SambaNova, an AI chip startup that has amassed over $1.1 billion in venture capital funding to date, is targeting OpenAI and other competitors with a new generative AI product tailored for enterprise clients.
Today, SambaNova unveiled Samba-1, an AI-driven system engineered for tasks such as text rewriting, coding, language translation, and more. The company refers to the architecture as a “composition of experts” — a technical term for a collection of 56 generative open-source AI models.
Rodrigo Liang, the co-founder and CEO of SambaNova, asserts that Samba-1 enables businesses to fine-tune and cater to multiple AI use cases, bypassing the hurdles of ad hoc AI system implementation.
“Samba-1 is entirely modular, allowing companies to add new models asynchronously… without discarding their prior investment,” Liang shared in an interview with TechCrunch. “They’re also iterative, expandable, and easy to update, offering our clients the flexibility to adapt as new models are incorporated.”
Liang is an effective promoter, and his claims sound enticing. However, the question remains: Is Samba-1 truly superior to the numerous other AI systems for business tasks available in the market, not least of which are OpenAI’s models?
The suitability of Samba-1 largely depends on the specific use case.
The apparent primary advantage of Samba-1 is that it’s a compilation of models trained independently, rather than a single large model. This gives customers the ability to control how prompts and requests are directed. A request made to a large model like GPT-4 follows a single path — through GPT-4. However, a request made to Samba-1 can take one of 56 paths (to one of the 56 models that constitute Samba-1), depending on the rules and policies specified by the customer.
Liang asserts that this multi-model approach also diminishes the cost of fine-tuning on a customer’s data, as customers only need to focus on fine-tuning individual or small groups of models rather than a colossal model. Theoretically, this could lead to more reliable responses to prompts (for instance, less driven by hallucinations), as the answers from one model can be cross-checked with the answers from the others, albeit at the expense of additional computation.
“With this architecture, there’s no need to fragment larger tasks into smaller ones, allowing for the training of numerous smaller models,” Liang explained, adding that Samba-1 can be deployed either on-premises or in a hosted environment, depending on a customer’s requirements. “With a single large model, the computational cost per request is higher, thus increasing the cost of training. The architecture of Samba-1 reduces the cost of training.”
I would argue that numerous vendors, including OpenAI, offer competitive pricing for fine-tuning large generative models. Additionally, several startups, such as Martian and Credal, offer tools to direct prompts among third-party models based on either manually-programmed or automated rules.
However, what SambaNova is offering isn’t innovation in the traditional sense. Instead, it’s a comprehensive package — a full-stack solution that includes everything, even AI chips, needed to build AI applications. This might be more attractive to some enterprises compared to other options available.
“Samba-1 provides each enterprise with their own customized GPT model, ‘privatized’ based on their data and tailored to their organization’s needs,” Liang stated. “The models are trained on our customers’ private data, hosted on a single server rack, at a cost that is one-tenth of alternative solutions.”