Large models aren't the answer to every enterprise AI problem: Articul8 CEO

Arun Subramaniyan says Articul8's India heritage-AI venture aims to raise $30-50 million by March, and explains why enterprises need domain-specific models to control token costs and improve context

Articul8, an enterprise AI company spun out of Intel in 2024, says domain specific models are more effective than large language models for most companies. Founder and chief executive officer Arun Subramaniyan tells Avik Das that the company’s Indian venture, where it has partnered with Madras Sanskrit College to launch a Sanskrit heritage model and platform, is close to raising between $30 million-$50 million as it looks to operate independently. Excerpts:

 

Q: What are the top three challenges that you see when it comes to enterprise AI adoption? 

A: While AI adoption has increased individual productivity, multiple studies have shown that the organisational productivity has gone down because there's so much of AI generated content that is mostly junk. The other problem is token costs have exploded across the board. Every chief information officer (CIO) says their token costs have gone up five to six times in the last three months. This is because you don't to throw large models for everything, And the third is to know the context.

 

 

Q: So large models do not work effectively for everybody? 

A: A misconception is domain specific models are by design small. That's not necessarily true. The size of the model for us is dictated by two things. One is the size of the data set. And the other is the complexity of the task the enterprise is going after. For example, in pharmaceuticals, some of the domain-specific models that we are building are close to half a trillion parameters And the reason for that is they have data that is so voluminous for about 50 years that is so unique. So there you need a large model to assimilate all that data.

 

Q: Articul8 was looking to raise about $50 million for its India venture. When do you plan to close that round? 

A: We are looking to raise between $30-50 million as a seed funding for it and my expectation is given the conversations that are going on, we would close that by March from existing and new investors. We are setting up this heritage model as a separate entity for two reasons. The first is to have the intellectual property (IP) that is extracted from India to stay in India. And it is an effort that could go much broader than Articul8’s current mission of being domain specific. But even though it's a separate entity, they would collaborate with Articul8 in the early days to make sure we do the technology transfer and over time have its own management team to scale up.

 

Q: What are some of the key sectors you operate in? 

A: Some of those include manufacturing, large scale industrials, semiconductors, aerospace, automotive, and large scale energy systems with both production and transmission. Most of our revenue is from the US but our second largest market is Asia. The underlying common theme is all of them require deep expertise in the domain and they need not only like production scale software but they also have very low tolerance for failures. Our hypothesis is general purpose models are necessary. But the more powerful they become, they are not sufficient to get you to your final outcome in complex applications. Even the frontier firms acknowledge the need for domain-specific models to augment their general purpose models.