EN
Generative AI will be the defining technology of this century, fundamentally reshaping how businesses, governments, and societies operate. It is streamlining processes to make people more efficient, reduce costs, improve operations, and more. This technology is enabling us to better fight climate change, make scientific breakthroughs, cure diseases, and reduce poverty, but like every significant technology its capacity for improving society is dependent upon how it is applied.
At SambaNova, we believe that generative AI must be applied fairly and transparently, while ensuring security and data privacy. Samba-1, the first trillion parameter model built for the enterprise and government organizations, was designed to ensure that these standards of Responsible AI are always applied.
Fairness and bias mitigation
Any generative AI model is only as good as the data that it was trained on. SambaNova Samba-1 delivers the first trillion parameter model that can be quickly, easily, and affordably fine tuned on customer data.
SambaNova uses the latest open source models to build Samba-1, so the training data and model weights are always known. This provides our customers with maximum transparency as well as helping them meet regulatory compliance standards.
In contrast, large monolithic models are typically trained on general internet data, but exactly what that data is remains unknown to the organizations that use it. Since these models are trained on general data from the internet, bias can be introduced. Beyond that, both the training data and the model weights are part of the IP of the vendor that licenses it. They will not expose this information to their customers, effectively making their solution a “black box” that is accessed through an API.
A Composition of Experts Model
Samba-1 is able to achieve trillion parameter scale by combining multiple smaller expert models into a single instance that operates as one large model. This is a fundamentally different architecture than the monolithic design used with other trillion parameter models. If those models have a built in bias, then the only way to remove it is to retrain the model with new data, which is a costly and complex process that may not be practical.
With Samba-1, if a model is determined to have bias, the design of the SambaNova system enables the model to be replaced with a different model, quickly and easily.
Security and Data Privacy
SambaNova has always prioritized security and data privacy for our customers. SambaNova platforms are available as fully integrated on-premises systems which enables organizations to install these systems behind their firewall and in their datacenter. For customers that want to use an as-a-Service model, this is also available, with systems dedicated to customers, so data is never shared.
Further, once a model has been fine tuned with a customer's data, that model becomes the sole property of that customer in perpetuity. This ensures that the customer is always in control of their data for absolute privacy and data sovereignty.
SambaNova remains committed to providing customers with responsible, generative AI that is safe, secure, and transparent, and free from bias. Learn more about what SambaNova Suite can do for your organization.
