OpenAI’s first developer conference was held today and many new features and updates were announced for their GPT and DALL·E platforms. These updates include a new GPT-4 Turbo model, Assistants API, and many updates to DALL·E 3.
Next Generation of GPT-4: GPT-4 Turbo with 128K Context
With the release of GTP-4 in March which was made widely available to all developers in July, a new version has already been released called GPT-4 Turbo, this next generation of GPT is more capable and is up to date on events through April 2023. GPT-4 Turbo now has a 128k context window in order to fit more than 300 pages of text in a single prompt, this also has lead to an optimization in overall performance. OpenAI is now offering GPT-4 Turbo at a 3x cheaper price for input tokens, and 2x cheaper price for output tokens when compared to GPT-4. GPT-4 is already available to all paying developers via passing the gpt-4-1106-preview in the API, and a stable model will be coming within weeks of this post.
What is a Language Model Context Window?
In language models like GPT (Generative Pre-trained Transformer), a context window refers to the number of preceding tokens or words that the model considers when generating predictions for the next token in a sequence. The model processes input data in chunks, and the context window determines how much of the past information the model takes into account. Imagine you’re trying to predict the next word in a sentence. The context window is like the number of words you look at before the word you’re trying to predict. If the context window is, say, 10 words, you’re considering the last 10 words to make your prediction.
A larger context window in a language model like GPT can be helpful in capturing longer-term dependencies and understanding more nuanced relationships between words in a given sequence.
Improved GPT-3.5 Announced with Introduction of Updated GPT-3.5 Turbo
With the updated GPT-4 Turbo OpenAI has also announced a new verion os GPT-3.5 Turbo which will support a 16K context window by default. The new GPT-3.5 Turbo will support improved instruction following, JSON mode, and parallel function calling. Open AI has state that internal evaluations have show n a 38% improvement on format following tasks such as generating JSON, XML and YAML. Current developers can access this new model by calling gpt-3.5-turbo-1106 in the API.
When working with tasks like generating JSON, XML, or YAML using a language model like GPT, the context window plays a role in how the model understands and generates the structured output. These formats often require a specific structure and adherence to rules, and a larger context window will help developers with improved semantic understanding and context-aware assistance during code and structured data generation.
OpenAI Assistants API Released
OpenAI has also released the new Assistants API which they hope will help developers building agent-like experiences in their own applications. An assistant in this context refers to a specialized AI designed with specific instructions, harnessing additional knowledge and the ability to invoke models and tools for task execution. The recently introduced Assistants API comes with enhanced features like Code Interpreter and Retrieval, along with function calling, alleviating much of the intricate work that was traditionally manual, leading to an easier way to build high quality AI apps in the future.
OpenAI Assistant will lead to the creation of many new applications such as:
- Natural Language-Based Data Analysis Apps: Allow users to interact with and analyze complex datasets using natural language queries.
- Coding Assistant: Coding assistant that understands natural language queries about programming tasks. Developers can ask for code snippets, explanations, or debugging help, and the AI assistant can provide relevant suggestions and solutions.
- AI-Powered Vacation Planner: Build a vacation planning app that understands user preferences expressed in natural language. The AI assistant can suggest personalized travel itineraries, recommend attractions, book accommodations, and provide real-time updates based on user feedback.
- Voice-Controlled DJ: Design a DJ application that responds to voice commands for song selection, mixing, and genre preferences.
- Smart Visual Canvas: Develop a visual design tool where users can sketch or describe their ideas, and the AI assistant interprets these inputs to generate or enhance visual elements.
- Health and Wellness Coach: Create a personalized health and wellness assistant that understands users’ fitness goals, dietary preferences, and health conditions.
- Language Learning Companion: Develop an AI-powered language learning app that engages users in natural conversations, provides instant feedback on pronunciation, suggests relevant vocabulary, and adapts lessons based on individual learning styles.
- Automated Customer Support: Implement an AI-driven customer support system that understands and responds to customer queries in natural language. The assistant can provide information, troubleshoot issues, and escalate complex problems to human support when necessary.
DALL·E 3 Improvement and ChatGPT Integration
Developers can now integrate DALL·E 3, as a ChatGPT Plus and Enterprise user, directly into their apps and products cia OpenAI’s Images API by specifying dall-e-3 as the model. This helps companies to generate images and designs for customers and campaigns mor easily.
ChatGPT will also be integrated more fully with DALL·E 3 to help create new images from simple conversations, this new feature is already available to Plus and Enterprise users. This will allow you to describe your image, and ChatGPT will generate it by providing a selection of visuals you can refine and iterate upon. You can also ask for revisions right in the chat.