OpenAI & Microsoft's HUGE AI Announcements (Dev Day Supercut)

Summary

OpenAI's Dev Day event highlighted several groundbreaking advancements, including the launch of GPT-4 Turbo, with a massive context length of 128K tokens, better accuracy, and new features like JSON mode and reproducible outputs. The event introduced new capabilities such as vision, voice, and text-to-speech, enhancing AI's multimodal abilities. Custom models and the new Assistance API allow developers to create more specialized, intelligent applications. With a focus on AI's transformative potential, OpenAI aims to empower users and developers through cost-effective, customizable solutions, deepening their partnership with Microsoft to push the frontiers of AGI.

Takeaways

• 😀 OpenAI launched GPT-4 as a research preview a year ago, followed by GPT-4 and major updates including voice, vision, and image capabilities.
• 😀 GPT-4 Turbo is the new model introduced with significant improvements, including a massive increase in context length (up to 128,000 tokens), better accuracy over long contexts, and more control for developers.
• 😀 Developers can now use new features like Json mode for valid JSON responses, improved function calling, and reproducible outputs for consistent results.
• 😀 Retrieval capability has been added, allowing models to pull knowledge from external documents and databases, with the knowledge cutoff updated to April 2023.
• 😀 The new GPT-4 Turbo model includes improved vision and text-to-speech features, enabling image inputs and natural-sounding audio outputs for a wide range of use cases.
• 😀 The Whisper V3, the next version of OpenAI's open-source speech recognition model, is being released with better performance across many languages.
• 😀 Custom models program has been launched, allowing companies to build tailored models with proprietary data and support from OpenAI's research team, though initially at a higher cost.
• 😀 OpenAI has increased rate limits for GPT-4 customers, offering higher tokens per minute and introducing a copyright shield to help defend against legal claims related to copyright infringement.
• 😀 GPT-4 Turbo offers lower pricing than GPT-4, being three times cheaper for prompt tokens and twice as cheap for completion tokens, making it more cost-effective for developers.
• 😀 A new GPT store will be launched where users can list their custom-built GPTs, with revenue sharing for the most popular and useful GPTs, promoting a vibrant ecosystem of tailored AI models.

Q & A

What significant milestone did OpenAI reach before the OpenAI Dev Day?

Before the OpenAI Dev Day, OpenAI reached the milestone of launching ChatGPT in November 2022 as a research preview, which went well, and followed it up with the launch of GPT-4, the most capable AI model at the time.

What is the main improvement in GPT-4 Turbo compared to previous versions?

GPT-4 Turbo features a significant improvement in context length, supporting up to 128,000 tokens, which is 16 times longer than GPT-4's 8K context. It also delivers better accuracy over long contexts, making it more versatile.

How does the new Json mode in GPT-4 Turbo enhance developer control?

Json mode ensures that GPT-4 Turbo responds with valid JSON, which makes the model more predictable and reliable for developers working with structured data, offering a higher degree of control.

What update has been made to the model's knowledge cut-off?

The knowledge cut-off for GPT-4 Turbo has been updated to April 2023, addressing the previous frustration that the knowledge of GPT-4 only went up to 2021.

What new capabilities are being added to the GPT API in terms of modalities?

The GPT API now supports image inputs, allowing for tasks like captioning, classification, and analysis. Additionally, a new text-to-speech model has been introduced for generating natural-sounding audio, along with an updated version of OpenAI’s speech recognition model, Whisper V3.

What are the key features of the Custom Models program launched by OpenAI?

The Custom Models program allows researchers to collaborate closely with companies to create specialized models using proprietary data and custom knowledge domains. Although initially expensive, it offers tailored AI models for specific use cases.

How is OpenAI addressing the issue of pricing for GPT-4 Turbo?

GPT-4 Turbo is considerably cheaper than GPT-4, with prompt tokens priced at 1 cent per 1,000 tokens and completion tokens at 3 cents per 1,000 tokens. This new pricing model makes GPT-4 Turbo more affordable, being more than 2.75 times cheaper than GPT-4 for most customers.

What role does Microsoft play in OpenAI's infrastructure?

Microsoft has been instrumental in supporting OpenAI’s infrastructure by helping build the system that powers the models. This includes collaboration on hardware like data centers, racks, accelerators, and networks to ensure optimal performance for OpenAI’s models.

What new feature in ChatGPT is aimed at improving the user experience?

A new feature in ChatGPT eliminates the model picker dropdown menu. ChatGPT will now automatically determine the appropriate model to use based on the user's needs, streamlining the user experience.

What are 'GPTs' and how can they be used?

'GPTs' are tailored versions of ChatGPT designed for specific tasks. Developers can customize a GPT with specialized instructions, knowledge, and actions. These GPTs can be published for others to use, unlocking various new use cases.