OpenAI DevDay Updates: November 2023

A summary of the major model, API, multimodal, customization, and copyright-related announcements, along with a personal view on what may matter most for non-developers.

GPT-4 Turbo

A new GPT-4 Turbo model was announced. It is faster and cheaper, with input pricing reduced to one-third and output pricing reduced to one-half.

Its context window has expanded from 8K and 32K options to 128K. Rate limits have also been doubled. The model includes knowledge updated through April 2023.

GPT-4 Turbo supports function calling, improved instruction following, and JSON mode.

GPT-3.5 Turbo Updates

GPT-3.5 Turbo has also been updated. Input pricing has been reduced to two-thirds of the previous price, while output pricing remains unchanged.

Its context window has moved from 4K and 16K options to 16K. It also offers improved instruction following, JSON mode, and parallel function calling.

Fixed Seeds and Log Probabilities

Both GPT-4 and GPT-3.5 now support fixed seeds and log probability output. Fixed seeds make reproducibility experiments possible, while log probabilities can be used to calculate the probability of the next token appearing.

Assistants API

The Assistants API makes it possible to call and use a custom GPT configured with a GPT-based model, instructions, relevant documents, and tools.

It can use Code Interpreter, Retrieval, and Function Calling. Conversation history is managed in the form of threads.

Custom GPT Builder

The Custom GPT Builder lets users configure a custom GPT through a natural-language conversation about the role it should perform. Files containing relevant information can also be uploaded in advance.

A Custom GPT demonstration video begins at 26:25.

Playground link: https://platform.openai.com/playground?mode=assistant

Multimodal API

The announcements also point toward multimodal capabilities beyond text-only language models.

  • GPT-4 Turbo with Vision: image input is available.
  • DALL·E 3: image generation is available.
  • Text-to-Speech: text can be converted into speech.
  • Whisper v3 for speech-to-text and Consistency Decoder for image generation were released as open source.

Model Customization and Copyright

GPT-4 fine-tuning is now available; previously, fine-tuning had been available only for GPT-3.

The Custom Models program is a service for developing models tailored to every stage of training, using data intended for internal company use.

Copyright Shield is a program that provides compensation if copyright infringement issues arise.

Personal Takeaways

From a non-developer perspective, the most notable development is the beginning of the custom GPT era. An ecosystem may emerge in which specialized GPTs can be selected and used as needed, much like choosing apps from an app store.

Cheaper, faster models with larger context windows are also important. The Assistants capabilities—file uploads, code execution, and image-generation capabilities—appear to leave significant room for automation.

Overall, the trend is moving from text-only LLMs toward multimodal AI.