Blog

In the heart of San Francisco, OpenAI's inaugural developer conference, DevDay, marked a significant milestone in the AI industry. As a landmark event, it brought together the brightest minds in AI, offering a platform for keynotes, live demos, and in-depth sessions. This conference not only showcased OpenAI's latest advancements but also set the stage for the future trajectory of AI applications in various domains.

GPT-4 Turbo with 128k Tokens

OpenAI's DevDay conference introduced GPT-4 Turbo, a significant enhancement over its predecessor, GPT-4. This new model represents a leap in AI capabilities, providing developers and businesses with a more efficient and cost-effective AI solution. 

Here are the key features and advancements of GPT-4 Turbo:

Enhanced Capabilities and Knowledge Base

• Extended Context Window: GPT-4 Turbo boasts a 128,000 token context length, translating to the ability to process the equivalent of over 300 pages of text in a single prompt. This vast improvement enables the model to handle larger and more complex datasets with ease​​​​​​.
• Updated Knowledge: The model incorporates world knowledge up to April 2023, offering more contemporary and relevant insights compared to GPT-4's September 2021 cutoff​​​​​​.

Developer-Friendly Features

• Function Calling Updates: GPT-4 Turbo introduces improvements in function calling. It allows for the description of functions within apps or external APIs, enabling the model to output JSON objects containing arguments to call these functions. This feature now supports multiple function calls in a single message, enhancing the model's efficiency and accuracy​​.
• Improved Instruction Following and JSON Mode: The model excels in tasks requiring the careful following of instructions and supports a new JSON mode. This mode ensures the generation of valid JSON responses, useful for developers creating JSON outside of function-calling scenarios​​.
• Reproducible Outputs and Log Probabilities: A new 'seed' parameter allows for reproducible outputs, aiding in debugging and unit testing by providing consistent completions. Additionally, the feature to return log probabilities for the most likely output tokens is being introduced, which can be utilized in applications such as autocomplete search features​​.

Assistants API for Enhanced AI Interactions 

‍

‍

OpenAI's Assistants API, announced at their developer conference, represents a significant stride towards empowering developers to create sophisticated, agent-like AI experiences within their applications. This innovative API offers a suite of tools and capabilities designed to enhance the functionality and interactivity of AI assistants.

Here are the key features and capabilities of the Assistants API:

• Agent-Like Experiences: The API enables the building of AI agents with specific instructions, allowing for a more purpose-driven and tailored AI experience. These agents can leverage extra knowledge and call on models and tools to perform various tasks, effectively handling complex, multi-step processes​​​​​​.
• Code Interpreter: A crucial component of the API, the Code Interpreter, allows for the writing and execution of Python code in a secure, sandboxed environment. This feature enables AI assistants to iteratively run code, solve problems in coding and mathematics, and even generate graphs and charts​​.

• Retrieval Component: The API includes a retrieval component that enhances the knowledge base of AI assistants with external information. This feature is particularly useful for incorporating product information or company-specific documents, allowing for a more informed and relevant AI response​​​​.
• Function Calling: Assistants API supports function calling, enabling AI agents to invoke specific programming functions defined by the developer. This capability allows for a more dynamic and responsive AI experience, as the assistant can incorporate responses from these functions into its messages​​​​.
• Use Cases and Applications: The Assistants API opens up a wide range of potential applications, from natural language-based data analysis apps to coding assistants and AI-powered vacation planners. Its versatility makes it suitable for various domains, enhancing both consumer and enterprise AI applications​​​​.

Availability and Pricing

The Assistants API is currently in beta and available to all developers. Usage of the API is billed based on the chosen model's per-token rates, with tokens representing parts of raw text processed by the API​​.

Custom GPTs and the GPT Store

OpenAI's introduction of Custom GPTs and the GPT Store marks a significant evolution in AI personalization and accessibility. These features allow users to tailor ChatGPT to specific needs and share these custom versions with a broader community. 

Customizing ChatGPT for Specific Purposes

Users can create Custom GPTs (GPTs) without any coding requirements. These GPTs can be tailored for individual use, internal company use, or public sharing, catering to a variety of needs and contexts​​.

GPTs also empower users to customize ChatGPT for specific tasks, incorporating their preferences and requirements. This feature extends the utility of ChatGPT beyond general use, adapting it to specific scenarios like education, work, or personal projects​​​​.

Enterprise customers can create internal-only GPTs tailored to specific business use cases, departments, or proprietary datasets. These custom GPTs can support various functions such as marketing, customer support, or employee onboarding​​.

The GPT Store: A Marketplace for AI Tools

The GPT Store will allow users to share their custom GPTs with a wider audience. It will feature creations by verified builders, enabling users to search, access, and potentially monetize their GPTs based on user engagement​​.

The store will spotlight GPTs in various categories such as productivity, education, and entertainment. It will also feature leaderboards to showcase popular or effective GPTs​​.

Privacy and Safety Considerations

OpenAI ensures that chats with GPTs are private and not shared with builders. Users have control over whether their data is sent to third-party APIs used by GPTs. OpenAI has also implemented systems to review GPTs against usage policies, aiming to prevent the sharing of harmful content​​.

Integration with External APIs

Developers can enhance GPTs by integrating them with external APIs, allowing them to connect to databases, email systems, or facilitate e-commerce orders. This feature significantly broadens the scope of GPTs, making them more versatile and practical for real-world applications​​.

Advancements in API Functionality

‍

GPT-4 Turbo with Vision

GPT-4 Turbo is an enhanced version of GPT-4 that is not just an advanced language model,but it integrates visual perception, enabling AI to process and interpret both text and images. This integration opens up new possibilities for AI applications. For instance, in content creation, GPT-4 Turbo with Vision can provide more contextually rich and relevant outputs by understanding both textual descriptions and visual cues. In educational tools, it can offer more immersive learning experiences by combining textual information with relevant visual aids.

The enhancement in user interfaces can lead to more intuitive and interactive systems, where AI can respond to visual inputs alongside textual queries. This integration marks a significant step towards more holistic and versatile AI applications, blending the realms of visual and textual understanding in a single AI model.

DALL-E 3

DALL-E 3, another highlight from OpenAI's DevDay 2023, takes image generation to new heights. Building on its predecessors, DALL-E 3 offers more nuanced and contextually aware visual outputs, enabling users to create highly detailed and specific images based on textual descriptions. The implications of this technology are vast, especially in creative and commercial domains. For artists and designers, DALL-E 3 allows them to generate unique visuals that can complement or inspire their work. In the commercial sphere, businesses can leverage this technology to create bespoke marketing materials, product designs, and other visual content with unprecedented ease and efficiency. DALL-E 3 offers tools that can transform imagination into visual reality with a few text prompts.

Text-to-Speech

The advancements in Text-to-Speech (TTS) technology showcased at DevDay 2023 mark a substantial improvement in AI-generated audio. These new TTS capabilities bring naturalness and realism to AI-generated voices that were previously unattainable. This technology has profound implications for various sectors. 

In terms of accessibility, advanced TTS can provide more engaging and easier-to-understand audio for visually impaired users, making digital content more inclusive. In the entertainment industry, realistic voice generation can enhance the user experience in video games, audiobooks, and virtual assistants, offering more immersive and interactive engagements. The improved TTS technology makes AI interactions more human-like and relatable. This advancement in TTS technology represents a significant stride towards bridging the gap between AI and human-like communication, enhancing the way we interact with technology using voice. 

Lower Prices and Higher Rate Limits

‍

Lower Prices Across the Platform

GPT-4 Turbo: The pricing for GPT-4 Turbo has seen a significant reduction, making it more accessible for developers. The cost of input tokens is now 3 times cheaper than the standard GPT-4, priced at $0.01, while output tokens are 2 times cheaper at $0.03.

GPT-3.5 Turbo: For GPT-3.5 Turbo, input tokens are now 3 times cheaper than the previous 16K model, priced at $0.001, and output tokens are 2 times cheaper at $0.002. Developers who were using the GPT-3.5 Turbo 4K model benefit from a 33% reduction in input token prices. These lower prices are exclusive to the new GPT-3.5 Turbo model introduced at the conference.

Fine-Tuned GPT-3.5 Turbo 4K Model: The fine-tuned version of the GPT-3.5 Turbo 4K model now offers input tokens at a 4 times reduced rate of $0.003 and output tokens at 2.7 times cheaper, priced at $0.006. Importantly, fine-tuning now supports a 16K context at the same price as the 4K with the new GPT-3.5 Turbo model. These new prices are also applicable to fine-tuned gpt-3.5-turbo-0613 models.

Higher Rate Limits for Enhanced Scalability

Doubling Tokens Per Minute for GPT-4 Customers: In a move to support the scalability of applications, the tokens per minute limit for all paying GPT-4 customers have been doubled. This change is designed to accommodate larger-scale operations and more intensive usage.

Transparent Usage Tiers and Rate Limits Increases: OpenAI has now published usage tiers that determine automatic rate limits increases. This transparency allows developers to understand how their usage limits will scale automatically.

Facilitating Requests for Increases in Usage Limits: Developers can now easily request increases to their usage limits directly from their account settings, providing more flexibility and control over their application's operational capacity.

Copyright Shield

A new initiative by OpenAI aims at providing legal protection to users of the OpenAI platform. This feature is specifically designed for users of ChatGPT Enterprise and the developer platform. 

Copyright Shield is a response to the complex legalities associated with AI-created content. OpenAI commits to defend and cover the costs for its customers if they face legal claims of copyright infringement related to the use of OpenAI's generally available features. By providing this layer of security, OpenAI aims to lower the barrier to entry for developers and businesses interested in exploring AI. This move is seen as a significant step towards fostering innovation and creativity in AI applications.

Whisper v3 and Consistency Decoder

Whisper v3 represents the latest evolution in OpenAI's advanced speech recognition technology. Building upon its predecessors, Whisper v3 offers enhanced accuracy and efficiency in transcribing spoken words into text. This version is particularly adept at handling various accents and dialects, making it a more inclusive tool for global users. The technology's improved noise reduction capabilities ensure clearer transcription even in less-than-ideal audio conditions. Whisper v3's advancements are significant for applications ranging from real-time transcription services to voice-controlled interfaces, providing a more reliable and accessible speech-to-text solution.

The Consistency Decoder is another noteworthy development, primarily focused on improving the coherence and relevance of AI-generated text. It functions by ensuring that the responses generated by AI models like GPT-4 are not only contextually appropriate but also maintain consistency throughout longer conversations or content pieces. This advancement is particularly crucial in applications like customer service chatbots, educational tools, and content creation platforms, where sustained coherence is essential. The Consistency Decoder enhances the user experience by providing more reliable and logically consistent interactions with AI, fostering a more natural and engaging dialogue between humans and AI systems.

Conclusion 

OpenAI's DevDay 2023 highlighted significant advancements in AI, with the introduction of GPT-4 Turbo featuring a 128k token limit and updated knowledge. The Assistants API was a key development, offering tools for sophisticated AI interactions, including a Code Interpreter and function calling capabilities. Custom GPTs and the GPT Store marked progress in AI personalization, allowing users to tailor and share AI models. Notably, GPT-4 Turbo now includes visual perception and DALL-E 3 advanced image generation. Text-to-Speech technology also saw improvements, enhancing realism in AI voices. Additionally, OpenAI introduced more accessible pricing models and higher rate limits, alongside a Copyright Shield initiative for legal protection. Whisper v3 and the Consistency Decoder represented leaps in speech recognition and text consistency, underlining OpenAI's commitment to evolving AI technology.