1Mby1M Virtual Accelerator AI Investor Forum: With Yash Hemaraj, General Partner at BGV (Part 2)

Sramana Mitra: Yash, I agree with you on workflow automation and very precise domain-specific AI being a much better opportunity to build businesses and do startups in than boiling the ocean and trying to do something much broader. In general, in startups, I don’t think broad brushed approaches work very well. They just burn a lot of capital. They don’t really produce very effective companies.

There are a few things that I want to pursue in this context and then I want to do some examples of what you have invested in touching upon those aspects. Before generative AI came into play, people were looking for data specific to a particular workflow, a particular domain, or a particular problem solving. We were not looking at very large data repositories.

With generative AI, the concept of big data has expanded significantly, and now we’re dealing with much larger, more general data repositories. This introduces challenges like hallucination in large language models (LLMs), which didn’t exist in the pre-generative AI era. Back then, machine learning models worked with big data but didn’t hallucinate because they weren’t based on language models.

As we discuss examples from your portfolio, I’d like to explore two key points. First, how do LLMs, which are prone to hallucination, compare to earlier machine learning models? Second, do smaller, domain-specific language models with constrained vocabularies and workflows have advantages in being more precise and less error-prone? I’d love to see examples from your portfolio that illustrate these concepts.

Yash Hemaraj: That’s a constant debate we’re having. We’ve formed a Human-AI Alliance, a collaboration of experts from various industries, including enterprise adopters, startup founders, universities, and investors, to craft this thesis. This helps us understand where AI is over-invested, where opportunities lie, and where new AI-native businesses are emerging.

On your question about large versus smaller language models, it really depends on the specific business problem. In an enterprise context, there will be place for LLMs as well as some domain-specific models. It depends on where in the workflow that you have to use.

For example, despite the billions of dollars invested in large language models, their capabilities are multiplying rapidly. We are already discussing GPT-5, and the projected capabilities of these models are mind-blowing. It would be a missed opportunity for many enterprises not to leverage these foundational models. Over time, we believe that given the level of investment required and the substantial amounts already committed, it will become increasingly difficult for smaller venture funds to invest in this space.

LLMs, despite their growing capabilities, may become commoditized and serve more like utilities, providing a foundation for other applications. However, challenges like hallucination need to be addressed, especially in enterprise settings. For example, we’ve invested in Encrypt AI, a company that provides hallucination management, ensuring models are checked for bias, hallucinations, profanity, and other risks. While this may be acceptable in a consumer context, in an enterprise context, it can damage your brand, so it’s crucial to have mechanisms in place. Encrypt AI helps ensure proper governance by evaluating prompts and outcomes and testing whether the model is functioning as intended.

We’re also big believers in the governance, risk, and compliance (GRC) domain. As more regulations come into play, companies need to comply with local laws, and startups are actively addressing these needs by building enterprise-grade applications. When it comes to LLMs, there are still valuable investment opportunities in smaller, domain-specific models or in areas like GRC and agentic software. We liken this generative AI movement to the evolution of browsers. Just as browsers enabled a multitude of enterprise-grade capabilities, generative AI offers the opportunity to build fully autonomous AI agents. These agents require identity control, governance mechanisms, and role-based access control—all essential capabilities being developed today.

On the flip side, there’s tremendous opportunity in smaller, domain-specific models that can offer more controlled, precise, and less error-prone solutions. For example, in our portfolio, we have LYZER.ai, an India-US cross-border company helping enterprises build agentic software. They provide pre-built AI agents and an end-to-end solution for enterprises to adopt these capabilities on day one, integrating tools like GRC systems and identity management.

This segment is part 2 in the series : 1Mby1M Virtual Accelerator AI Investor Forum: With Yash Hemaraj, General Partner at BGV
1 2 3 4 5 6