TruthArchive.ai - Related Video Feed

Perplexity aims to reinvent how we find information by offering an answer engine that grounds responses in sources, rather than simply returning links. Johnny Ho, co‑founder and chief strategy officer, explains that the product is additive to search, designed for queries that are long, multi-step, or require structured answers beyond traditional search. Ho’s background spans engineering and competition math, including IOI medals and time at Quora before returning to Harvard, and he emphasizes staying close to engineers to understand what the product must do. He describes Perplexity’s two‑track approach: synthesize an answer from sources using an appropriate language model, and provide interactive elements, follow‑ups, or actions that users can take. Under the hood, Perplexity grounds answers by offline crawling and online retrieval, then uses the right large language model for each scenario. The company has an in‑house model called Sonar and also leverages external models; latency and accuracy drive model selection, especially for coding and objective queries. A key innovation is the publisher program, which credits every source used in an answer equally, aligning incentives to surface high‑quality content rather than chase clicks. Today the business relies on subscriptions, with a path toward advertising as user value grows, while avoiding biasing answers or diluting trust. Ho explains the product’s future as a sequence of small, consumer‑friendly improvements rather than dramatic shifts. They see voice and multimodal interfaces as enhancements that still preserve source citations and enable follow‑ups, not a pure audio experience. The Discover feed and iPhone app are examples of polished, early products that came from scrappy experimentation. The team emphasizes iterative delivery, quarterly goals broken into weekly milestones, and a culture that prioritizes user motivations and real‑world usefulness over top‑line metrics alone. They acknowledge that major platforms like Google face slow responses to AI disruption, but Perplexity believes in distinct value through trust, transparency, and a publisher‑driven ecosystem.

Google's AI turnaround is real: Gemini just hit the number one app in the app store, and the internal energy at Google has changed, says Robby Stein, VP of Google Search. The company maintains that its core mission—making information universally accessible—remains, but the AI moment has created a tipping point where models can genuinely deliver for consumers. The shift is not about replacing search but about multiplying its reach through AI overviews, AI mode, and multimodal tools like Lens, all designed to deliver faster, more accurate answers while weaving live data into results. There's three big components to what we can think about AI search: AI overviews at the top, which provide quick answers; multimodal and Lens for visual search; and AI mode, which binds it all into a single conversational experience. AI mode uses all of Google's information, including 50 billion products in the shopping graph updated two billion times per hour, 250 million places in Maps, and the entire context of the web, so you can ask anything and follow up. It can be accessed at google.com/ai and is integrated into core experiences so you can ask follow-ups directly or take a photo and go deeper in AI mode. Stein emphasizes three big features of AI search: AI overviews at the top, which provide quick answers; Lens for visual queries; and AI mode, which binds it all into a single conversational experience. He notes that Google’s data backbone—shopping graph, Maps, finance, and web signals—allows the AI to understand context and surface authoritative sources. The interface aims for a consistent, simple experience; you can start in core search and have follow-ups, then dive deeper in AI mode or Lens as needed. The goal is to make the transition between AI and traditional search seamless rather than a toggle. Looking ahead, AI is expanding into inspiration and multimodal creativity, with live AI search and 'AI corner' experiments such as visual inspiration boards and Nano Banana-like tools. The team emphasizes testing with labs and trusted testers, then scaling to IO launches and global rollout. Public examples include live conversational search and ongoing integration across products, all aimed at giving users effortless access to knowledge with reliable sources.

Dualingo's mission is to provide the best education and make it universally available. From day one, Dualingo was technology first; back then it was software, and now we call it AI. We were one of the launch partners of OpenAI when they first launched Chip D4 and we immediately saw the potential of this technology to help our mission. The first insight was that AI could accelerate content production. It took us what we recall as twelve years to build the first 100 courses, and within one year we built another 148 more. That acceleration changed what we could deliver and how quickly we could scale the platform. Two things became clear about AI inside Duolingo. First, content generation is transformative; 'there's a lot of the sentence content within these courses' now produced with AI, and the curriculum design remains human-made. Second, new AI features were possible—like Lily, the interactive video call with a purple-haired character who remembers you and speaks in a natural way. Third, AI also boosts productivity company-wide. They mention ‘content generation’ changing learning content, ‘video call with Lily’ as a feature, and general productivity tools. They see personalization as the future of education and expect it to be multimodal—voice, typing, video, and dynamic, on-the-fly course design that adapts to you. On strategy and markets, the founders discuss fundraising and resilience. ‘It's harder to raise 3 million than 100 million,’ and the Series A was 3 million at 15 with Union Square Ventures leading; there was ‘one offer,’ Union Square or back to university. They contrast the push to relocate to SF with their success staying in Pittsburgh; Europe at the time wouldn't have supported similar growth. They credit Union Square for legitimacy; while marketing had been lean with ‘the green owl’ and ‘unhinged things’ that built a strong brand, they underpriced their brand—‘dramatically underprice brands’—yet built a highly efficient marketing machine. They explain why they didn't monetize early and how monetization is now essential but done with a mission to keep learning accessible. Regarding education's future, they frame higher ed as three things: instruction, credentials, and social networks, and they discuss the chess course as proof of AI-enabled development: prototype in nine months, then integrated into the app. They emphasize the secret sauce as not a single feature but a process, with retention driven by the streak. They discuss the social dimension—potential social features and Duo social, or even 'Duo dates'—and the ongoing balance between founder-driven detail and delegation. They assert that AI will augment humans, not replace them, while stressing the enduring value of motivation, guidance, and credentials in education.

In this conversation, Lex Fridman speaks with Aravind Srinivas, CEO of Perplexity, about the innovative approach of combining search engines and large language models (LLMs) to create an answer engine that provides reliable, citation-backed responses. Perplexity aims to reduce hallucinations common in LLMs by ensuring that every answer is grounded in verifiable sources, similar to academic writing. Aravind shares insights from his background as an AI researcher and discusses the challenges faced when building a startup, emphasizing the importance of understanding user needs and creating a product that genuinely helps people discover knowledge. Perplexity operates as a knowledge discovery engine, focusing on providing direct answers rather than just links to websites. This approach encourages users to explore related questions, fostering a continuous journey of learning. Aravind highlights the significance of user experience and the need for a product that evolves with user interactions, aiming to guide users toward deeper understanding and curiosity. The conversation delves into the technical aspects of Perplexity, including retrieval augmented generation (RAG), which combines search and LLMs to generate answers based on retrieved documents. Aravind explains the indexing process, the importance of maintaining a fresh and accurate database, and the challenges of ensuring the model's accuracy in understanding queries. He emphasizes the need for a hybrid approach that incorporates traditional search techniques alongside modern AI methods. Aravind reflects on the future of search and the internet, envisioning a world where AI tools enhance human curiosity and knowledge discovery. He believes that as AI systems become more capable, they will empower individuals to seek truth and understanding, ultimately leading to a more informed society. The discussion also touches on the potential for AI to form deeper connections with humans, though Aravind expresses caution about the implications of such relationships. Throughout the conversation, Aravind shares his passion for knowledge and the mission of Perplexity to facilitate learning and exploration. He believes that by harnessing AI's capabilities, we can create tools that not only provide answers but also inspire curiosity and foster a deeper understanding of the world. The overarching theme is the importance of prioritizing human curiosity and knowledge in the development of AI technologies, ensuring that they serve to enhance, rather than diminish, the human experience.

Perplexity positions itself as an answer engine that grounds AI responses in sources and aims to do more than a traditional search. Johnny Ho, co-founder and chief strategy officer, explains that Perplexity's goal is to answer long, multi-step queries with trustworthy sources, while preserving the ability to let users explore further. The conversation tracks his path from IOI medalist and competitive programmer to product and strategy leadership, and how that background informs a user-centric approach. Ho emphasizes that Perplexity's product remains additive to current search: it surfaces sources, supports follow-up questions, and offers interactive elements, rather than simply replacing Google with a single answer. The two-step model emerges: a traditional search or signal layer to collect relevant material, then an LLM to synthesize and present a concise response with provenance. Under the hood, Perplexity combines offline crawling, online retrieval, and careful selection of snippets from multiple sources to ground answers. They have an in-house fine-tuned LLM called Sonar, and they also use external models such as OpenAI's GPT family, choosing the right model for each scenario. A key focus is avoiding biasing answers and ensuring that the user can verify the origins of the information. The product envisions an evolving role for AI: the "80/20" rule means most queries get a good answer quickly, while the remaining 20 percent trigger deeper exploration, actions, or interactions that could affect real-world tasks. They discuss how speed, latency, and formatting affect adoption, especially in coding or action-driven use cases. On the business side, Perplexity monetizes through consumer and enterprise subscriptions, while also experimenting with a publisher program that credits sources used to generate an answer. The model favors a fair, non-biasing approach where publishers gain credit proportional to how often their content contributes to an answer, a concept designed to align incentives and sustain high-quality content. Ho notes the potential future of advertising but stresses user value and trust as core, with gradual experimentation rather than a radical overhaul. The team's culture centers on rapid iteration, with weekly goals and a focus on small, continuous improvements that compound into product progression.

Aravind Srinivas founded Perplexity AI about 1.5 years ago to solve a specific problem: building a great natural language to SQL tool. It wasn’t intended to dethrone Google; it focused on search over databases rather than a coding co-pilot. Early experiments included a natural language to Stripe Sigma tool and Bird SQL, a Twitter crawl whose data helped attract initial investors, including Jeff Dean; he tried their Twitter search demo and liked it, though they keep him separate due to conflicts. As models improved from GPT-3 to 3.5 and beyond, they moved to online inference, pulling links and summarizing them with citations, achieving fast, concurrent search and LLM calls, and optimizing tail latency and perceived latency via UX streaming. They grew to about 45 people with a trial-based hiring approach; cadence includes Monday plans, Friday all-hands, and standups. They see word-of-mouth growth and metrics like 10 million MAUs in 2023, aiming to 10x. Partnerships include Arc browser; approaches to ads, enterprise, and open-source models are evolving.