For five hours in August 2022, Denis Yarats and Aravind Srinivas waited outside Yann LeCun's office in lower Manhattan. They skipped lunch. The thing they hoped to show him was not yet Perplexity, at least not the Perplexity that millions would later recognize. It was a piece of social search, clever enough to rummage through LinkedIn, GitHub and Twitter and answer the sort of question that makes a demo suddenly personal: who replies to your posts most often?
LeCun eventually opened the door. The demo ran. He invested. The scene has the furniture of startup folklore, but its most revealing object is the waiting itself. Yarats had spent years becoming the sort of engineer who could make an idea executable. Once it existed, he was prepared to sit in a hallway until someone important saw it.
His route there began far from Manhattan and Silicon Valley. Born Dzianis Yarats in Gomel, Belarus, he studied at Belarusian State University and entered the competitive-programming circuit as frost_nova. In 2011, he joined Raman Udavichenka and Yury Pisarchyk at the ACM-ICPC World Finals in Orlando. Their team name, “Kuplyu Moped!”, translates roughly as “I'll buy a moped!” They placed 30th. It is difficult to imagine a more useful apprenticeship for a future startup CTO: a clock, a stubborn problem and no committee to ask for an extension.
A career built in both directions
Yarats went first to Microsoft, where he worked on Bing from 2011 to 2013. Then came Quora, still a compact company, where he worked on machine learning and ranking and met Johnny Ho, an engineer he would later recruit as a Perplexity co-founder. The jobs put him close to a problem that appears simple only to people who have never touched it: how do you rank a sea of possible information quickly enough that the useful item feels inevitable?
In 2015, a machine-learning conference convinced him that deep learning would matter enormously. He joined Facebook AI Research the next year, initially as an engineer, and pursued doctoral research at New York University. The transition did not erase his earlier identity. It enlarged it. Research taught him how to set a problem; engineering had taught him how to make the answer run.
His papers often carried this practical accent. DrQ applied straightforward image augmentation to help reinforcement-learning agents learn directly from pixels. DrQ-v2 improved the method and reported strong visual-control results with an unusually modest computational footprint. Most tasks, the paper said, could train in eight hours on one GPU. The idea had rigor, code and a receipt for the electricity.
The public code mattered. Yarats's GitHub still resembles a compact museum of the period: Soft Actor-Critic implementations, an autoencoder variant, DrQ, DrQ-v2 and Proto-RL. Researchers could inspect the argument at the level where an elegant diagram meets an obstinate dependency. Some repositories collected hundreds of stars and forks. More important, they gave other people a runnable baseline. Yarats's later enthusiasm for open models follows the same instinct. A technique becomes more valuable when strangers can test it, find its weak spots and carry it somewhere its authors did not anticipate.
“It's trying things yourself rather than... read a paper and... ‘this makes sense.’”Denis Yarats on learning by implementation
One of those papers produced an unforeseen result. In 2020, Yarats at NYU and Srinivas at Berkeley released closely related reinforcement-learning work within days of each other. They connected by email. The overlap could have created rivalry; instead it created a conversation. They stayed in touch while Yarats worked at FAIR and Srinivas moved through DeepMind, Google and OpenAI.
The useful wrong turns
When the pair decided to build a company in 2022, they added Ho and Andy Konwinski, the Databricks co-founder who helped get the venture moving. They did not stride directly toward an answer engine. They tried a Jupyter Notebook assistant. They explored text-to-SQL. Their Bird SQL prototype translated plain English into code and searched Twitter for relationships that ordinary web search missed. It attracted attention, but then Twitter shut off free API access.
Next came enterprise search across tools such as Salesforce and HubSpot. The internal data was poorly documented, the models were not ready, and every customer threatened to require its own detachment of engineers. The founders found the work painful. A wrong turn is irritating when one has driven slowly. They were moving fast enough to turn again.
ChatGPT's arrival made a general answer product newly legible, while its factual mistakes made a particular feature urgent. Yarats and Srinivas came from academic research, where a naked claim is an invitation to a stern reviewer. Citations felt natural. Perplexity joined web retrieval to generated prose so a reader could see the supporting links rather than accept a fluent paragraph on charm alone.
Citations did not magically settle truth. Pages disagree. A ranking system can retrieve a weak account. A language model can attach a source to a sentence that the source does not quite support. Yarats has described the solution as a company-wide mindset rather than one clever component: better models, better ranking, multiple viewpoints when a dispute cannot be resolved, self-verification and a data loop that learns from mistakes. The apparent simplicity of a cited paragraph is therefore a design promise, not proof that the technical problem has retired.
The interface concealed an awkward orchestra: retrieval, ranking, language models, latency, source disagreement and verification. Yarats has praised Google search as perhaps the most sophisticated system humanity has built. His argument is not that the old system was trivial. It is that a different product can remove some of the work between a question and an answer. Respect for the mountain need not forbid a new path up it.
Small things, done today
Perplexity's early culture carried Yarats's habits into the organization. The company hired slowly at first. Candidates sometimes worked beside the team for a week, long enough for ambiguity to arrive and interview polish to expire. Yarats looked for coding strength, curiosity and the ability to learn an adjacent field. If the fit was clear, the evaluation quickly became recruitment.
The structure was strikingly flat. In a 2024 interview, Yarats said that a roughly 65-person engineering organization had only just hired its first engineering manager. He and Ho functioned more like technical leads. The arrangement was unusual and plainly suited to a particular phase, but its logic was consistent: hire people who can find the next useful task, then remove whatever slows the work.
There was another advantage to hiring for range. Perplexity could not outspend larger laboratories for every established researcher. Yarats instead looked toward strong engineers, mathematicians and competitive programmers who could acquire the missing AI knowledge. He rejects the tidy border between researcher and builder. In his account, the valuable person can form a hypothesis, write the system that tests it and recognize when the result has contradicted the theory. It is a demanding profile, but also a democratic one: reputation matters less than the speed and depth with which someone can learn.
“If you can do something today, you have to do it today rather than doing it tomorrow next week.”Yarats on maintaining startup momentum
Speed, in his vocabulary, is paired with accuracy. Search tolerates neither a leisurely answer nor a confident fiction. That tension explains his attraction to systems that can retrieve multiple accounts, attribute their claims and check their own output. It also explains his support for open-source models. A large community can discover bugs and efficiency tricks that no single company will find alone. The user's task, however, should remain above the machinery. Most people do not want to select a model. They want the job finished.
By 2025, Yarats described the job in broader terms. Quick answers, Pro Search, Deep Research, Labs and the Comet browser were steps toward one experience. A person could ask a simple question or a complicated one. The system would scale the effort, reach the permitted public or private information, choose a useful form for the result, help with a decision and eventually take action. Search would become less like receiving a map and more like completing the errand.
The quiet idea beneath the answer
There is a temptation to tell this story as a contest with Google. Yarats himself is less theatrical. He notes that people can use Google, ChatGPT and Perplexity for different tasks. A new category does not require an old one to vanish. It requires a sufficiently useful reason to exist.
His own career makes the case in miniature. Competitive programming did not disappear when he became an engineer. Engineering did not disappear when he became a scientist. Research did not disappear when he became a founder. Each layer made the next one more useful. At Perplexity, he could set a direction, then sit beside an engineer and unblock the implementation. The abstraction never became an alibi.
That may be why the five-hour wait remains so apt. The future did not arrive while two founders stood in a hallway. A door merely opened on something they had already built. Yarats's aspiration now is larger than the original demo, but the method remains almost comically plain: choose a tractable piece, go deep, test it, cite what deserves citation, and do today what would be easier to schedule for tomorrow.