We’re here at the SVN West facility on a rainy Monday morning in San Francisco, as OpenAI, the creator of ChatGPT and GPT-4, holds its first-ever conference for developers, Dev Day.
The only swag so far are a bunch of nifty pins with the OpenAI logo and labels such as “Engineering” and “Research” and “Go to market.” Another set of pins represents personal pronouns.
Sam Altman takes the stage. Recaps the various milestones — ChatGPT a year ago, followed by GPT-4, “still the most powerful model”
The company disclosed it has over 2 million developers building on its APIs “for a wide range of use cases,” and 92% of the Fortune 500 companies. ChatGPT itself gets about 100 million weekly active users, the company said.
Altman goes into a rapid-fire itemization of the many innovations being announced. The various announcements receive healthy rounds of applause — it’s an upbeat, enthusiastic crowd.
Altman brings a special guest on stage: Satya Nadella, CEO of Microsoft. “The first thing we have been doing in partnership with you is building the system,” says Nadella. “The shape of Azure is drastically change in support of this,” he noted. “We want to build our co-pilot as developers on OpenAI API.”
“A couple of things are going to be very, very key for us,” said Nadella. “We intend fully to commit ourselves deeply to make sure you have not only the best models but also the best compute,” said Nadella. “Our mission is is to empower every individual.”
The venue of Dev Day, the SVN West event space at Van Ness and Market in downtown SF. Tiernan Ray
The key product and technology news includes:
- GPTs: custom versions of ChatGPT that OpenAI says “anyone can easily build” for specific tasks. The company is offering two initial custom GPTs, Canva and Zapier AI, for the popular design app and the workflows software, respectively. The company plans to offer additional GPTs;
- GPT Store: Later in November, OpenAI will open the GPT Store, for obtaining the GPTs others have built, where developers can earn money for their creations;
- Copyright Shield that will absorb cost of defending customers;
- Fine-tuning service for GPT-4 for developers;
- Custom Models program for enterprises, a team of OpenAI researchers who will work with “selected organizations” to “train custom GPT-4 to their specific domain”;
- A new user interface for ChatGPT, a simple, dark background with the OpenAI logo, and the phrase, “How can I help you today?” The new user interface will make it easier to juggle between ChatGPT and DALL*E, the image-creation program from OpenAI, the company said;
- GPT-4’s information source gets updated to April, 2023, a big step past the traditional limit imposed on the program of September, 2021. ChatGPT also gains an ability to search PDFs and other documents;
- The GPT-4 program gets its “context window,” the amount of input it can take into account when formulating an answer, four-fold, from 32,000 to 128,000, in a new “Turbo” version of the program. (For more on the various features of GPT models, see OpenAI’s Web site.);
- GPT-4 Turbo can now accept images as part of the prompt, and can generate “human-quality speech” as its output;
- Assistant API: a function-calling mechanism that makes it easier for developers to plug specific “assistant” functions into their apps, such as “natural language-based data analysis app, a coding assistant, an AI-powered vacation planner, a voice-controlled DJ, a smart visual canvas.”
- A new “seed” parameter makes GPT return “reproducible outputs” “most of the time”;
- A new version of GPT-3.5 Turbo that gains greater function handling and JSON handling;
- Cut the price on GPT-4 Turbo and GPT-3.5 Turbo, based on the price per input and output tokens, and doubled the rate of “tokens per minute” that can be used.
Said Altman of custom models, “We won’t be able to do this with many companies to start, and it won’t be cheap.”
The Copyright Shield program, said Altman, “means that we will step in to defend our customers” in the instance of litigation, “and absorb the cost.”
OpenAI says a key element of the Assistant API is “persistent threads,” which “allow developers to avoid re-sending the entire conversation history with every new message and work around context window constraints.
The new GPT-4 Turbo can be accessed immediately in preview form, OpenAI said, by passing the command gpt-4-1106-preview to the to the OpenAI API. A stable version is planned to be released “in the coming weeks.”
GPT-4 Turbo does a better job, the company said, of following specific instructions, such as to output a response in XML. It also gains a capability for responses in JSON via a new parameter, “response_format.”
The new seed parameter for GPT-4 is a beta feature that “is useful for use cases such as replaying requests for debugging, writing more comprehensive unit tests and generally having a higher degree of control over the model behavior,” the company said.
The company disclosed it has over 2 million developers building on its APIs “for a wide range of use cases,” and 92% of the Fortune 500 companies. ChatGPT itself gets about 100 million weekly active users, the company said.
The cuts mean that GPT-4 Turbo, for example, is now a penny per input token, versus three cents before, and three cents per output token, versus six cents before — which the company bills as “three times cheaper” and “two times cheaper,” respectively.
