Today, more than 40 thousand enthusiasts joined OpenAI in their very fist developer conference: DevDay. It was packed with more than a dozen new and exciting announcements. Here's a summary of the most relevant ones if you missed it!
New Model: GPT4 Turbo
As of today, a new model named GPT4 Turbo is available for users.
It has an huge 128k token context! This makes it officially the LLM with the largest context commercially available today. As a reference, plain GPT4 had a context of 16k or 32k tokens.
It's knowledge cutoff is April 2023. OpenAI says they will commit to having more recent cutoffs from now on.
It's supposedly smarter and more accurate over time.
More Model Control
There's a new JSON Mode, that ensures that GPT answers with a valid JSON.
There's a new seed parameter, that allows you to get reproducible outputs.
Log probs will be included along with the outputs. Ideal for recommender systems, for example.
You'll be able to bring your own documents to the knowledge base.
Better function calling, allowing multiple calls simultaneously.
Models will be generally better at following instructions
API Improvements
DALL-E will be available through the API
Turbo with Vision will be available through the API
Text-to-speech will be available through the API
A new Whisper v3 was announced.
Pricing and Rates
The API rates for GPT4 were doubled!
GPT4 Turbo is 3x cheaper on input tokens and 2x cheaper out output tokens. As a reference: 1000 input tokens will cost 1 cent, and 1000 output tokens will cost 3 cents.
GPT-3.5 Turbo 16k had the same price reduction. As a reference: 1000 input tokens will cost 0.003 cents, and 1000 output tokens will cost 0.004 cents.
Fine tuning will also be cheaper.
OpenAI prioritized price over speed, so while GPT4 turbo is cheaper, its probably slower.
ChatGPT
Will start using GPT4 Turbo!
No more annoying model picker, it will have all modalities enabled by default.
Build your own ChatGPT using GPTs: a new service that allows you to create new chatbots using knowledge from your documents, connect to services and make them public, private or restricted. Built using natural language!
Assistant API
New API that eases building assistive experience using OpenAI LLMs.
Will handle threading automatically so assistants are capable of keeping multiple conversation threads.
Assistants will be capable of:
Retrieving information from your documents.
Invoking the code interpreter
Calling functions accordingly
Again, you build them with natural language!
Other Announcements
OpenAI stated that they will provide legal help for companies using their services that are currently being accused of copyright infringements.
OpenAI and Microsoft strengthen their partnership to make AI more accessible to everyone.
OpenAI will work with a small set of companies to create custom models for their specific needs and knowledge base.
Every attendee got a $500 coupon to use the OpenAI platform!
Undoubtedly, a great day for OpenAI and developers in general. We are excited to start building and see what innovative products are accelerated by these announcements.
What are you planning on creating?