From a roughly $500 million valuation in January to a potential $9 billion this November – this company had four funding rounds so far this year, with the fifth on the horizon!

Today, in our GenAI Unicorn series, we explore Perplexity AI, an AI-driven search engine that leverages large language models to provide answers with cited sources. Parsing the internet real-time, it aims to deliver relevant results. This approach has sparked controversy, with debates over the validity and copyright of the sources. Despite – or perhaps because of this – the startup has drawn millions of users and is now seeking to raise $500 million and double its valuation to $9 billion.

Perplexity AI's journey is remarkable. Two years ago, they secured initial funding through cold emails – before even having a product. Starting as a Twitter-based search tool, they soon had to pivot to challenging the very foundation of online search. Just recently – in November – they offered to New York Times to provide services during a tech workers' strike, which of course sparked backlash as Perplexity’s CEO, Aravind Srinivas, was seen as undermining the workers' collective bargaining efforts.

How do they manage to thrive among controversies? Can they truly challenge Google? What does the future hold as SearchGPT emerges? Could an acquisition be on the horizon? Let’s dive into the details and uncover the trends shaping this unique story. It’s a long read.

In today’s episode:

Aravind Srinivas – the co-founder and CEO of Perplexity – initially wanted to study Computer Science at Indian Institute of Technology (IIT) Madras but was admitted to Electrical Engineering. That didn’t stop his interest in algorithms programming.

“A friend mentioned a machine learning contest to me. At the time, I didn't even know what ML was. It turned out to be fun, and I ended up winning the contest without spending too much time on it – it just came naturally. That’s when I decided to dive deeper into it,” tells Aravind Srinivas his story.

The same 2017 when he earned his a Bachelor's and Master's degree in Electrical Engineering from IIT Madras, he went to Berkeley, to work on his Ph.D. in Computer Science. But the most brilliant part was his strategic internship.

Internship: 2018, May-Aug – OpenAI, Research on Policy Gradient Algorithms

“I came to Berkeley thinking I was definitely one of the top AI PhD students. Then I joined OpenAI, and it hit me hard – everyone was so much better than me. It was a big reality check.”

That summer of 2018, OpenAI published its first GPT (Generative Pre-trained Transformer) model.

"We realized there was a new way of learning – using all the internet data to learn from it – and I felt that was going to be crucial. I told my advisor in Berkeley, ‘This is the right direction; we should pursue it.’ Surprisingly, he was open-minded and said, ‘Alright, I’m not a specialist in this, but let’s give it a try.’ So we spent a lot of time – holidays, weekends – just learning, coding, and understanding everything we could. We did this for two years, which eventually led me to a new research focus: combining generative AI and reinforcement learning. This approach powers technologies like ChatGPT, which doesn't just predict the next word but ensures it knows how to communicate effectively with humans."

Internship: 2019, May-Sept – DeepMind, Large-Scale Contrastive Learning: CPCv2

In 2020, Aravind Srinivas met Denis Yarats over email after independently publishing similar research papers on AI training methods, each two days apart – Srinivas at UC Berkeley and Yarats at NYU. This shared academic interest sparked an ongoing dialogue between them on AI advancements.

Internship: 2020-2021, May-April – Google, Transformers for Vision: Bottleneck Transformers, HaloNet. SoTA Vision Models: Copy-Paste Augmentation, ResNet-RS.

While at Google, Srinivas started thinking: “How could we innovate in search with Google so dominant?” At the moment, he was also reading the book “In the Plex: How Google Thinks, Works, and Shapes Our Lives”, and was very inspired by that. Transformers seemed to hold massive potential for a breakthrough in search. Aravind even reached out to Ashish Vaswani, one of the creators of transformers at Google, saying, “I want to work on this with you – it’s the next big thing.”

Unfortunately, the timing wasn’t right. So 2021, he joined OpenAI as a Research Scientist, focusing on language and diffusion generative models. Then GitHub Copilot – a tool enabling programmers to complete code as they write – achieved real adoption and profitability.

The timing was finally right to bring generative technology to market. In 2022, Aravind Srinivas, then at OpenAI, and Denis Yarats, at Meta AI, teamed up with Andy Konwinski, a co-founder of Databricks, and Johnny Ho, a former Quora engineer and Wall Street quant trader, to develop… just something new.

In August 2022, Perplexity was born. The four visionary founders had a mix of ideas about search, inspired by Google, Copilot moment, academic rigor with correct citation, transformer capabilities, manifested in brewing ChatGPT that Srinivas, working at OpenAI, most likely knew about.

“I thought the startup moment had arrived where you could actually translate it to products.”

Yet, they had no clear idea or product.

Nonetheless, Aravind reached out via cold emails to a few investors, like Nat Friedman and Elad Gil. Within a month of its inception, in September, with a strong belief in the transformative power of transformers (remember, this was before ChatGPT), they raised their seed round: $3.1 million. And the work began.

Their initial idea was to create a tool that translated natural language into SQL queries to interface with databases, initially focused on public Twitter data. The founding team built a demo that allowed users to ask questions about structured data, like querying Elon Musk’s follower metrics. It was called Bird SQL, text-to-SQL solution, and they’ve been working on it for at least four months.

Both Elad Gil and Nat Friedman were impressed by the speed at which the team iterated on ideas and delivered. “You’ll mention an idea to them, and then a day later, they’ll have something built,” said Elad Gil. “They were sending me demos twice a day,” added Nat Friedman.

And then, November 30 happened, ChatGPT was unleashed onto the world.

“I remember that day very clearly,” says Dennis Yarats. “I had just woken up and saw a lot happening on Twitter. We quickly recognized that this was groundbreaking – it wasn’t something that would just come and go. We immediately began considering what we had: a prototype addressing early feedback that ChatGPT was receiving, like concerns about hallucinations and lack of source transparency.

We decided to act on the opportunity. Within two days, we put together a simple website as a joke and posted it on Twitter, not expecting much attention. But to our surprise, it gained traction quickly, with people retweeting and praising it, despite its limitations – it was slow and didn’t work perfectly,”

Aravind: “It happened on December 7. It was a unique launch – we didn’t have a sign-up process or a waitlist, just a simple search bar. Michael Dell even sent me a LinkedIn message congratulating us. That was a huge moment for me.”

The holiday season approaching, and they were hesitant to continue, thinking interest might fade. But in January, the traffic hadn’t dropped – instead, it had grown. That was a clear sign they had something valuable. So, they unanimously decided to pause text-to-SQL and fully focus on general search, setting aside months of work – a choice that proved right.

“Every sentence that it says also has corresponding reference, or we call it a citation. This is all coming from our academic background. Like my co founder, Dennis, and I are PhDs, we figured that we would use this principle that everything in a paper that you write in academia, you have to back it up with reference from some other paper. And that's how Perplexity works,” explains Srinivas.

Soon they realized that user may start with curiosity without a clear idea of what they want. Rather than expecting them to be expert prompt engineers, how AI can assist users in refining their inquiries? The responsibility, the founders argued, lies with the AI to guide users toward effective prompts. To address this, they developed a "copilot" feature on their site, designed to ask clarifying questions as users type, helping to expand and refine the prompt interactively – similar to conversing with a friend while making decisions. That also makes communication more “human-like”. That’s how they came up with their proprietary intent-recognition algorithm. They call it Search Companion and it’s available for their Premium subscribers.

Citations are a cornerstone of Perplexity's strategy to build trust and transparency in AI-generated responses. By attributing information to reliable sources, Perplexity empowers users to validate content and make informed decisions. This focus on accountability sets it apart from traditional language models, which often lack the ability to provide verifiable references.

Beyond RAG and source citation, Perplexity's flexibility with multiple language models enhances its reliability, allowing users to cross-check outputs and identify potential hallucinations. Specialized searches in platforms like Reddit or Wolfram Alpha further refine its ability to address niche queries with precision.

Despite its advancements, Perplexity acknowledges the need for continuous refinement. The team actively works to address challenges such as outdated or inaccurate sources and the risk of “second-hand hallucinations” (occur when AI generates responses based on incorrect or hallucinated information from another AI model or source.)

Perplexity AI has faced significant legal and ethical challenges over content usage. The New York Times and News Corp accused the company of copyright violations, with lawsuits alleging unauthorized data scraping and verbatim reproduction of journalists' work. Additionally, Wired reported Perplexity’s disregard for robots.txt protocols, raising concerns about its web crawling practices. Accusations of plagiarism further spotlight ethical concerns, as Perplexity allegedly republished content without proper attribution. These controversies reflect broader tensions between AI companies and traditional media over content rights.

To address these issues, Perplexity launched a program sharing ad revenue with publishers like Time and Fortune. However, CEO Aravind Srinivas sparked further debate by arguing that facts should be universally accessible, rejecting ownership over knowledge. His controversial offer of technical support to The New York Times during a strike added to the company’s criticism, highlighting the complex relationship between AI innovation and media ethics.

Perplexity is now generating approximately $50 million in annual recurring revenue, a significant jump from $2.5 million last October, as reported by The Information. However, the company has yet to achieve profitability.

Overall, the platform handles 100 million weekly queries and attracts major enterprises like Databricks and Zoom, contributing to its $50 million ARR as of October 2024.

In the interview to Lenny Rachitsky, Srinivas said: “Big challenges today revolve around scaling from our current size to the next level, both on the hiring side and in execution and planning. We don’t want to lose our core identity of working in a very flat and collaborative environment. Even small decisions, like how to organize Slack and Linear, can be tough to scale. Trying to stay transparent and scale the number of channels and projects without causing notifications to explode is something we’re currently trying to figure out.”

Srinivas once remarked, “Our competition isn’t even Google; it’s user awareness.” Small scale competitors face the same awareness problem:

Big guys:

The growing competition indicates that Perplexity’s success may hinge on how well it can either differentiate itself further or expand into niches that competitors have yet to explore.

In 2023, Perplexity received acquisition offers ranging from $150 million to $200 million from players such as X (formerly Twitter), OpenAI, Notion, and Microsoft. Elon Musk expressed interest in integrating Perplexity's technology to enhance X's search functionality. These discussions did not culminate in a deal back then. Now, with a potential $9 billion valuation, only X (formerly Twitter) might be bold enough to pursue this deal. Another appealing factor for Musk is how much traditional media hate Perplexity

Potential Synergy with Meta: As Meta develops its own AI-powered search engine to reduce reliance on Google and Bing, Perplexity emerges as a potential partner. Its proprietary search technology could accelerate Meta's initiatives, offering advanced web crawling and indexing solutions. However, Meta would need to assess the extent of Perplexity's proprietary technology versus third-party dependencies. A partnership or acquisition would hinge on Meta's confidence in Perplexity's consumer appeal and the justification of its valuation.

Traditional search is quickly becoming a relic of the past, and Perplexity embodies the shift toward something entirely new – a platform that prioritizes human-AI collaboration and transcends the boundaries of a mere search tool. By turning search into a conversation, Perplexity is redefining how we engage with information, emphasizing context and interactivity. This vision sets it apart. Could its approach be replicated? Likely – but Perplexity’s founders are moving fast, overcoming obstacles and shipping features aligned with their mission to build a dynamic knowledge hub where anyone can contribute and interact.

The company is reportedly finalizing a $500 million funding round, potentially elevating its valuation to $9 billion—an impressive leap from its $150 million valuation in March 2023. This substantial capital infusion could fuel technological advancements, market expansion, and talent acquisition, reinforcing its position in the AI search sector.

On the flip side, profitability looms as a significant hurdle. The coming year will be crucial in determining whether Perplexity can stand on its own or will be acquired – potentially bolstering a larger player in the race to redefine search.