OpenAI has introduced a new feature called GPTs (Generative Pre-trained Transformers) that allows users to create customized versions of ChatGPT for specific purposes. These tailored GPTs can be developed without coding and cater to various tasks, such as learning board game rules, teaching math to children, or designing stickers. Users can create their GPTs, keep them for personal use, share them within their company, or make them public.
To create a GPT, users need to start a conversation, provide instructions and additional knowledge, and specify its capabilities, such as web searching, image creation, or data analysis. Examples of GPTs are already available for ChatGPT Plus and Enterprise users, including Canva and Zapier AI Actions. OpenAI plans to make GPTs accessible to more users shortly.
The GPT Store, launching later this month, will showcase publicly shared GPTs, allowing them to be searchable, climb leaderboards, and be featured based on usefulness and appeal. Users will also have the opportunity to earn money based on the popularity of their GPTs. OpenAI emphasizes privacy and safety, ensuring that user chats with GPTs are not shared with builders and that users have control over data sharing with third-party APIs.
OpenAI envisions GPTs becoming more sophisticated over time, eventually handling real-world tasks. Users can connect GPTs to the real world by defining custom actions, integrating external data, and interacting with databases or emails. For enterprise users, GPTs provide the flexibility to create internal-only versions for specific use cases, departments, or proprietary datasets.
The introduction of GPTs aligns with OpenAI's commitment to involving the community in shaping AI behavior and promoting safer and more aligned artificial general intelligence (AGI). OpenAI has also updated ChatGPT Plus with information up to April 2023, simplified the user interface, and eliminated the need to switch between models, offering a more seamless experience for users.
OpenAI has launched the OpenAI Data Partnerships program, inviting organizations to collaborate to create public and private datasets for training AI models. The initiative aims to broaden the training datasets for AI models, enhancing their understanding of various subject matters, industries, cultures, and languages. This broader dataset is crucial for developing Artificial General Intelligence (AGI) that is safe and beneficial to humanity.
OpenAI is already engaged in partnerships with several organizations. For instance, the Icelandic Government and Miðeind ehf are collaborating to improve GPT-4's proficiency in Icelandic by integrating curated datasets. OpenAI has also partnered with the Free Law Project, a nonprofit aiming to democratize access to legal understanding, to include an extensive collection of legal documents in AI training.
Data Partnerships encourage organizations to contribute unique data, guiding the future of AI research and creating models that are more useful to them. OpenAI is particularly interested in large-scale datasets reflecting human society, spanning text, images, audio, or video. The emphasis is on data expressing human intention, such as long-form writing or conversations, across different languages, topics, and formats.
OpenAI currently offers two collaboration avenues:
1. Open-Source Archive: OpenAI seeks partners to assist in developing an open-source dataset for training language models. This dataset will be accessible to the public for AI model training, contributing to the open-source ecosystem.
2. Private Datasets: Organizations can collaborate on private datasets to train proprietary AI models, including foundational, fine-tuned, and custom models. OpenAI ensures that the sensitivity and access controls preferred by the partner are respected.
OpenAI encourages partners to help teach AI to comprehend diverse aspects of the world, aiming to advance AGI that benefits humanity collectively. The program invites organizations to join the effort and contribute to the future of AI development.
OpenAI has announced several significant updates, additions, and reduced pricing across its platform in a series of major announcements. The key highlights include the introduction of GPT-4 Turbo, a more capable and cost-effective model with a 128K context window. GPT-4 Turbo supports a substantial amount of text in a single prompt, with prices 3x cheaper for input tokens and 2x cheaper for output tokens than GPT-4. Developers can preview GPT-4 Turbo by using the gpt-4-1106-preview in the API.
The Function Calling feature, which allows the description of app functions or external APIs to models, has seen several improvements. GPT-4 Turbo now enables the calling of multiple functions in a single message, enhancing efficiency. The accuracy of function calls has also improved, with GPT-4 Turbo better at returning the suitable function parameters.
OpenAI has introduced a seed parameter that allows for reproducible outputs, making the model return consistent completions most of the time. This benefits debugging, unit tests, and achieving more control over model behavior. Log probabilities for the most likely output tokens generated by GPT-4 Turbo and GPT-3.5 Turbo will be available soon, which will be helpful for features like autocomplete in search experiences.
GPT-3.5 Turbo has been updated, supporting a 16K context window by default and offering improved instruction following, JSON mode, and parallel function calling. The new model presents a 38% improvement in format following tasks, such as generating JSON, XML, and YAML. The updated GPT-3.5 Turbo can be accessed by calling gpt-3.5-turbo-1106 in the API.
The Assistants API has been introduced, allowing developers to build agent-like experiences within their applications. Assistants have specific instructions, leverage extra knowledge, and can call models and tools to perform tasks. Persistent and infinitely long threads facilitate seamless interaction. The Assistants API includes features like Code Interpreter, Retrieval, and function calling.
In addition, OpenAI has rolled out new modalities in the API, such as GPT-4 Turbo with vision, DALL·E 3 integration for image generation, and Text-to-Speech (TTS) API for generating human-quality speech from text. The platform has introduced a copyright shield to defend customers facing legal claims around copyright infringement. Additionally, Whisper large-v3, the next version of the automatic speech recognition model (ASR), and the open-source Consistency Decoder have been released.