Google created a terrible internet; we can create a win-win solution: Perplexity AI’s Aravind Srinivas

Perplexity AI co-founder Aravind Srinivas.

In 1998, Jeff Bezos persuaded the Google co-founders to accept his $250,000 investment while the company still operated out of a rented garage. Since then, Google has evolved into a tech powerhouse, extending its dominance far beyond its core search business.

Aravind Srinivas, the founder of Perplexity AI, hopes the Bezos touch — the Amazon founder invested in Srinivas’s startup in January — will work its charm again. The AI-powered search engine seeks to set itself apart from Google’s keyword-based search by providing direct answers rather than offering users a list of web links.

Perplexity’s strategy has yielded promising early wins. Founded in August 2022, it has already notched 10 million users and is one of the buzziest AI startups in the Silicon Valley. The company recently raised $73.6 million in Series B funding at a valuation of $520 million from investors, including Nvidia, Amazon founder Jeff Bezos, Institutional Venture Partners, and Shopify CEO Tobi Lutke.

Srinivas, an IIT Madras and UC Berkeley alumnus, comes across as someone one would meet in the corridors of an Indian engineering college. He even mentions his GPA minutes into the conversation.

Watch the full interview

In an interview from the US, Srinivas discussed Perplexity’s competitive advantage over Google, plans for India, business model, and why the AI gold rush will continue. Edited excerpts:

It’s been quite a journey for you, from Chennai to Silicon Valley: IIT Madras- BTech and MTech, then Berkeley, and now co-founding Perplexity. What got you started?

I started around 2014. It was called machine learning at the time. I did an internship at a startup in Bengaluru. They took me because I got a 10 out of 10 GPA in my second year at IIT Madras.

I had one month to study machine learning because I got so interested in it, and I watched Andrew Ng and Geoffrey Hinton’s lectures on Coursera.

I then took Professor Balaraman Ravindran’s classes at IIT Madras and did a lot of research with him. I got to do an internship in Montreal with one of the Turing Award winners of deep learning, Yoshua Bengio.

Also read: Perplexity AI has more than 1 million users in India, says co-founder Aravind Srinivas

All these things got me admission to UC Berkeley for a PhD. I was noticed by OpenAI’s researcher, John Schulman, and that got me working at top research labs in the US. Over time, that translated into starting Perplexity.

How did you come up with the idea for Perplexity?

There wasn’t a clear 'one moment' where I had an idea in my head and translated it to a company.

We were trying to translate many different ideas about using large language models to build commercial products.

We just realised how we were approaching the problem of scraping a lot of data was a roundabout way and a more efficient solution to scrape the links and let the large language model, the AI model, do the reasoning on the fly when the user is asking for a question and give the answer.

That was an insight that my co-founder had. And that got launched as Perplexity.

Over the years, there have been many attempts to disrupt Google search - from Bing Search to DuckDuckGo to Neeva research, which shut down recently. But none of them have really taken off. Why do you think you will have staying power?

There is something to be said about the right place and right time. Everybody else who tried to disrupt Google did it through a focus on privacy or ad-free models. And those were not things that a large section of the population wanted or were not truly disruptive to Google either.

But when it came to offering an answer-based experience where links were no longer the prominent part of the user experience, you could just get answers directly. That was true disruption because it basically threatened their core user interface.

Google’s user interface is tied to its business model. Now, when that gets disrupted, when you do not need the interface any more when you can just converse and ask more questions, that disrupts the whole incumbent in a way that even if they wanted to do the exact same thing, it’s going to take them a lot of time to work around the existing business model, existing users, and Wall Street expectations.

And that’s the perfect mode in which a startup can operate because you can take all the risks that big companies cannot.

Now, the time to do this was 2022, when there existed sufficient progress in generative AI, especially in this one tool called the large language model. That was the right moment for a startup to start and ride the next wave.

As Bezos says, you always want to start companies where you’re not riding the existing wave, but you’re riding the next one.

You have over 10 million active users. Some of them pay you $20 a month. How many such users do you have?

We haven’t made that statistic public, but I would say that fraction is increasing by more than 50% month on month. You might think $20 a month is expensive, but it is four or five lattes in the US a month.

And are you seeing a lot of usage from India as well? Will you have a different pricing strategy for India?

We have more than a million users in India, and we want that number to expand to like 10 to a hundred X more because the Indian population is so big, and we need to do a little more work supporting Indian languages.

I know you think $20 a month is expensive for India today, and I probably think you’re right, but India’s GDP is growing so much that I believe the optimal pricing is something we will arrive at maybe in some time to come.

Will you also look at tying up with, say, Sarvam, which is building an LLM that supports local languages?

I have a good regard for Sarvam. I certainly hope that the models you train specifically for India end up being a better model to use than the generic model like GPT or Claude or the models that we built. And I’m excited to be a customer of LLMs built by Sarvam or other Indian language model companies.

What are the other revenue models you will look at? Is it advertising?

Advertisement is something we shouldn’t rule out, and we should do it the right way. If we go the way of Google, where we corrupt the quality of the answer to make more revenue and satisfy our shareholders, then we are also going to end up with the same fate.

And when we figure it out and have a sufficient user base to try out these experiments, we are going to try them out.

How is the future of the web going to be, and how is it going to be monetised? What happens to search engine optimisation?

I think the way it’s going to evolve is, you’re not going to value a generic visit to your site.

So, the amount of time a user spends on Moneycontrol and the quality of the referral is more important than the quantity.

Instead of measuring referral traffic, you’re going to measure how many times your domain got cited in an answer bot like ours. And we are forced to also give you the statistics because we rely on you creating good content for our products to be good too.

The kind of internet that Google created might seem good to you because that’s what they want you to believe, but honestly, they’ve created a terrible internet where they scrape all your content without giving you any money or creating any incentive for you to benefit from their ad business.

How does the NYT lawsuit against OpenAI impact you? What do you think about striking data licensing deals with content providers?

I don’t believe in data licensing deals. I think we can establish a better win-win situation than what Google has done with the current publishers on the internet.

Where do you think India is in the AI race?

India is certainly not very behind. There are a couple of startups trying to train large language models in India (Sarvam and Krutrim).

The awareness of the Indian public about generative AI is pretty good. We are going to see faster movement in the AI revolution in India compared to the previous web and mobile revolution.

When Sam Altman came to India last year, he said founders here or companies shouldn’t waste time building a foundational model. Do you agree? And do you have plans to build an LLM?

He’s probably saying that because building a great foundational model isn’t just about raising a lot of capital. You need top talent to work on training these models, utilising the hardware effectively, and making sure all the data mixes are done carefully and well. This requires a lot of discipline and rigour, and that needs good training in machine learning and large language models and stuff.

Most of those people who know how to do that, except for one French startup, are present in four companies today: OpenAI, DeepMind, Anthropic, and Meta. People outside these companies don’t really know how to train these models effectively, and that includes us, too.

Indian startups can be good at post-training.

How long will the AI funding gold rush last?

It’s going to last for much longer than we think because AI is the final frontier. Most of the things that we have built in software are built through a lot of hard-coded rules. Once AI is able to learn all of that and do most of it on its own, building software just gets exponentially easier. So, the value AI is creating is exponential.

What would you advise an 18-year-old in India or a 21-year-old who’s just graduated who wants to build a startup in AI and maybe do something like what you’ve done?

Follow your heart. Don’t look at markets and decide "oh, the market wants a sales analyst product. So I’m going to build it."

Nobody cares about the product you’re building initially other than you and your co-founders. And it’s hard to care about something if you don’t like it.

So, follow your heart and get some users, and then the market will automatically guide you to building something that has some commercial value.