OpenAI_Keynote_TLDW.md

OpenAI DevDay Key Announcements

Past Accomplishments

1. A year ago, OpenAI launched ChatGPT as a "low-key research preview".
2. In March, they launched GPT-4, the most capable model in the world.
3. They have since launched voice and vision capabilities for ChatGPT.
4. DALL-E 3, the world's most advanced image model was recently launched.
5. For enterprise customers, they launched ChatGPT Enterprise with increased security, privacy, and speed.
6. OpenAI has around 2 million developers building on their API and about a hundred million weekly active users on ChatGPT.

New Announcements

1. Launching a new model, GPT-4 Turbo, with improvements based on user feedback.

   • Increased context length: GPT-4 Turbo supports up to 128,000 tokens of context.
   • More control: A new feature called JSON Mode ensures valid JSON responses. A feature for reproducible outputs is also introduced.
   • Better world knowledge: GPT-4 Turbo's knowledge about the world is updated up to April 2023.
   • New modalities: DALL-E 3, GPT-4 Turbo with vision, and new text-to-speech model are integrated into the API.
   • Customization: Fine-tuning has been expanded to the 16K version of the model.
   • Higher rate limits: OpenAI is doubling tokens per minute for all established GPT-4 customers.
   • Copyright shield: OpenAI will defend and pay the costs of its customers facing legal claims on copyright infringement.

2. Introduction of a new pricing model: GPT-4 Turbo is considerably cheaper than GPT-4 by a factor of 3x for prompt tokens and 2x for completion tokens.

3. Future improvements: OpenAI will work on improving the speed of GPT-4 Turbo.

Partnership with Microsoft

1. Microsoft is instrumental in the development of OpenAI's infrastructure.
2. Microsoft is using the OpenAI APIs to build their own products, such as GitHub Copilot.
3. Looking forward, Microsoft plans to continue building the best infrastructure in Azure with OpenAI's API support.# OpenAI's GPT-3.5 Turbo and GPT-4

• OpenAI's mission is to empower every person and organization on the planet to benefit from AI.
• Safety is a key priority for OpenAI.
• OpenAI has made significant improvements to ChatGPT.
  • ChatGPT now uses GPT-4 Turbo, which includes the latest knowledge cutoff.
  • ChatGPT now has the capability to browse the web, write and run code, analyze data, and generate images.
  • The model picker in ChatGPT is no longer required.

ChatGPT's Enhanced Capabilities

• Future AI systems, often referred to as "agents", will be smarter, more personal, more customizable, and capable of performing tasks on behalf of users.
• OpenAI believes that gradual incremental deployment is crucial to addressing safety issues associated with AI.
• OpenAI is introducing GPTs (tailored versions of ChatGPT for specific purposes).
  • GPTs can be customized with instructions, expanded knowledge, and actions.
  • GPTs can be published for others to use.
  • GPTs can be programmed via conversation, making them easily accessible.

Examples of GPTs

• Code.org has created a GPT to help teachers provide engaging experiences for middle school students.
• Canva has developed a GPT that allows users to start designing by describing what they want in natural language.
• Zapier has built a GPT that enables users to perform actions across 6,000 applications.

Building GPTs

• GPTs can be easily built and customized.
• GPTs can be created using a GPT builder.
• Developers can provide specific instructions and upload documents to guide the GPT's responses.
• GPTs can be published privately or shared publicly.

The GPT Store

• OpenAI is launching the GPT Store, where users can list and discover GPTs.
• OpenAI will share a portion of its revenue with creators of the most useful and popular GPTs.

OpenAI's Assistants API

• OpenAI is introducing the Assistants API, which includes persistent threads, built-in retrieval, a code interpreter, and improved function calling.
• The Assistants API simplifies the process of building assistive agents.# Function Calling Feature

• Function calling in AI is a powerful tool
• It now guarantees JSON output with no added latency
• It allows for invoking multiple functions at once
• The assistant can use functions to interact with components and features of an app
• Creates harmony between AI and UI

Retrieval Feature

• Retrieval gives the assistant more knowledge beyond immediate user messages
• The assistant can parse long-form documents including text and product specs
• The assistant can read and extract information from uploaded files

Assistants API

• Manages state for user conversations
• Leverages external tools like knowledge, retrieval, and Code Interpreter
• Invokes your own functions to make things happen
• Not a black box - developers can see the steps that the tools are taking within the developer dashboard

Code Interpreter

• Now available in the API
• Gives the AI the ability to write and execute code on the fly, and even generate files
• Can answer complex queries by writing and executing code

Text to Speech and Voice Recognition

• Six unique and rich voices to choose from in the API, each speaking multiple languages
• Uses Whisper to convert voice inputs into text
• Uses TTS API to convert text into speech

Launch of GPT-4 Turbo Model

• Delivers improved function calling, knowledge, and lowered pricing
• Offers new modalities

Partnership with Microsoft

• Deepening partnership with Microsoft to advance AI technology

Future of AI

• AI agents will gradually be able to plan and perform more complex actions on behalf of the user
• The goal is to empower individuals and elevate humanity
• AI will provide superpowers on demand and change the world in many ways.