Software-as-a-service (SaaS) company Zoho, which has been working on its own large-language models (LLMs), will be focussing on training its models based on domain-specific datasets such as those focused on finance, sales and marketing, customer relationship management, etc, the company’s AI research director Ramprakash Ramamoorthy said.
Since the company is building LLM models for its enterprise suite of products, it will first utilise the already built narrow models for training them in order to save on costs.
This means that unlike an LLM model which is being used by ChatGPT and Google Bard, Zoho’s LLM will be domain specific, catering primarily to their enterprise customers.
“Our AI models will be a combination of narrow models and LLMs. For LLMs, instead of training it on raw datasets, we will first utilise our narrow models to finetune them,” Ramamoorthy told FE in an interaction. “Leveraging internal repositories of data is not only crucial to save on costs of training these models, but also crucial in order to mitigate risks of AI biasness,” Ramamoorthy said.
Zoho, in June this year, announced that it is working to develop its own large-language model. The company has been currently leveraging generative AI across its stream of enterprise apps, through integration with ChatGPT. However, going forward the company plans to host everything in-house once its LLM model is developed.
With regard to its own generative AI, the company has an AI assistant product Zia that is being used by customers to manage customer records, producing documents, providing insights into processes, and forecasting.
In a bid to fasttrack the process to launch its own LLM, the company has started leveraging two types of graphics processing units (GPUs) – one is for training the data, and the other one is for inference, that is to make models generate predictions, make decisions, or produce outputs based on specific input data and contexts, Ramamoorthy added.
Zoho has over 100 million users across its over 55 business applications. The US is the company’s largest market in terms of customers, and India is within the top five markets.
When asked about whether integration with OpenAI platforms like ChatGPT will lead to privacy risks, Ramamoorthy said, “Open AI has a privacy policy that says, if you are a paying customer, your data will not be used to train, or make the model better. Further, we also have sufficient prompts that say that your data is leaving the Zoho ecosystem”.
With regard to the first step, the company will launch its AI LLM model trained on English language and then also work on coming up with similar models in other languages going forward, Ramamoorthy added.