When we hear about the massive GPT-3 language model, it is natural to assume it is open source and ready for the taking. Assuming, of course, there is enough capacity to actually run it. However, there is a waitlist for GPT-3 access — and even for large companies it can span a year or more, spurring several efforts to work around OpenAI’s list.
For the AI hardware startups that are lucky enough, here is the best workaround we’ve seen to date — but it’s probably not the cheapest. Instead of waiting to deploy on your own systems, you can jump ahead and buy SambaNova’s Dataflow-as-a-Service appliances to run it at scale ahead of the rest. According to the company’s VP of Product, Marshall Choy, there are various configurations for GPT-3 specifically, but based on time to train requirements they can scale the infrastructure for GPT-3 training from one to 16 or more racks.
“The beta program is advancing nicely and a couple of tasks, text completion and search have now come out of beta” from OpenAI, says Choy. “Flexibility is also improving. [We can] soon make available a pre-training option in addition to the current inference-only offering. For those users looking for simple API access, GPT-3 is a great option.” He says SambaNova’s own hardware aims to provide low/no-code development options inclusive of API/SDK access in addition to GUI and CLI interface to perform both training and inference.
“SambaNova is bringing accessibility to the enterprise beyond OpenAI by achieving the highest levels of accuracy with domain-specific pre-training, without a waitlist,” he adds. SambaNova also announced the availability of the SambaNova GPT banking model, focused on high accuracy for banking and financial services tasks, which is available only as part of the SambaNova Dataflow-as-a-Service GPT package.
OpenAI’s GPT-3 language model uses deep learning to produce human-like text, but the long waitlist for GPT-3 limits its availability to a handful of organizations — and that’s even with Nvidia and Microsoft’s recent language announcement. Another company is bringing a competing GPT offering to the market and introducing a full suite of NLP for production and deployment of language models, making it possible to be up and running with a customized language model within one month as opposed to nine months or a year.
Choy says that making it to the top of the waitlist is only part of the challenge. Deploying large language models takes expertise and significant staffing resources. “Organizations have embarked on DIY efforts to build GPT-style solutions at high cost of deployment staff and infrastructure with high complexity. In addition to effectively establishing an AI practice, user have needed to select, tune & optimize, and finally maintain models over time. Before that they need to evaluate and select infrastructure and integrate with the aforementioned ML stack.”
Instead of integrating infrastructure and then training, tuning, deploying, and optimizing models using open source frameworks as the interface, Dataflow-as-a-Service interacts at the API level with the service, which thereby abstracts away all the complexities of infrastructure including the ML model itself. A public dataset pre-trained model is delivered ready for domain-specific corpus pre-training to generate a language model that is then fine-tuned for task-specific understanding, and finally deployed for use in domain-specific applications at the highest levels of accuracy.
Use cases for SambaNova’s Dataflow-as-a-Service GPT include:
- Sentiment analysis — Enable enterprises to save costs and time while implementing the most accurate sentiment analysis in scenarios such as customer support and feedback, brand monitoring and reputation management.
- Document classification — Enable solutions such as sorting articles or texts and routing them to relevant teams.
- Named Entity Recognition and relation extraction — Enable enterprises to save substantial time and money, and implement more accurate named-entity recognition applications in areas such as invoice automation, managing EHRs by identifying patient information and their prescriptions, and extracting import information from financial documents.
“Enterprises are insistent about exploring AI usage for text and language purposes, but up until now it hasn’t been accessible or easy to deploy at scale,” said Rodrigo Liang, CEO and co-founder of SambaNova. “By offering GPT models as a subscription service we are simplifying the process and broadening accessibility to the industry’s most advanced language models in a fraction of the time. We are arming businesses to compete with the early adopters of AI.”