Never Cite More Than Five Links: Perplexity's Bet on Answer Engines

Sarah Guo and Elad Gil open this March 2023 episode of *No Priors* with Perplexity's two operating founders: CEO Aravind Srinivas, a former OpenAI research scientist with a Berkeley CS PhD, and CTO Denis Yarats, who had been a research scientist at Facebook AI Research and a machine-learning engineer at Quora. A third co-founder, Andy Konwinski, helped them incorporate. Elad, listening on the other side of the table, is the company's first investor — the person who offered to put in money during the founding days.

The company started around August 2022. At recording time it was seven months old. Aravind's first picture of the founding is almost cinematic: an open space in San Francisco's Noe Valley, opposite a café called Martha's, where he told Elad it would be cool to build a visual search engine. The only way to disrupt Google, he thought then, was to skip text search and go straight from camera pixels. Elad's reply was blunt. That path would not work. He needed to think about distribution.

Aravind still thinks Elad was right. Search, he says, is tremendously a distribution game as much as a technology game. What they actually had, in those early weeks, was not a product. It was a shared itch: large generative models were exciting, and search — general or vertical — was the motivation. ChatGPT had not yet rewritten the consumer internet. They were bouncing.

Elad once suggested they try text-to-SQL. They got excited and started prototyping. Denis was hacking a Jupyter notebook extension that would put a copilot in every cell, then pointing that at databases. They talked about indexing pieces of Twitter and other feeds and searching on top. Aravind's phrase for the period is "a bunch of non-linear pathways" that eventually led to the product they have now.

Guo is impressed by the rate of iteration: point at a direction, prototype, see if it works, move on. Both founders credit academic training. In a PhD lab you sit on hundreds of ideas, run experiments quickly, read the results, and iterate. Product work adds two extra loops they had to learn: get the engineering to a state you can show someone, then actually put it in front of users, companies, and potential customers and collect feedback. Those two things, Aravind says, are why the company exists.

Denis is the reason the engineering and recruiting keep up. He brought in co-founder Johnny Ho, once ranked world number one in competitive programming — Aravind's analogy is being Magnus Carlsen for code. Denis showed him what they were playing with; Johnny got excited and joined. A few people help you accelerate a lot, Aravind says, and Johnny is a big reason Perplexity moves this fast. Denis kept repeating the move through the fall, pulling in more of the same kind of person.

Denis's first hiring filter is curiosity, not a large-model résumé. Plenty of people have heard of LLMs without a specific background; he wants the ones who are burning to learn and will put a lot of energy into it. He would rather take someone with that desire than someone with more experience who will not work as hard. The second filter is a trial period. Once they see a real chance of hiring, they ask the person to work with them for a while — closer to a few months of living together than a sequence of 45-minute loops.

They intend to keep headcount very small, which makes every hire expensive. Big-company interviews of four or five meetings sometimes work and sometimes do not; the trial, Denis says, gives both sides more confidence. Candidates get to see whether they want this work and this pace. The pace is the point he will not paper over: work-life balance is not ideal. That is the only way he sees to beat the competition — trade very fast, do great things — so the alignment has to happen at the beginning.

Aravind makes the hiring bar a strategy. If you do not yet know which product you will build or which market you are going after, iteration speed is the only thing you can hope for. Until they figured out the business, "people should iterate really quickly with us" was non-negotiable. Elad adds the startup version of the same math: incumbents have more people, more money, more distribution, more product. The only real advantage a startup has is speed.

Elad remembers a dumb meme from Twitter's mobile days: only hire people with prior mobile experience. He thought you should find great engineers and let them learn iOS. He asks whether the LLM world has grown a similar dogma. Aravind's answer was already settled from watching OpenAI. The people who did GPT-3 and later founded Anthropic came from physics: CEO Dario is a physics PhD, Jared Kaplan a physics professor who wrote the scaling-laws paper. OpenAI succeeded by pulling extremely talented physicists who wanted into machine learning, then into language models.

The engineering bench was Dropbox and Stripe people — solid software engineers who could build infrastructure and front-end, now doing AI work. For Aravind the lesson is clean: to work in AI research or AI product, you do not already have to have been in AI. Johnny Ho is the in-house proof. He was not an AI person. He was a competitive programmer and a trader who had spent a year at Quora, and he is as good as anyone at picking up new things. The stranger observation is about the models themselves. People who use LLMs to build products, Aravind says, often understand them better than the people who actually ran gradient descent. A Stanford or Berkeley PhD student may know how to train the model and still not be the best person to ship a product with it.

Denis adds the research-lab version. Early Facebook AI Research in Menlo Park had an exclusive culture: no PhD, no research-scientist title. He disliked it enough that it became one reason he later did a doctorate. Experience taught him the opposite of the academic snobbery. The best research scientists are also very good engineers. DeepMind and OpenAI made the most progress over the last five or six years in places stuffed with extremely good engineers. The academic view that engineers are not quite smart enough is, he says, backwards.

Aravind closes the loop with Google's early hiring. Jeff Dean came from compilers and systems; Urs Hölzle was a systems professor. Google told those people it had the most interesting computer science problems and could scale them — come work on search. Most of the celebrity researchers were people who wanted into that space for the first time, not information-retrieval veterans. Elad notes how fast the desired pedigree has flipped again: a few years ago companies wanted lab PhDs; now they want people who have run multi-billion-parameter training, and those people come from OpenAI and DeepMind.

The steepest learning curve for Aravind, coming out of research, has been how to run a company. He is not doing much core engineering; Denis is. The usable method has been rapid advice from people including Elad, plus being brutally honest with himself when he is wrong, listening to feedback, and course-correcting fast. For Denis the hard part is building a team: who does what, how to prioritize, what to work on. On a small team you have to identify the one thing and put all your focus on it.

Early on they made the opposite mistake. At six or seven months they had already built Twitter search, a Salesforce integration, a HubSpot integration, and more — most of it never released. As a leader, Aravind says, if you cannot pick, people lose faith; they think you have no clarity and that is why you make them do a new thing every week. Do one thing wrong at a time and correct, rather than five things at once to see which wins. Academia taught them to hedge: one first-author project and three or four co-author projects, hoping one is a hit. Startups do not work that way. Focus, then iterate on one thing, or a few.

Perplexity's stated mission is the world's most trusted information service. Guo asks how they think about factual accuracy, bias, and mashed-up sources. Aravind reaches for the first paper a PhD advisor marks up: you only write what you can cite; everything else is opinion, not a scientific fact. That is why the first version was citation-powered search. The subtle point, he insists, is that they are not retrofitting citations onto a chatbot. It is a citation-first service that will never say anything it cannot say. Ask it who it is, ChatGPT-style, and it still goes to a search engine, pulls sources, and comes back with an answer. It will not say it is Perplexity, a bot, how are you. They do not care about personality or character. They care about truth and accuracy.

Aggregation is the remaining mess. Mash multiple sources and the model can hallucinate identities. Aravind's example is sitting across the table: there are several Elad Gils, one a venture capitalist, another a doctor; if the model also reads LinkedIn it might decide the Elad Gil with a biology PhD is a different person from the investor and write a funny summary that collates the wrong lives. They do not have a proper fix. Models are getting better at the subtleties, and they give users the power to drop sources they think are irrelevant, Wikipedia-style. He does not think one LLM will magically solve it. You need an end-to-end platform where users can correct mistakes, with incentives designed so the same tools cannot be used to hide information.

On reinforcement learning they are still in the RLHF neighborhood: user feedback on summaries, contractors rating completions, and now language models prompted to do some of the contractor work — Anthropic has a paper on that. They have not gone on to agents and browsers; Aravind wants to stay on the first part for at least six months to a year. Denis adds that you do not need full-blown RLHF to move quality. Rejection sampling, a discriminator on top of several Llama samples, ranked by another model, is in a sense one step of reinforcement learning and is already very effective.

They believe chat UI is the future, and people are using follow-ups heavily. The caveat is a duty of care: if you can get the answer right on the first attempt, you should. Google already does a kind of implicit chat with "people also ask"; you click into a follow-up without a chat box. Complex questions make a chat UI more honest than ten links. A question like whether it will rain in the Bay Area over the next week does not need a conversation. Perplexity wants to support both. Since they shipped chat, the share of queries that go to at least one follow-up has been rising.

Asked to sketch search five years out, Aravind reaches for a phrase Andrej Karpathy tweeted and Satya Nadella also uses: answer engine — something that answers the question instead of handing you a pile of links. He believes Perplexity is the first conversational answer engine, and that answer engines will become their own market segment, the way people have a default search engine today. Follow-ups become normal. Actions sit on top of answers, the way Google already lets you book a flight from a result; the product becomes an assistant. And much less traffic hits the underlying content sites, because far fewer links need to be opened.

That last point is a product decision, not a limitation. Perplexity does not cite more than five links. Users have asked for ten or twenty, or for the old ten-blue-links stack under the summary; the team said no. They want people to feel the difference. The first version shipped with three or four links and a fifty-word summary. Fewer tabs, only the ones you really need. Pull-to-push is closer than a five-year science-fiction gap: Google Now already tried notifying you about your football club; language models can do that better, and Aravind thinks it happens in a year, not five. Denis, looking at Google as a user, already skips the first two or three results because they are SEO or ads. An answer engine means fewer clicks, and clicks are how Google makes money. That, he thinks, is why it may be very hard for Google to fully launch the same system. The new paradigm is fewer clicks, each more expensive, and a faster path from question to a targeted action.

Guo asks what an era of fewer clicks does to publishers. Aravind is relatively optimistic. Citation brings PageRank's original logic back: PageRank was inspired by academic citations, and an important paper gets cited. In a citation-based interface, better content is more likely to be cited by a language model, unless people invent "answer-engine optimization" as cheap as keyword SEO — which he thinks is unlikely, because models are much better at relevance. The incentive, in his telling, looks more like Substack: own the content, make it high quality, keep your own subscribers. How that monetizes at Google's click-ad scale is still an open problem, and he is glad many companies are trying it at once.

Elad offers the historical reminder. Google's first money was selling search results, paid per thousand queries syndicated to others; then an enterprise hardware appliance; then ads. People said YouTube and Facebook would never make money. Perplexity has several cards and has not picked. Thousands of people have asked for an API, the way Bing sells search. The Chrome extension is heading toward 100,000 users, a Grammarly-like free-plus-prosumer path. At hundreds of millions of visitors, ads become possible — but they do not want Google's mistake of mixing ads into the core search product; Facebook's alternative pathway looks healthier. Subscription search is something ChatGPT is trying; if Google or Bing puts out something 80 percent as good for free, a subscription product dies, so they are likely to stay away.

A platform play is on the list too: if people want Perplexity over their own links, bookmarks, or company data, that can monetize. Short term they are focused on users, traffic, and the experience. Google and Microsoft, Aravind says, will pretty much build the same thing. The only way to live is the thing they started the podcast with: more velocity, more shipping, stay ahead on the core value of the product.

Guo has been impressed by their bias to action; most academics she knows pre-plan. Both founders share an advisor, Pieter Abbeel, who is a different kind of academic: he pushes students to get results fast. That, plus Denis's time in industry, is why they operate this way. Aravind's advice to people choosing between a PhD, an industry research role, or a company is unsparing. Being a PhD student right now is one of the worst jobs you can take. AI is hot and highly paid; you can join or start a company. Doing a doctorate means giving all that up, ignoring the daily buzz on Twitter, staying poor, and working on hard problems. It is more a mental test than a matter of picking the right problem, and very few people can do it.

He is honest about his own reason. He came to the United States for a PhD because there was no other way in: he could not take a loan for a master's, and a doctorate is fully funded. If he had been able to get an industry job, he would have taken it. If those obstacles still do not stop you, he would not spend the years improving Transformers. Existing labs have enormous incentive to make GPT's core architecture more efficient and to train even bigger models. The better bet is radical alternatives — even questioning the Transformer itself, which is the paper he would try to write. He points at Stanford work on state-space models and FlashAttention, the latter already used to make large models more efficient, and at video generation, which is still open.

Denis would add a practical step: do not go to a PhD right after undergrad. Spend a couple of years in industry. The best researchers are also very good engineers, and that experience propels the doctorate. AI has become more of an engineering field; if you want to work on large models or large-scale systems you have to be a strong engineer. If you want to stay on the purely academic side and think deep ideas, he agrees with Aravind: ignore LLMs and do something radical.