What did OpenAI announce at its first DevDay, and why does it matter?

Yesterday OpenAI held DevDay, the first conference it has ever organized for developers, the people who build apps and services on top of its models. In one keynote and a long blog post, the company announced a new flagship model, a new way to build assistants, custom versions of ChatGPT, several new developer features and price cuts across its platform.

If you only use ChatGPT in a browser, some of this sounds technical. It still matters to you: most of the AI features appearing in other apps run on these same tools, and many of the announcements make those tools cheaper, faster or able to read far more text at once.

When and where it happened

OpenAI first announced DevDay in September as a one-day event in San Francisco on November 6, 2023. It promised to bring together hundreds of developers from around the world, with breakout sessions run by its technical staff and a keynote streamed live for everyone else. On Monday, Sam Altman, OpenAI's CEO, took the stage to open it.

The company opened with numbers. About 100 million people now use ChatGPT every week, it said, and more than two million developers build on OpenAI's API (the programming interface that lets software send requests to its models). New features began reaching customers at 1 p.m. Pacific time the same day.

What was announced

Here is the day in one table, from OpenAI's announcement:

A few smaller items round it out: a new GPT-3.5 Turbo that handles 16,000 tokens by default, a "JSON mode" that makes the model reply in valid structured data, a seed setting for repeatable answers, an experimental program for fine-tuning GPT-4 (further training it on a company's own examples), a Custom Models program for a few large organizations, and the release of Whisper large-v3, the newest version of OpenAI's open-source speech recognition model.

How the pieces fit together

The easiest way to see DevDay is as three layers.

The bottom layer is the model. GPT-4 Turbo is the engine: it reads your text, works out a reply and writes it. Two changes stand out. Its context window, the amount of text it can consider at once, grows to 128,000 tokens (a token is a word or part of a word), roughly a long book. And it is cheaper: input text costs a third of GPT-4's price, output text half.

The middle layer is the assistant. On its own, a model forgets everything between requests. The Assistants API adds the parts developers used to build themselves: a thread that stores the conversation, a retrieval tool that searches documents you upload, a Code Interpreter that writes and runs Python, and function calling, which lets the assistant ask your app to do something, such as look up an order.

The top layer is the product. GPTs package instructions, files and tools into a custom ChatGPT that people can build without writing code. OpenAI says the Assistants API rests on the same abilities that power GPTs, so a developer can build in code what a non-programmer builds in ChatGPT.

Why it matters

For builders, cost and size change what is practical. A model that is two to three times cheaper and can read a whole contract or codebase in one prompt makes projects affordable that were too expensive last week. Tools such as retrieval and Code Interpreter, offered as ready-made parts, mean a small team no longer has to assemble them from scratch.

For everyone else, DevDay signals a shift from one general chatbot toward many specialized helpers. OpenAI calls the Assistants API its first step toward "agent-like" experiences, meaning AI that can pursue a goal and use tools rather than just answer. Copyright Shield, meanwhile, addresses one of the main worries businesses have raised about using generated text and images.

There are open questions. GPT-4 Turbo is labeled a preview, with a stable version promised in the coming weeks. The Assistants API is a beta, and OpenAI has asked developers for feedback. Many expect rough edges before these tools settle.

What can developers and everyday users try first after DevDay?

If you write code, the quickest start is to switch an existing GPT-4 project to gpt-4-1106-preview and compare cost and quality on your own examples. Then open the Assistants playground on OpenAI's platform, which lets you set up an assistant with instructions, files and tools without writing code, and see how threads and retrieval behave before you build them into an app.

If you use ChatGPT Plus or Enterprise, look for GPTs. example GPTs are available to you from today, and OpenAI plans a store where builders can publish their own later this month. Building one lets you turn a prompt you reuse often into a custom assistant.

If you use the free version, nothing changes overnight, but the apps you already use may soon gain new AI features built on these tools. Knowing the three layers, model, assistant and product, will help you understand what those features can and cannot do.