Introducing Space
At Reka, we develop our models with the beliefs that AI models should understand the multimodal world we live in and perform actions to accomplish tasks for us. As our models become more capable, we continue to explore better ways for everyone to benefit from them. We think the current approach to interact with these AI models based on turn-based messaging-style interface is more suited for text-based chatbots.
We are excited to introduce Space, a new interface that redefines how users interact with our multimodal models. Our goal is to empower AI models to reason over diverse multimodal inputs, remember and learn from past interactions, organize background knowledge, and perform actions on your behalf.
Space is powered by Reka Flash, our 21 billion parameters model. It is still in its early stage. We make it available to you in private beta today in order to gather your feedback, so that we can improve it further. Space is free to use.
We focus on adding a few key features for the beta version of Space:
Board/canvas layout: Space starts as an empty space that allows you to add, remove, and edit elements as needed.
File organization: Space provides an easy and intuitive way for you to upload and organize your files. Your files can now be interleaved, and provided as an input to our model to extract valuable information or get insights. Our intention is to equip our model with appropriate background knowledge that can be accessed as selectable memory, providing you with precise results, without relying on the model to memorize and recall the information on the fly.
Interaction style: Unlike traditional turn-based chats that are linear in nature, you can work with any part of AI outputs by highlighting, editing, or quoting them in Space.
Collaboration: Space makes collaboration possible. You can design together or share existing space with your friends and colleagues.
We think Space can be useful for learning new subjects, coordinating projects, organizing resources, and many more.
On the research side, one of our objectives is to train a model that learns from demonstrations and interactions (with humans, itself, and other models) to accomplish tasks for you. We believe that next-generation AI models are proactive, adaptive, and self-improving (we plan to share more about our technical progress soon).
Space is a starting point that allows us to evaluate our model in a consistent graphical user interface with fixed tools and actions. This will allow our model to adapt to your workflow more quickly. With your consent, our model will learn from your interactions, actions and content organization, among others, and tailor its performance to better meet your needs.
This beta launch is only the beginning of Space. We have several more Space features that we are excited about currently under development. We think improvements in model capabilities should be contextualized in the form with which we interact with these AI models.
If you are excited about human-AI interactions and improving Space, our Product team is hiring! We are also hiring across all roles including research, applied, and business. You can check them out and apply here!
Space is currently in private beta. Read more about what you can do with Space and sign up to our waitlist here.