom versions of ChatGPT.
While some users have taken to maintaining curated libraries of carefully crafted prompts and instruction sets, they could in theory put everything into standalone GPTs. In addition, external developers can offer APIs for access by GPTs, allowing them to connect to databases, access emails, integrate with travel listings, or even facilitate e-commerce orders.
A GPT Store will be rolling out later this month for verified “builders”, with app store-like leaderboards and categories. OpenAI says chats with GPTs are not shared with builders. Eventually, it will also be possible for builders to earn money based on how many people are using a GPT – though details were not offered.
A range of more than 10 OpenAI-supplied GPTs has already been added to ChatGPT in the meantime, including a Creative Writing Coach, Math Mentor, and Tech Support Advisor.
GPT-4 Turbo
Many of the new features appear geared toward developers. A new GPT-4 Turbo model was unveiled at the same conference, accessible via API for now. Billed as more capable, faster, and cheaper, it also supports a much larger 128K context window, pulling ahead of Anthropic’s Claude-2 and its 100k context window.
“We also optimized its performance so we are able to offer GPT-4 Turbo at a 3x cheaper price for input tokens and a 2x cheaper price for output tokens compared to GPT-4,” wrote OpenAI in a blog.
Another notable new feature is a new “seed” parameter that OpenAI says will enable reproducible outputs by making the model return consistent completions “most of the time” for reproducible outputs. This is useful for replaying requests for debugging, unit tests, and giving users a higher degree of control over the model behavior.
GPT-4 Turbo is currently in preview for paying developers who are encouraged to give it a spin. OpenAI says the stable production-ready model of GPT-4 Turbo will be released in the “coming weeks”.
Developers
Another major update is the ability of developers to access the multimodal capabilities in GPT-4. This includes GPT-4 Turbo with vision, image creation (DALL·E 3), and text-to-speech (TTS).
This means developers can finally upload images through the regular API to get GPT-4 to analyze real-world images, including reading documents with figures. This long-awaited feature was first demonstrated at the start of this year at the unveiling of GPT-4.
Using the TTS feature, developers can generate human-quality speech. Six preset voices to choose from and two model variants are currently offered, and the demonstration startlingly human voices with correct intonation.
OpenAI also released Assistants API, which it says will help developers build agent-like experiences within their applications. Dubbed a purpose-built AI designed with specific instructions and extra knowledge, it can call models and tools to perform various tasks.
Right now, Assistants have access to tools such as a Code Interpreter to write and run Python code in a sandbox, the ability to retrieve data from external repositories, and the ability to invoke custom functions.
The Assistants feature is currently in beta and can be created without writing any code at the Assistants playground.
Paul Mah is the editor of DSAITrends. A former system administrator, programmer, and IT lecturer, he enjoys writing both code and prose. You can reach him at [email protected].
Image credit: iStockphoto/OpenAI
