This post first appeared in the Constellation Insight newsletter, which features bespoke content weekly.

Generative AI and large language models (LLMs) have received plenty of buzz, but enterprises need to stay focused on how domain-specific models develop. Why? That's where the returns will be.

At Constellation Research, we've kicked around small language models, safeguarding corporate data while tuning LLMs and how the real wins will be for specific use cases. Generative AI is nice for search and consumer use cases, but the magic happens when enterprises drive returns.

Fortunately, domain specific models are developing quickly. For instance, Google Cloud outlined Med-PaLM 2, a medically tuned LLM, that's aimed at healthcare. Google Cloud launched a bevy of AI tools at Google Cloud Next for CIOs to digest.

Bayer is using Google Cloud’s Vertex AI Search and exploring Med-PaLM 2 use cases. Hospital giant HCA is working with Google Cloud to leverage Med-PaLM to support doctors treating patients. Google Cloud fleshed out the HCA collaboration a bit more in a release, but the gist is that the two parties, along with Augmedix, is tuning models to support doctors and nurses. HCA is targeting:

  • Generative AI to improve patient handoffs between nurses.
  • Speech to text with ambient medical documentation.
  • Google AI to document key medical information more easily from conversations during patient visits.

According to Google Cloud, prompts were designed to guide the LLM toward topics such as medication changes, lab results, vital sign fluctuations and patient concerns. HCA is collecting nurse feedback to refine the tool. Ultimately, HCA wants to use Med-PaLM 2 LLM to support caregivers.

8 takeaways from Constellation Research's Healthcare Transformation Summit

My bet is that these domain-specific LLMs are going to be the real win for enterprises since they can leverage models and refine with their own data. These large LLMs such as PaLM, ChatGPT and LLama will have versions for various industries and use cases.

And this trend is going to go beyond healthcare. See:

Aneel Bhusri, CEO of Workday, said during the company's second quarter earnings call that it is using customer anonymized data to train LLMs. "We can then do domain-specific large language models, and those are smaller and less expensive. And we turn around and use those models to either make our products more competitive or they're the basis of new SKUs like the Skills Cloud," explained Bhusri. The real takeaway is that Workday isn't going to go add-on happy with charges. He said:

"I think you see us more in the mode of new SKUs like Skills Cloud rather than actually charging for any insight from the data -- that it's the customer's data. They allow us to use it in an anonymized way and we give them the results back. But I think what it allows us to do is train these large language models and then domain specific ones that will create new SKUs."

Intuit CEO Sasan Goodarzi was also bullish on domain-specific LLMs. Goodarzi said that Intuit "has incredibly rich longitudinal, transactional and behavioral data for 100 million customers."

Goodarzi added:

"For small businesses, we have a 360-degree view of their business and customers. We have 500,000 customers and financial attributes per small business on our platform and this data gives us insights into behaviors, income streams, expenses, profitability, and cash flows, enabling us to provide personalized experiences and recommendations to help them prosper.

Additionally, we have 60,000 financial and tax attributes per consumer on our platform. We are using our data to fine-tune our own financial large language models that specialize in solving tax, accounting, cash flow, marketing, and personal finance challenges."

Another key thought about domain specific LLM and AI use cases is that workloads will be spread around across multiple industries. 

Dell Technologies' Jeff Clarke, Chief Operating Officer, said the domain-specific use of LLMs and AI models will touch every industry. He said:

"What we think really happens on the enterprise level and in business is sort of the notion of domain-specific process-specific or field of study type of AI, where we actually use customers' data business will use their data they will tune the model and then run inference at site on edge, whether that be in a smart factory, smart hospital in a transportation network. So when you think about the vertical nature of this and how it will actually work in the real world, we think that technology makes its way all the way out to the edge, AI follows where the data is going to be created."