Join our daily and weekly newsletters to receive the latest updates and exclusive content on industry-leading AI coverage. Learn more
Transformers are a cornerstone of the newfangled era of generative AI, but they’re not the only way to build a model.
AI21 is available today with recent versions of its Jamba model, which combines transformers with a Structured State Space Model (SSM) approach. The recent Jamba 1.5 mini and Jamba 1.5 enormous build on the initial innovations the company introduced with the release of Jamba 1.0 in March. Jamba uses an SSM approach known as Mamba. The goal of Jamba is to combine the best features of transformers and SSM. The name Jamba is actually an acronym that stands for Joint Attention and Mamba (Jamba) architecture. The promise of the combined SSM Transformer architecture is better performance and accuracy than either approach can provide alone.
“We got incredible feedback from the community because it was essentially the first and still is one of the few Mamba-based production models that we got,” Or Dagan, VP of product at AI21, told VentureBeat. “It’s a novel architecture that I think started some debates about the future of architecture in LLM and whether transformers are here to stay or whether we need something different.”
Both Jamba 1.5 models are available under an open license. AI21 also provides commercial support and services for the models. The company also has partnerships with AWS, Google Cloud, Microsoft Azure, Snowflake, Databricks, and Nvidia.
What’s Modern in Jamba 1.5 and How It Will Speed Up Agent AI
Jamba 1.5 Mini and Huge introduce a range of recent features designed to meet the evolving needs of AI developers:
- JSON mode for handling structured data
- Quotes for increased responsibility
- Document API for better context management
- Function calling capabilities
According to Dagan, these additions are particularly crucial for developers working on agent-based AI systems. Developers commonly exploit JSON (JavaScript Object Notation) to access and create application workflows.
Dagan explained that adding JSON support allows developers to more easily build structured input/output relationships between different parts of a workflow. He noted that JSON support is key for more complicated AI systems that go beyond just using a language model. The citation feature, on the other hand, works in conjunction with the recent document API.
“We can teach the model that when you generate something and you have documents in your input, you need to assign the appropriate parts to those documents,” Dagan said.
How Citation Mode differs from RAG in providing an integrated approach to agent-based AI
Users should not confuse citation mode with Retrieval Augmented Generation (RAG) technology, although both approaches base answers on data, which can improve accuracy.
Dagan explained that Jamba 1.5’s citation mode was designed to work with the model document API, providing a more integrated approach compared to established RAG workflows. In a typical RAG setup, developers would connect a language model to a vector database to access the relevant documents for a given query or task. The model would then need to learn how to efficiently incorporate this retrieved information into its generation.
In contrast, the citation mode in Jamba 1.5 is more tightly integrated with the model itself. This means that the model is trained not only to retrieve and include relevant documents, but also to explicitly cite the sources of information it uses in its output. This provides greater transparency and traceability compared to the established LLM workflow, where the model’s reasoning can be more unrecognized.
AI21 also supports RAG. Dagan noted that his company offers its own end-to-end RAG solution as a managed service, which includes document search, indexing and other required components.
Looking ahead, Dagan said AI21 will continue to work on improving its models to meet customer needs. It will also continue to focus on enabling agent-based AI.
“We also understand that we need to push the boundaries of agent-based AI systems and how planning and execution are handled in that domain,” Dagan said.