Generative AI models like ChatGPT have revolutionized content creation by producing highly convincing prose. This technology has tremendous potential for businesses seeking to unlock insights from the mountains of structured and unstructured data they collect.
The Challenge of Truth in ChatGPT for Enterprise
However, ChatGPT models are prone to "hallucinations" that can lead to misinformation, making it challenging for enterprise use cases where ground truth is crucial.
According to Marshall Choy, Senior VP of Product at SambaNova, generative AI has so far been limited to the augmentation stage, with reinforcement learning and human feedback in the loop. "The reality is, the higher the cost of a mistake, the less willing we are to fully automate," he said.
Solving the Challenge with Specialized Models
To address this challenge, SambaNova has trained a collection of open-source generative AI models, including GPT, Bloom, and StableDiffusion. These models are optimized for domain-specific enterprise use cases, whether in finance, legal, healthcare, or the public sector.
For instance, they can help analyze contact center interactions or understand text in large volumes of documents. However, to ensure maximum accuracy, generative AI models need to be specialized, and training them with domain-specific data is crucial.
According to Anton McGonnell, Senior Director of Product at SambaNova, the focus is on making models that are highly specialized to provide the ground truth required by enterprises. While specialization is essential, it may not always meet all enterprise needs for ground truth.
Ensuring Accuracy with Citations and Sources
SambaNova is currently developing a version of GPT that can cite its sources, allowing users to verify the accuracy of generated text. The model can link the generated text to the documents from which it was sourced, ensuring the user has confidence in the data obtained.
Choy emphasizes that ground truth is essential in enterprise applications, where citations are necessary to avoid plagiarism or infringement on copyrights and misinformation. To avoid the risk of making things up, generative AI models like ChatGPT should be programmed to indicate when they don't know the answer.
Addressing Security and Privacy Concerns
SambaNova is addressing enterprise concerns such as security and privacy by enabling customers to deploy its DataScale hardware systems behind their firewall and retain ownership of their data and model.
Conclusion
While generative AI models like ChatGPT have tremendous potential, they still require development to meet the ground truth requirements of enterprise applications. Specialization and training with domain-specific data are necessary, and models need to be credited to source their information accurately.
As with any new technology, there are concerns regarding security and privacy, which need to be addressed to enable businesses to make the most of generative AI.
