1Mby1M Virtual Accelerator AI Investor Forum: With Rajeev Madhavan, Clear Ventures (Part 2)

Sramana Mitra: Absolutely. So let me synthesize a little bit. The semiconductor layer is very expensive to do startups in. There’s quite a bit of activity going on there, but for our audience, where our philosophy is more lean startup, bootstrapping kind of activities, this is not really possible to do as it is capital intensive. We work in the capital-efficient domain.

Secondly, the layer above that is the platform layer where it’s also very expensive to build these platforms. Most of us agree on this thesis that the large cloud providers who were previously providing cloud platform as a service are now providing AI cloud platform as a service. Azure, OpenAI, Google, IBM, and AWS are all providing cloud platform as a service. There are some startups in that field as well. So you are seeing players like Cohere and Anthropic, etc.

Then comes the application layer, where the vertical AI activity is. This is really the richest, most highly active area of entrepreneurship today. Most of it germinates out of people’s intense domain knowledge. You do need intense domain knowledge to be able to build the value on top of the lower layers where the horizontal AI exists and then build highly contextual solutions.

Now, I’m going to ask you a question, Rajeev, in this context.

The companies that you’re investing in, what is the distribution or what is your analysis in generative AI companies or where the empowering technology is generative AI versus just regular AI, machine learning, deep learning, etc.?

Rajeev Madhavan: Generative AI is great in things like writing reports and assessing data, but if you’ve very specific data, you really need to build your own private LLM in conjunction with a tool. For example, if I need to find the address of somebody from an insurance company using some software, I can use any of the LLM models like OpenAI or Google Gemini. You are using those for a certain functionality, but for the knowledge base and the combination of what you’re building, you probably will end up having to build your own custom models.

So, several companies that started couple of years before the generative AI trend first started taking hold had built their own knowledge graph, but they use the LLMs for various auxiliary functionalities. Building your application entirely on that may be possible only for things like marketplaces and open generic data availability solutions.

Suppose you are a founder and you’re selling to an enterprise customer, which is almost 100% of our customer base. Say, the model answers a question that it should not have answered. For example, if it’s asked the salary information of XYZ in the company, the model answers, “That’s not allowed.” Then there is authenticity of data, Authorization needs to happen.

There are hallucination issues. There are issues of what answers are correct for the corporation based on the enterprise policies. There’re a lot of issues that actually have had an impact on the speed with which many of the companies can deliver their tools.

For example, take our portfolio RPA company Cognitos. The biggest delta they can do is what they call trusted AI. They have an AI language-based interpreter, which is sitting right in the middle of the solution and allowing people to create, communicate, and handle exceptions in English. It’s also a filtration layer that ensures that bad things don’t happen because enterprises cannot tolerate that.

So trusted AI is needed, not just large models or old school, knowledge graph-based AI models.

Sramana Mitra: Well, the other side of that coin is that you don’t really need the large models for a lot of the vertical AI solutions. What you really need is a domain-specific small model that does have the full understanding of the vocabulary of that domain and can do learning, algorithms, and workflows within that domain, but doesn’t necessarily need the whole world’s language capability.

Rajeev Madhavan: Yes, this is where the open-source Llama 3 kind of models will clearly make it easy to build that specialized small model for a particular company. They can use that technology to build it, but for generic things like I want to find addresses in the application, or I want to generate a report out of this, I may still use some aspects of very large models because they have coverage. For just writing a good English report, for example, this one is doing better than that.

You really don’t have to spend everything into one model and build into that, because that actually is going to be a tougher problem for you. Every time you update, you’ll be updating a mother of all models, whereas you’d want a lot smaller models that you can use efficiently in developing your platform.

This segment is part 2 in the series : 1Mby1M Virtual Accelerator AI Investor Forum: With Rajeev Madhavan, Clear Ventures
1 2 3 4