No PriorsNo Priors Ep. 9 | With Perplexity AI’s Aravind Srinivas and Denis Yarats
CHAPTERS
- 0:00 – 2:22
Founding Perplexity: from visual search ideas to an LLM-powered answer engine
Aravind and Denis recount the early genesis of Perplexity, including initial (rejected) ideas like camera/visual search and experiments like text-to-SQL and notebook copilots. They emphasize that “search” was the consistent underlying motivation, with generative models making a new interface possible.
- •Early founding context (Aug 2022) and first conversations with investors
- •Initial exploration: visual search, text-to-SQL, data/Notebook copilots
- •Nonlinear prototyping path that still anchored on “search” as the core thesis
- •Search framed as both a technology and distribution problem
- 2:22 – 4:38
Building a fast-iteration culture: academic experimentation meets user feedback loops
The discussion turns to why Perplexity can iterate unusually quickly. Aravind attributes it to an academic mindset of rapid experimentation, but notes that product iteration adds operational work: shipping usable prototypes and collecting real user/customer feedback.
- •Academic roots: many ideas, fast experiments, quick iteration cycles
- •Product iteration requires shipping, distribution, and direct user feedback
- •Engineering speed + user learning loops as a compounding advantage
- •Early team composition as a force multiplier for execution velocity
- 4:38 – 5:38
Hiring for intensity and learning speed: curiosity over credentials
Denis describes what he optimizes for when hiring: strong intrinsic motivation to learn and apply LLM technology, even without prior pedigree. The team prioritizes small size, high output per person, and alignment with a demanding pace.
- •Primary hiring signal: curiosity and “burning desire” to work on LLMs
- •Preference for effort and learning velocity over prior domain experience
- •Keeping headcount small to maintain speed and coherence
- •Need for explicit alignment on pace/work intensity to compete
- 5:38 – 7:49
Trial periods as a hiring filter: de-risking fit for both sides
Perplexity uses extended “trial” work periods to validate collaboration before committing to a hire. Denis contrasts this with standard interview loops and argues trials reduce surprises while giving candidates a realistic view of expectations.
- •Trial periods used after initial confidence to validate real collaboration
- •Reduces mismatch risk compared to traditional short interview loops
- •Creates mutual signal: candidates assess the team’s pace and style
- •Acknowledges occasional surprises but fewer than typical processes
- 7:49 – 9:01
Why speed is a startup’s edge—and why focus is the hard part
Aravind and Denis argue speed is often the only durable startup advantage versus incumbents. They also describe the early mistake of pursuing too many directions at once, and why startup leadership requires clearer focus than academia’s hedging style.
- •Speed as a non-negotiable competitive advantage when direction is uncertain
- •Early over-breadth: many integrations and prototypes that never shipped
- •Leadership clarity matters—thrash erodes team faith
- •Academia encourages hedging; startups demand focused iteration on fewer bets
- 9:01 – 13:17
You don’t need an “AI pedigree”: great engineers and cross-domain talent win
They challenge the belief that only people with prior LLM training can contribute to AI companies. Using examples from OpenAI/Anthropic and their own cofounder Johnny Ho, they argue strong engineers and scientists from other fields can ramp quickly and often outperform in product contexts.
- •OpenAI/Anthropic examples: physicists and strong engineers driving breakthroughs
- •Engineering excellence as a differentiator in modern AI progress
- •Product-building with LLMs is distinct from training them via gradient descent
- •Rejection of exclusivity culture around PhDs; capability > credentials
- 13:17 – 16:34
Transitioning from research to running a company: learning management, prioritization, and self-correction
Sarah asks about the steepest learning curves moving from research/engineering into leadership roles. Aravind highlights learning to run a company and becoming brutally honest about mistakes; Denis emphasizes prioritization, organization, and choosing the single most important thing.
- •Aravind: learning CEO skills via advice, feedback, and rapid course correction
- •Denis: team organization, prioritization, and focus as critical in small teams
- •Avoiding parallel overreach; do fewer things, iterate deeper
- •Leadership accountability: clarity and sequencing matter more than exploration
- 16:34 – 17:35
Citation-first product philosophy: only answer what can be cited
Perplexity’s trust thesis is grounded in a research-paper norm: only state what you can cite. Aravind distinguishes Perplexity from chatbots that add citations afterward—Perplexity is designed to avoid making claims without sources, even at the cost of “personality.”
- •Core rule: don’t say anything you can’t cite
- •Not “retrofit citations into a chatbot”—citations are first-class in the system
- •Trustworthiness prioritized over bot character/personality
- •System behavior: retrieves sources even for questions typical chatbots answer directly
- 17:35 – 19:25
Handling aggregation errors and bias: source control and user correction loops
They discuss failures that arise when combining multiple sources, such as entity conflation and hallucinated mashups. Proposed mitigations include improved model understanding over time and giving users tools to remove irrelevant sources—while noting incentive and abuse risks.
- •Aggregation risks: conflating entities with similar names; mixed-source hallucinations
- •No single “better LLM” is sufficient—needs end-to-end product design
- •User controls to curate/remove sources (Wikipedia-like), plus governance concerns
- •Acknowledgement that trust/bias handling will require many iterations
- 19:25 – 20:56
Reinforcement learning at Perplexity: RLHF, rating pipelines, and rejection sampling
Sarah probes their reinforcement learning approach. Aravind outlines RLHF-style feedback loops using user/contractor ratings (and potentially LLM-assisted labeling), while Denis describes intermediate techniques like rejection sampling and ranking multiple candidate outputs to boost quality efficiently.
- •RLHF framing: collect preferences on summaries/completions to improve outputs
- •Contractor ratings and emerging LLM-assisted labeling workflows
- •Near-term focus on quality improvement vs. longer-horizon agentic browsing
- •Rejection sampling/reranking as a pragmatic “in-between” step to raise answer quality
- 20:56 – 22:41
Chat vs. search UX: conversational follow-ups with a ‘get it right fast’ ethos
They argue chat is the future interface, but also stress responsibility to answer accurately on the first attempt when possible. The chat format helps users refine poorly formed queries via follow-ups, while simple queries should be satisfied immediately without extra interaction.
- •Chat UI seen as the long-term default for complex information needs
- •Responsibility to minimize user time: aim for first-try correctness where feasible
- •Chat helps users discover/clarify what they really meant to ask
- •Google already approximates conversation via related questions; chat makes it explicit
- 22:41 – 26:02
The future of search: answer engines as a distinct market and the ‘fewer tabs’ shift
Aravind predicts “answer engines” will become a standalone category, offering direct answers with fewer links and more follow-up behavior. He also forecasts more action-oriented outcomes (booking, executing tasks) and a structural reduction in traffic to content sites as users consume fewer links.
- •“Answer engine” framing: direct answers instead of link lists
- •Answer engines as a new market segment with default providers
- •Behavior shift: more follow-ups; fewer opened tabs; fewer cited links by design
- •More actions integrated into results (transactions, assistants)
- 26:02 – 28:33
From pull to push: agentic, intent-aware information delivery—and where monetization breaks
They discuss a move toward push-based assistance (agents anticipating needs), arguing it’s feasible soon and improved by LLMs. Denis highlights a tension for incumbents: answer engines reduce clicks, challenging ad-based models, implying monetization must evolve toward fewer but higher-value clicks.
- •Push model precedents (e.g., Google Now) and expectation of near-term adoption
- •Question of fragmentation vs. consolidation across personal/work tools (Drive, GitHub, email)
- •‘Fewer clicks’ creates monetization pressure for Google’s current model
- •New ad paradigm: fewer clicks, more targeted, higher value per click
- 28:33 – 33:59
Publishers in a citation-based world and Perplexity’s monetization options
They argue fewer clicks won’t eliminate publishing incentives; instead, citation-driven discovery may reward higher-quality content, echoing academic citation dynamics and PageRank’s origins. Aravind then lays out possible monetization paths—API access, prosumer tiers via extensions, ads that don’t corrupt core UX, and enterprise/internal-data offerings—while emphasizing near-term focus on growth and product quality.
- •Citation-based incentives could favor higher-quality content over SEO keyword gaming
- •PageRank analogy: important content earns citations; LLM relevance may be harder to game
- •Monetization ideas: API offering, prosumer subscription features, enterprise/internal data platform
- •Caution on ads in core UX; uncertainty about subscriptions if incumbents offer ‘good enough’ for free
- 33:59 – 39:10
Advice to researchers: the reality of PhDs during AI booms and choosing radical problems
They close with career advice for researchers weighing academia, industry, or startups. Aravind notes the mental difficulty of staying in a low-paid PhD during an AI boom and recommends working on radical directions (even beyond transformers), while Denis advises spending time in industry first to build engineering strength—now essential for impactful AI research.
- •PhD as a demanding tradeoff amid high startup/industry opportunity and pay
- •Best academic leverage: controversial/radical directions beyond incremental LLM work
- •Engineering skill increasingly central to modern AI research impact
- •Suggestion: do industry first, then PhD; or pursue efficiency/alt-architectures (e.g., state space models, attention innovations)
