1Mby1M Virtual Accelerator AI Investor Forum: With Ashish Gupta, Partner at Clearvision Ventures (Part 4)

Sramana Mitra: So what of platforms and this euphoria – this absurd amounts of funding that are going into a handful of very large platform companies? How do you analyze them? I can understand that you don’t want to invest in them.

Ashish Gupta: Yes, I’m too poor! I’m missing four zeroes!

Sramana Mitra: Certainly. No question! But how do you analyze them?

Ashish Gupta: First and foremost, I don’t have a deep enough understanding to have an opinion that you should take too seriously. It’s based mostly on reading. Ideally, we should have invited my wife to this. She’s the AI PhD. She does machine learning research at Google. So you have the wrong family member for this conversation!

The little that I have been able to piece together, I come back to the disproportionate advantage that the larger guys, like the Amazons and the Facebooks, have. They have access to untold amount of data, which is going to be a crucial element in the future, especially to keep these models current.

On the natural language understanding portion, one reaches an asymptote. However, on the contextualization, which comes back to your vertical LLMs or Gen-AI plays question also, why should I take on the overhead of an all-knowing beast and then try and constrain its power by putting things around it? I would rather start with a smaller beast and beef up its biceps, if that is all I need, as opposed to contain the other beast’s quads and brain. It’s upside down.

Sramana Mitra: And introduce inordinate amounts of hallucination.

Ashish Gupta: Absolutely. So to contain that hallucination, I first take on this nuisance of controlling context, instead, why don’t I start with the much smaller, thinner model, if that’s all I want?

Sramana Mitra: My instinct also. I can’t claim to know enough computer science. Even though I am a computer scientist, you are a computer scientist, but we are not working at this level at the moment in the computer science level. But I am going by just pure instinct. You talked about the database perspective, right? Databases need metadata and how you organize the data. There already you’re constraining it.

Your learning models will take that constraint – the data structure and the metadata – from that organization. Then why not just work with a smaller vocabulary, smaller language model that is domain specific and just work with that? Then there is no hallucination, none of the extraneous stuff.

That’s where I stand as well.

Ashish Gupta: Same here. If I’m in the pharma world, I don’t want to keep telling the model that MS is not Masters of Science or Microsoft or 40 other things. Also, the company is asking, “I’ve got all this data, which in free text can now, for the first time, be aggregated. I don’t need to structure it. I can stick it all into an LLM. Why should I worry about how is this data going to get used, abused, mixed up with all kinds of other folks?” I’m a large Fortune 100 company. I have more than enough data for this model to become rich and be useful. So I struggle with the large money use cases being built on top of.

Sramana Mitra: Yes. Same here. I think I’m not seeing it.

Now, having been in the industry for a while and SaaS being quite active for many years now, right? I would say around 2008, SaaS became really successful as an investment category and lots of money went in there. And we’ve seen an explosion of cloud companies and SaaS companies. Vertical SaaS has been a very successful field with good workflow automation, etc.

Now comes a combination of SaaS and AI. So there is this new company starting from scratch that is AI plus SaaS. Then there’s an old vertical SaaS company that maybe is doing $30-$40 million in revenue, but doesn’t have AI.

What does does the incumbent do?

Ashish Gupta: So I think pretty much every company has no choice but to start using some parts of AI. I think that is fait accompli. SaaS companies were built assuming an RDBMS as the underlying platform. Some of that portion will also get fragmented if I have unstructured data.

Let’s take a customer support app. Customer support apps take unstructured data, forcibly structure it, and in the process, lose fidelity. There is a strong reason for reconsidering an RDBMS as the backplane on which one builds a knowledge-oriented app. It was being used because there was no free text-based compute platform before AI arrived. So I think in the example that you portrayed, the incumbent has an advantage.

The advantage that the incumbent has is that AI is still a technology. The business problem on top of it is the real thing that causes people to buy that solution or not. That understanding is deeper with the incumbent than with the person who’s got the new tools. Technologies, at the end of the day, don’t end up determining, especially business-facing companies’ future. They get adopted horizontally by everybody. If somebody is lazy about it, then it’s a different issue.

But this $30- $40 million company, I think, has a better chance if they consciously start adopting AI than a person who starts brand new and modulo, those classes of applications where the entire architecture has to be revisited.

Sramana Mitra: That’s part of the issue though, Ashish, because the AI stack and a non-AI cloud stack are different. Offering the full power of AI in an old architecture is a non-trivial problem to solve.

Ashish Gupta: So if it is more of a knowledge-intensive, free text-based application, I would be cheering for the newbie. Let us say this was a company that was dealing with financial data where you really cannot walk away from inherently structured data.

Sramana Mitra: Yes, that’s different.

Ashish Gupta: Then, I would be cheering for the incumbent. Then, there would be the real hard problems, which I think you’re referring to, which are people in the middle, where I think one would have to lose some money before one finds the right answer.

Sramana Mitra: No, I think the nuance that you are pointing out is correct, in that it depends on the problem and it depends on the domain and the nature of the data. If it’s a more structure-oriented data, then it’s easier to make the transition. If it’s largely unstructured data that really has the power to solve problems, then it’s the newbie that is at an advantage.

Very cool. Ashish, it’s a pleasure. We will continue the conversation.

Ashish Gupta: Thank you for inviting me.

Sramana Mitra: Thank you. Pleasure.

This segment is part 4 in the series : 1Mby1M Virtual Accelerator AI Investor Forum: With Ashish Gupta, Partner at Clearvision Ventures
1 2 3 4