OpenAI Day 9: Honoring Global Developers: Enhancing Developer Experience
Opening Introduction
Welcome Speech by Olivia Gar
Olivia Gar, the platform product lead at OpenAI, kicked off Day 9 with great enthusiasm. She emphasized that the focus of the day was on developers and startups, expressing pride in the achievements of the OpenAI community: since the API was launched four years ago, it has been used by 2 million developers from over 200 countries!
Event Highlights
Today’s focus is to honor the developer community by unveiling new models and features, and to thank everyone for their continued support and contributions.
New Models and Features Release
O1 Model Officially Launched
The O1 model, after incorporating developer feedback, has officially ended its preview period. Since its announcement in September, it has seen several improvements:
- Function Calling: The model can directly call backend functions for more efficient task execution.
- Structured Outputs: Results are presented in a predefined format, making data processing easier.
- Developer Messages: A new type of instruction has been added, allowing developers to guide the model’s behavior more precisely.
- Reasoning Effort Parameter: Developers can control the model’s computational effort on a problem to balance performance and cost.
- Vision Inputs: The model can process visual data, which is widely used in manufacturing and scientific fields.
Live Demo: New API Features Showcase
The live demo showcased several capabilities of the O1 model, especially in form error detection:
- Uploading a tax form scan with intentional errors, the model successfully detected:
- Arithmetic errors (e.g., addition should be subtraction).
- Errors in calculating standard deductions.
- The Function Calling technology demonstrated how the model interacts with backend APIs to retrieve correct tax data.
- The Structured Outputs feature generated correction results in JSON format, making it easy for the front-end UI to display error details clearly.
O1 Model Evaluation Results
Internal evaluations show that O1 has significantly improved in the following areas:
- Function Calling Accuracy: The model calls the correct functions more accurately.
- Structured Outputs: Improved format compliance, reducing inconsistencies.
- Code Processing Capability: Excellent performance in coding tasks, demonstrating its robustness.
- Latency: Compared to the previous model, O1 uses 60% fewer reasoning tokens, resulting in faster responses and lower costs.
Real-Time API Updates
WebRTC Support Launched
Real-time APIs now support WebRTC, a protocol designed for real-time communication. Its advantages include:
- Adaptive Bitrate: Automatically adjusts based on network conditions for smoother audio and video streaming.
- Echo Cancellation: Removes background noise to improve audio quality.
A simple example application demonstrated how to implement WebRTC with minimal code, highlighting its ease of use.
New Features for Real-Time API
- Reduced Audio Token Costs: The audio token cost for GPT-4 has been reduced by 60%, making it more affordable for developers.
- New Python SDK: Makes integrating real-time APIs simpler and more intuitive.
- API Optimization: Further improvements in function calling and usability.
New Fine-Tuning Method: Preference Fine-Tuning
Preference Fine-Tuning
Preference fine-tuning is a new method compared to traditional supervised fine-tuning, allowing developers to provide pairs of responses (preferred and non-preferred) to optimize the model’s output.
- Application Scenarios: Improves performance in customer support and content moderation.
- Simple Demonstration: The API demonstrated the simplicity and efficiency of the fine-tuning process.
Additional Announcements
- Go and Java SDKs: Official support added, expanding developer tool options.
- Simplified API Key Process: Improved the API key application process to enhance user experience.
- AMA Event Announcement: An upcoming “Ask Me Anything” event where developers can interact with OpenAI team members.
Closing Remarks and Outlook
Highlights of the Day
- Official launch of the O1 model
- WebRTC support for real-time APIs
- Introduction of preference fine-tuning
Humor and Teasers
Olivia concluded the event with humor and teased the exciting content for Day 10.
Looking Forward to Tomorrow’s Announcements
Let’s look forward to the new features and innovations revealed on Day 10!