1Mby1M Virtual Accelerator AI Investor Forum: With Ashmeet Sidana, Chief Engineer at Engineering Capital (Part 2)

Sramana Mitra: All right, let’s discuss the second example.

Ashmeet Sidana: Sure, another example I’m very proud of is a company called Robust Intelligence. Again, I was the first investor and they did a good job. CEO Yaron Singer, PhD from Berkeley, was at Harvard when he observed the hallucination problems with the development and deployment of AI models.

This refers to errors in prediction that occur when you run these AI models, especially on the generative AI side. And so how do you make them robust? That is really the core technology behind Robust Intelligence. The company has been around for about four and a half years. It’s a very substantial business now, and I think we have some of the best technology in the world to do that.

Sramana Mitra: What use cases are they applying this on or how are they going to market?

Ashmeet Sidana: The initial use cases were in the financial services industry in embedded models or models that are not used by individual users and are not available to an outside user. A simple example would be credit card analytics, fraud analytics, or check deposit.

One of the first examples that Yaron liked to show was depositing a check using an image from your phone which is how most people deposit checks. Of course, checks are going away, but it’s a great example because we were able to show that by simply modifying just four or five pixels on that check, you could change the behavior of a machine learning system on the other end that was going to recognize what you’re doing. So our specialty at Robust Intelligence is an adversarial. That’s really where the technology is.

These are deeply technical ideas. No question about it. It’s a little bit arcane, a little bit technical, but I want to emphasize that both Robust Intelligence and Evinced, and I have other examples I can give, are companies that were started with a very small amount of capital of $2-$3 million. That is all the capital it took to get the product built and bring to market.

Sramana Mitra: I’m happy to do more use cases or case studies because it kind of gives us an idea about what kinds of things you’re investing in. I will ask you some questions based on what you’ve said so far. So let’s just carry on with more examples for the moment.

Ashmeet Sidana: Sure. I can give another example of a company called Concentric.ai, where I was the first investor in a PC round. It’s CEO and Founder Karthik had this insight that we have a separate problem on the internet in the digital age of privacy and security, for which there’s an enormous amount of political battling and regulation like GDPA and CCPA. Following these rules is relatively straightforward with structured data. You have a database with people’s addresses and social security numbers. Yes, you can mask them, delete them, change them, and manage all of that.

What do you do with all the unstructured data that exists in your organization? There’s much more unstructured data than structured data. It’s a HIPAA violation to send someone an email saying, “Hey, my friend just got diagnosed with cancer.” Now, suddenly you have that in your email, you haven’t named your friend yet, but perhaps there’s another way of identifying that friend in that email thread.

What Concentric built is a set of AI models to identify specific regulatory features that exist in unstructured data. It’s completely automated. It runs in your email and OneDrive, etc. and brings you into compliance. You know, under California law, you can send an email to a company and ask them to delete all your data. Well, that’s an easy email to send, but it’s much harder for a company to implement when they receive an email like that. So, Concentric will help you do things like that using AI.

Sramana Mitra: Okay. Let me ask you a question that is a bit more general question that will cut across all the things that we’ve just talked about.

AI versus generative AI. You know, we have been covering AI startups, I think since 2016. There have been some very good companies that have come up through the AI track over the last eight years. Then, of course, just over a year ago, this whole generative AI phenomenon broke open in the popular consciousness. Now when people talk about AI, people tend to refer to generative AI, which I think is a bit of a misunderstanding of what’s happening in the industry.

For example, there is a company that just went public, Tempus, which is a vertical AI company. It’s mostly diagnostic AI in medical imaging. This is is not a generative AI problem. You don’t need to converse with a chatbot. You need to have models that can diagnose on the basis of images and data. That’s a different problem. How do you see these different opportunities?

There are vertically AI opportunities that are also generative AI opportunities. That is also a class of opportunities. Then there is mission critical, vertical AI with generative AI technology built in that hallucinates. So that’s not an option.

How do you look at all this? By definition, these are companies that have to have technical insights and be able to apply technology to solve these problems. Their entire reason for existence is technology-driven solutions. How do you look at all this? How do you parse this set of opportunities? Do you have a preference for one or the other?

Ashmeet Sidana: What makes venture capital and investing such a fascinating business is that you have to stay in touch with the technology, and the technology advances very rapidly.

I’ll broadly define AI as these areas of machine learning that are used for prediction or recognition or classification. The original idea was to build classifiers using machine learning.

The medical imaging example that you’ve given is certainly a straight up classifier, right? You’re probably trying to identify some particular thing like a cancer or a thing in radiology or a particular marker in genes. That’s just a case of a classifier.

Generative AI is a subset of AI where you are trying to predict the next token in a language using a model. That’s the core case of generative AI. We can extend that and look at multi-modal models using sound, video, images, etc. Also, it turns out it’s very powerful over there.

But I think, the most important thing to remember about this space is that these are all specific techniques that are a point in time. There’s nothing magical about generative AI that says that it is a law of physics, that it is some core technology which will always be the important one. Tomorrow, someone could write a new paper. Tomorrow, someone could write a new model, which is more efficient, which works better, perhaps uses less energy, and perhaps makes higher quality predictions. Then we would not be using an LLM or what people today refer to as generative AI for doing it. So this is just a point in time technology that happens to be the state of the art today. Yes, it’s very powerful. Yes, it’s very useful. But it’s the state of the art today.

From an entrepreneurship perspective, for a founder, it’s important to understand that generative AI and LLM techniques are based on scale. They are based on using massive amounts of data to train models, and therefore, they’re very expensive. It costs hundreds of millions of dollars to train a foundational model from scratch. And so really, that’s not a venture-style business. While there are some VCs who have invested in that, I don’t consider that a venture business.

Sramana Mitra: It’s not a capital efficient venture style business.

Ashmeet Sidana: Absolutely, it is definitely not capital efficient to do that. If you need tens of thousands of GPUs and you have to run things for weeks and months to train your model, that is not something that is efficient. So it’s a different part of the market. That doesn’t make it wrong. It doesn’t make it bad. It just makes it different. It makes it outside the scope of traditional venture capital in my opinion.

Sramana Mitra: But there is a way to play this market, right? As we saw in the SaaS business, one of the most masterful strokes that Salesforce did early on was their Force.com platform. That kind of democratized that business. A lot of people developed applications on top of Salesforce platform and the Platform as a Service (PaaS) industry came about. Companies like Veeva were very capital efficient companies that were built on top of Salesforce’s platform. There are many others actually.

So my thesis is that same logic is going to apply here as well. The OpenAIs and Anthropics and Coheres and so on in that world who are raising very large amounts of money and training big data sets are going to operate as PaaS companies, and then there is a layer of businesses that will build applications on top of those. The applications built on the past are going to be venture style businesses. Do you agree with that?

Ashmeet Sidana: Yes, I absolutely agree with you. The way I describe the landscape of AI, opportunities today for entrepreneurs, is broadly speaking in three categories. So there are the foundational model types of opportunities where clearly the quality and the ability of the models is increasing, but that is a very capital inefficient business. So if you want to play in that space, you have to think from very deep capital pockets.

The second place where you can create value is by coming up with an application, and you could do it in the Veeva style where you could literally build an application on top of an OpenAI. If you have some great consumer insight, market insight, or business insight, you are able to solve a particular problem in a particular way better than anyone else and execute on it, I think a substantial business can be built very quickly because let’s face it, this form of AI is a new capability, is a new technology in the market today.

Then the third category is the picks and shovels business. Because AI is so transformational and a lot of people will be developing and building AI, there is an opportunity to build tools for AI, to help people develop AI, install, run, manage, and observe all of these AI applications they will be doing.

So the last two categories are much more amenable to the traditional venture model where I think companies will be created and I’m investing in both those categories.

This segment is part 2 in the series : 1Mby1M Virtual Accelerator AI Investor Forum: With Ashmeet Sidana, Chief Engineer at Engineering Capital
1 2 3 4