OpenAI Doubles Down on Agent Behavior and Hosts First Devday

OpenAI held its first Devday yesterday and it was amazing to be a part of. There was a keynote delivered and the Official Discord was jamming all day with thoughts, questions, use cases, and code. It was truly exciting to be a part of and see this released for the dev community. There were some announcements for ChatGPT as well and I will get to those later. Let's take a look at what the Devday had to offer.

The Assistant API

A new beta API has been released that while it is being called a precursor to agents, is agent-like behavior. I want to cover what this new API can do because it should be a real eye-opener as to where AI is heading. First though, the Assistant API can be broken down into these parts:

1. The Assistant Object
2. Threads
3. Messages
4. Runs

An assistant can be thought of as a chatbot on steroids I suppose. It is capable of running functions, retrieving content that is not in the model's training, and using tools. This means that you can upload files to an assistant to utilize across the board, upload files for it to use in threads via an individual message, and by the way, the assistant can create files for you too (think matplotlib possibilities here with Code Interpreter as a tool). This is known as retrieval (retrieval augmented generation - RAG).

Retrieval augments the Assistant with knowledge from outside its model, such as proprietary product information or documents provided by users. Once a file is uploaded and passed to the Assistant, OpenAI will automatically chunk your documents, index and store the embeddings, and implement vector search to retrieve relevant content to answer user queries.

Not only can you create Assistants via the API but you can create them in the OpenAI Playground and access them via the API (including threads and files). Similarly, Assistants you create via the API can be accessed in the Playground for the account owner as well. This is a nice at-a-glance feature that certainly helps with debugging if you are not retrieving run steps.

New Models on the Scene

I might as well just make a list here as it will be easier. I will give a few notes but this is not entirely comprehensive of what is possible.

• gpt-4-1106-preview (GPT-4 Turbo): The latest GPT-4 model with improved instruction following, JSON mode, reproducible outputs, parallel function calling, and more. Returns a maximum of 4,096 output tokens. This preview model is not yet suited for production traffic.
• gpt-4-vision-preview (GPT-4 Turbo with vision): Ability to understand images, in addition to all other GPT-4 Turbo capabilities. Returns a maximum of 4,096 output tokens. This is a preview model version and is not suited yet for production traffic.

Both of the above models benefit from a 128,000 token context window! Boom.

• gpt-3.5-turbo-1106 (Updated GPT 3.5 Turbo): The latest GPT-3.5 Turbo model with improved instruction following, JSON mode, reproducible outputs, parallel function calling, and more. Returns a maximum of 4,096 output tokens.
• dall-e-3: The latest DALL·E model released in Nov 2023

DALL·E 3 was almost glossed over but having this powerful image generation model available as an API is something that I know a lot of people were waiting on.

• tts-1 (Text-to-speech 1): The latest text-to-speech model, optimized for speed.
• tts-1-hd (Text-to-speech 1 HD): The latest text-to-speech model, optimized for quality.

Model pricing was also a topic of discussion yesterday and something the community appreciates. The pricing for GPT-4-Turbo (@ 128K token context) is:

Input: $0.01 / 1K tokens

Output: $0.03 / 1K tokens

Custom GPTs for ChatGPT

That's right! OpenAI announced and is slowly rolling out features where users can create (and sell) custom versions of ChatGPT, which are AI chatbot agents tailored for specific use cases. They allow users to create specialized AI applications without the need for coding, offering a personalized experience that can integrate with various services and data sources and is accessible through a dedicated GPT Store for subscribers and enterprise clients. These custom GPTs can perform a range of tasks, from content creation to event navigation, with the added benefits of web browsing, integration with DALL-E, and code interpretation capabilities.

Keep an eye out for the GPT Store, where these AI agents will be available. While specifics are under wraps, creators will earn based on usage. It was announced yesterday that there was a revenue share model that would be put in place. Trust me, if the Discord is any sort of signal, there will be a ton of GPTs available.

One word of caution would be that while OpenAI has announced security measures to be put in place, these GPTs should still be treated like ChatGPT. If you want to do some heavy lifting with your data, please do not toss unobfuscated data straight into ChatGPT. You have options though.

1. Hire someone to build a generative AI-powered app for you (likely in the form of an Agent/Assistant).
2. Purchase ChatGPT Enterprise.

Also, they announced that ChatGPT now has updated information up to April 2023!

Closing Thoughts

Yesterday Sam Altman said this in yesterday's Devday keynote

Eventually, you will just ask a computer for what you need and it will do all of these tasks for you.

While AGI is certainly in mind and potentially at play in that reality, AI Agents are the working actors behind this fruitful and time-saving scenario. I imagine a world where humans serve the role of conductor and individual agents work alone or together to complete tasks. Actually, I do not think we are super far away from that scenario today.

#aiassistants #assistants #openai #devday

------- ?? -------

Hey, I'm Dave! I'm a former digital agency owner, now the founder of Bold Crow AI. I help businesses and organizations implement customized AI solutions in a responsible way. I've built cool tools for nonprofits too, helping them build community and gather and leverage social proof.

Let's connect:?Dave Norris