Perplexity's Bet: Search Becomes Answers, and Answers Become a Business

The Perplexity CEO says the simplicity of Perplexity's interface obscures the complexity required to make it work. He frames the product around five dimensions that must all be strong: accuracy, reliability, latency, delightful UX, and constant iterative improvement, including personalization. He notes that roughly one in 100 startups can do one of these well, which makes doing all five simultaneously a deliberately ambitious target; without that ambition, he argues, it is hard to create a multibillion-dollar company.

Behind a query, Perplexity tries to understand and reformulate the question, decide which pages to use, choose which parts of those pages matter, and render the answer appropriately. The answer might be a paragraph, bullets, an image, or even a video, and each sentence needs supporting citations to reduce hallucinations. The company also introduced shareable permalinks so that the value of a query and answer can be useful to someone else without them asking the same question.

Perplexity suggests follow-up questions because most people are curious but few can articulate their curiosity into precise queries. The co-pilot feature follows the same philosophy: do not blame the user for not being a great prompt engineer. The CEO cites Larry Page's story about demoing Google against Excite and being told he had typed the query wrong; Page's response was that he was just the user, and the user is never wrong.

The company tries to stay vertically integrated so designers, product engineers, and backend teams talk constantly. A design choice can generate better interaction data, which can make the AI smarter and the product better, but only if the relevant teams are in the same conversation. The CEO says the culture is built around quality, truth, and velocity, values chosen because they mirror what the product must deliver.

Perplexity began with off-the-shelf OpenAI models, then fine-tuned smaller and faster models, and later increased its use of open-source models and released its own. The CEO says he would do it exactly the same way again: if you want to be product-focused, start with somebody else's models and do not waste time building your own at first. He notes that he and his CTO were former AI researchers who had trained models for half a decade, so they understood the work, but founders without that background should be even more product-focused.

His advice is to get a product out, get users, get returning users, and sustain usage before worrying about model ownership. That first barrier matters because it helps attract funding, but he says getting really good engineers is harder than getting investors and changes a company's destiny more. A foundation-model company needs hundreds of millions of dollars and scarce researchers who know how to train large models; he says H200 GPUs may be obtainable, but the people are harder to find.

Perplexity eventually built its own models because it did not want to depend too heavily on a company building products in the same space. The CEO says the timing was deliberate: the team waited for the Llama wave and expected open-source models to arrive, just as GPT-3.5 had created a wave and Nvidia's TensorRT promised faster, more efficient inference. He compares this to a Jeff Bezos idea: wait for the next wave, then position yourself to take advantage of it.

On moats, the CEO argues that a company only needs to think about moats once it has something to protect. He says he would rather be a rapper with 100,000 users than have a model inside a product that nobody knows, and he expects model consolidation around OpenAI and Anthropic on the closed side and Llama and Mistral on the open side. Perplexity wants to remain model-agnostic with the optionality to control its destiny, because users do not care which model produced the answer; they care about the best answer.

The CEO contrasts Perplexity with Quora and Wikipedia by saying those products were built to maximize knowledge, while Perplexity tries to maximize knowledge velocity. In his framing, Perplexity is optimizing delta knowledge divided by delta time, accelerating how quickly someone can get a relevant answer backed by sources. Wikipedia offers a static article and Quora waits for another human to answer, but Perplexity performs the research and returns a concise result in seconds.

Personalization is central to that velocity because different readers want different depths. Someone reading about black holes may want a high-level summary or may want to go deep into the physics, while someone reading about a celebrity may want either gossip or just the basic facts. Wikipedia has one article for everyone, and Quora is too slow; Perplexity aims to give fast, high-bandwidth access to knowledge at the level the user wants.

The CEO says Perplexity will not stop at research questions and wants to serve all of human curiosity, but he is careful about the vertical-versus-horizontal debate. He recalls being told by every VC to build a vertical search engine and by others that Perplexity was too horizontal. One investor, he says, told him that the same advice appeared in the late 1990s, when many companies tried to build a Google for a domain and most failed.

What worked in travel, he argues, was building an end-to-end experience: Booking.com is not just about getting hotel search right, because users can book and finish the task. Once a company takes that extra step, it is no longer just a search company; it becomes a travel company, and the moat is in supply and fulfillment rather than the search technology alone. That is the strategic choice Perplexity is trying to navigate as it expands.

The CEO admits Perplexity benefited from timing, but he says earlier Google challengers made a repeated mistake: they tried to run a Google competitor exactly like Google. They built crawlers, indexers, rankers, and go-to-market motions that looked like Google's, which meant they were competing against a company that had done the same work for 20 years. Anti-Google marketing and privacy positioning were also already taken.

He points to DuckDuckGo and Brave as the exceptions that captured privacy-focused search through go-to-market, branding, and product positioning rather than trying to out-build Google's core technology. DuckDuckGo marketed constantly around Google's privacy failures, while Brave built a privacy-focused browser, a token, and its own search index. The lesson is that a challenger needs a clever strategy and positioning, especially when building the exact same technology as the incumbent.

The AI moment was a rare failure in Google's history: for a period, another model provider had an amazing model and Google could not match it, and that model was accessible through APIs. The CEO says Perplexity took advantage of that opening. He does not claim Google will stay behind, but he treats the moment as the window that made a new search entrant plausible.

Looking ten years ahead, he expects search to become answers and agents that do tasks for users, more like talking to a knowledgeable friend than typing into a search box. The interaction должен be concise enough that users do not feel the system is rambling. Perplexity has not shipped voice-to-voice yet because eyes can read an answer faster than ears can hear it, and long spoken answers quickly become annoying.

Perplexity's model and inference work is partly about trust: some users worry that a wrapper company will run out of money or fail to build its own infrastructure. The CEO wants to show that Perplexity can serve its own models in simple steps, starting with models that rival GPT-3.5 Turbo on its traffic, and then ship them into production without users noticing a downgrade. He cites Paul Graham's idea that the AI companies that survive will be those that can ship their own models with no noticeable difference or an improvement.

He is explicit that Perplexity is not at GPT-4 level and says nobody else is either, pointing to Gemini's delays as evidence of how hard it is to reach that level. The hope is that another strong base model, such as Llama 3 or Mistral 2, arrives and Perplexity's data flywheel helps it get closer. The company's own advantages are query reformulation, query understanding, search index quality, summarization quality, and co-pilot interactions.

Serving models through someone else's GPUs and inference provider still leaves a company paying someone else and compressing margins. The CEO says that makes you a different kind of rapper unless you build an aggregator business like a Macy's for AI, which can be a good wrapper business but is not Perplexity's plan. Perplexity wants to be a search-focused business, so serving its own models matters.

On RAG, the CEO warns that being good at web search does not automatically translate into internal enterprise search. Google Drive search is weak, he argues, because indexing, embeddings, snippet selection, and inverted indexes are fundamentally different for internal documents. He is skeptical of claims that RAG has been solved by an embedding API or by one product, and he notes that throwing more information into long-context models can increase the chance of hallucination; retrieval, indexing, and ranking still need use-case-specific signals.

The CEO describes Perplexity's Discover tab as editorially curated, with a team choosing queries and content from web scraping. That led him to the lesson of Facebook's early growth: a product can grow by making existing users happier, but it also has to be intuitive for a new visitor who has never seen it. Facebook found that the number of real-world friends already on the network was the strongest predictor of whether a new user would retain, so it optimized for that new-user conversion rather than only adding features for existing users.

He says this is the hardest challenge for Perplexity, especially with vocal power users who ask for many features. The YC mantra of talking to users and shipping what they want works before product-market fit, but with millions of users a company cannot do everything the loudest users request without risking the brand and the next wave of users. He also argues that unintuitive landing pages and too many sign-up modals hurt growth, and he admires Apple for optimizing the user experience because it has distribution and lock-in to support that choice.

Perplexity's path to search was not direct. The team first tried text-to-SQL after an investor advised narrowing Google's search problem to searching over databases, but enterprise feedback was discouraging: dashboards already existed, and Databricks' chief architect told them that most SQL work is already written and run as periodic jobs. New queries were often built by drag-and-drop in Power BI, Tableau, or Metabase, so the team concluded text was not a universal interface and users would choose whatever gets the job done fastest.

They also scraped Twitter and built graph search for things like finding mutual follows for an introduction or seeing Paul Graham's AI tweets; investors liked it partly because people love searching for themselves. The team had three products, including a SQL editor, a Twitter search tool, and a summarization search, and launched the summarization search first because friends in their Discord said it was better than Google. The tool began as an internal Slack bot for questions about Ruby, bash, React, health insurance, Twitter news, and Facebook's stock; after the web UI shipped, usage exploded, Twitter search went viral, Jack Dorsey quote-tweeted it, the site crashed, and handle summaries spooked people with how much the AI seemed to know. Usage during winter vacation convinced the CEO it was real, and the team decided to build a fully conversational search product rather than slap a chat UI onto search.

In a quickfire segment, the CEO calls moats and vertical focus overhyped because he has not seen many AI-native products take over a vertical. He says delightful user experience is underhyped, and he notes that Perplexity's collections feature has usage but has not yet become the collaborative research workspace people wanted. Users do not want to manually create collections and move threads around; the feature needs to become more automatic and seamless.

On open source, he expects whatever GPT-4 makes possible today to become possible at cheaper prices and faster latency later. That does not mean the frontier stops moving: when the next model arrives, it will enable new things that were impossible before, such as co-pilot, dynamic prompt engineering, image upload and homework help, chains of reasoning, actions, and multimodal video understanding. Perplexity tries to move some workloads off GPT-4, but new capabilities keep resetting the frontier.

The CEO argues that AI regulation is premature because the widespread economic benefits have not yet appeared. He is not against discussing AI safety, but he worries that slowing builders too early prevents society from learning what safeguards are needed; he cites the Nvidia CEO's point that you cannot build safety around things that do not exist. He also sees flop thresholds and licensing proposals as a way for large labs to entrench themselves, and he wants many eyeballs and many builders rather than decisions left to three or four rooms.

If he were not at Perplexity, he says he would work on AI video and voice-generation companies, because realistic generated voices and cheaper movies could create powerful emotional experiences. His personal project is to record his parents for hours, make a podcast with them, and preserve their voices so he can talk to them or hear a congratulatory message from his mother even when she is not around. On OpenAI, he sees two schools of thought, one moving faster and one more cautious, and expects the company will not be the same as before; with limited GPUs, it cannot simultaneously be an app store, a first-party product, infrastructure, enterprise, and cutting-edge research, though it remains a leading AI organization.

Jacob Effron and Pat Chase close the episode by highlighting Perplexity's model sequencing: start on off-the-shelf models, get distribution and product resonance, then reduce dependence on external providers. They note that the CEO is optimistic about wrappers turning into large companies, and they like his framing that moats are a good problem to have because it means a company has something worth protecting. They also emphasize Perplexity's building-in-public advantage, where shipping products created viral distribution, including the Twitter search product that helped drive attention to the main search product.

The hosts return to the knowledge-work comparison with Quora and Wikipedia, especially the idea that users want different depths on different topics, such as black holes versus celebrity gossip. They discuss the CEO's view that RAG and hallucination reduction will look different for each end use case, so one company solving RAG broadly is unlikely. They also flag his claim that recruiting engineers is harder than raising capital, and they see search as a classic hard AI problem that remains attractive to engineers. Pat Chase says the personal AI project of recording parents gave him goosebumps, while Jacob Effron wonders whether such persistence is beneficial or a Black Mirror scenario.