The conversation revolved around the potential of AI-powered search and summarization tool, perplexity, to revolutionize search and knowledge discovery. Speakers discussed the importance of sourcing and citation in generating accurate answers and the limitations of perplexity. They also explored the challenges of personalized information retrieval and presentation, and the qualities of successful AI founders. Additionally, they delved into the intersection of technology and humanity, emphasizing the potential benefits and risks of emerging technologies and the importance of catering to human curiosity.
OUTLINE: 0:00 – Introduction 1:53 – How Perplexity works 9:50 – How Google works 32:17 – Larry Page and Sergey Brin 46:52 – Jeff Bezos 50:20 – Elon Musk 52:38 – Jensen Huang 55:55 – Mark Zuckerberg 57:23 – Yann LeCun 1:04:09 – Breakthroughs in AI 1:20:07 – Curiosity 1:26:24 – $1 trillion dollar question 1:41:14 – Perplexity origin story 1:56:27 – RAG 2:18:45 – 1 million H100 GPUs 2:21:17 – Advice for startups 2:33:54 – Future of search 2:51:31 – Future of AI
Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet | Lex Fridman Podcast #434
Action Items
- [ ] Improve retrieval of snippets from the index to reduce hallucinations.
- [ ] Explore techniques like self-supervised post-training and chain of thought reasoning to improve the model's abilities.
- [ ] Consider how to integrate ads in a way that doesn't interfere with seeking truth or the user experience.
Outline
AI-powered answer engine perplexity, combining search and LLMs for accurate and well-sourced answers.
- Speaker 1: perplexity is an answer engine that sources answers with LM.
- Speaker 2: perplexity combines search and LLM for citation-backed answers.
- Speaker 1: Orchestrated product built to cite internet sources accurately.
- Speaker 2: Forced chatbots to only say things backed by citations from multiple sources.
- Speaker 1: Integrated Slack bot, pinged GPT, but accuracy was a concern.
- Speaker 2: Viewed perplexity as a knowledge discovery engine, not just a search engine.
AI-powered search engine Perplexity and its capabilities compared to Google.
Perplexity AI provides direct answers, while Google focuses on links.
Speaker 1: Integrating real-time information into conversational AI is challenging.
Speaker 2: Custom UIs for specific queries are needed for optimal user experience.
Speaker 1 suggests using personalization to provide more accurate weather information, such as tailoring recommendations based on a user's habits and interests.
Speaker 2 agrees, highlighting the importance of understanding a user's preferences and habits to deliver a more personalized experience.
Google's search engine and advertising model, with a focus on perplexity's unique approach.
- Speaker 2 suggests perplexity's unique approach to search disrupts Google by rethinking UI, not just improving upon it.
- Speaker 1 bets on technology improving over time to provide better, cheaper, and more efficient search results.
- Google's search advertising revenue is based on bidding for AdWords, with high ROI leading to increased spending.
- Speaker 2 praises Google's sponsored links for providing good brands and avoiding trickery.
- Speaker 1 explains how Google benefits from advertising done elsewhere on the internet.
Google's ad system and its potential for perplexity.
- Speaker 2: Google's ad system is data-driven, but not all parts of the industry are (Speaker 2).
- Speaker 1: Google optimizes multiple objective functions simultaneously in its ad system (Speaker 1).
- Speaker 2 discusses perplexity in ad units and how it differs from Google's business model.
The business strategies of Amazon and Google, including their approaches to advertising, cloud computing, and AI.
- Speaker 1 highlights the programmatic reality of how companies are run, focusing on margin as opportunity.
- Speaker 1 argues that the growth of cloud and YouTube revenue reduces reliance on advertisement revenue, with potential for hybrid subscription and advertising models.
- Speaker 1: Integrating AI into perplexity could work on all fronts, but must be done carefully to maintain user trust.
- Speaker 2: There are ways to manipulate perplexity's output, such as injecting hidden text or using answer engine optimization.
Google's early days, PageRank, and academic roots.
- Page and Brin's innovation: ignore text, rank by link structure.
- Larry Page prioritized hiring PhDs to build core infrastructure, focusing on latency and compatibility.
Improving search engine user experience through prompt engineering and reducing latency.
- Speaker 1 emphasizes the importance of reducing latency in software products, citing Spotify as an example of a successful product with low latency.
- Speaker 1 and Speaker 2 discuss the user-centered philosophy of Google, including the idea that the user is never wrong and that products should always provide high-quality answers to user queries.
- Speaker 1: Product should be designed for lazy users, allowing them to be more lazy.
- Speaker 2: Sequence of questions is important, auto-suggest bar can help predict user intent.
Designing a product with minimal features while balancing user needs and growth.
- Designers debate simplicity vs. complexity in product design.
- Speaker 1: Balancing new user growth with retaining existing users is challenging, as new features for power users can confuse new users.
- Speaker 2: Silent frustration from confused or frustrated users can be difficult to identify and address.
Entrepreneurship, leadership, and innovation with insights from successful founders.
- Speaker 1 admires Jeff Bezos' ability to gain clarity of thought and make decisions efficiently.
- Speaker 1 values Bezos' focus on operational excellence and customer obsession.
- Speaker 1 highlights the importance of customer obsession and direct distribution in business success.
- Speaker 1 admires Elon Musk's ability to ignore criticism and push through with determination and first principles thinking.
The importance of details and planning in hardware development.
- Speaker 1: Importance of understanding every detail for good business decisions.
- Speaker 2: Questioning conventional wisdom and simplifying systems for efficiency.
- NVIDIA's CEO is paranoid about going out of business, planning 10-20 years in advance.
AI research and innovation, including open source models and unsupervised learning.
- Speaker 1: Zuckerberg's open-sourcing of AI models gives hope for more players in the field.
- Speaker 1: Lectin's slab has produced many next-generation AI researchers, including Koray and Sora.
- Speaker 1: RL is not the real deal, most compute should be spent on learning from raw data.
- Speaker 2: Auto regressive models might be a dead end, reasoning in latent spaces could be powerful.
Open-source AI, safety concerns, and the importance of transparency and guardrails.
- Speaker 2 argues that open source is the only way to keep AI safe, despite potential risks of malevolent actors.
- Speakers discuss the potential of transformer models and attention mechanisms in AI development.
Transformer architecture evolution, attention, parallel computation, and scaling up models for better performance.
- Speaker 1: Attention and transformers make better use of hardware than cons, applying more compute per flop.
- Speaker 1: Masking is a clever insight that allows parallel compute during training, reducing training time significantly.
- Insight from the history of transformer models: scaling up RNNs with transformer architecture and training on large datasets led to breakthroughs in natural language processing.
- Importance of RHF (reinforcement learning from human feedback) in controlling and improving transformer models, as demonstrated by GPT-3 and other recent advancements.
Improving language models' reasoning skills through post-training techniques.
- Speaker 1 argues that pre-training and post-training phases are crucial for AI models, with post-training focusing on fine-tuning and improving product quality.
- Researchers explore training small language models on specific datasets for improved reasoning skills.
- Researchers find that forcing AI models to explain their reasoning improves their performance on NLP tasks.
Using natural language explanations to improve AI reasoning abilities.
- Speaker 1 proposes using natural language explanations as a latent to train models for reasoning and math abilities.
- Speaker 1 believes this approach could lead to more capable agents, but corner cases may remain a challenge.
AI's potential to mimic human curiosity and reasoning abilities.
- Speaker 1 and 2 discuss the potential for AI systems to self-improve and achieve intelligence explosion through recursive self-improvement.
- They speculate on the need for human interaction and signal provision to facilitate this process, and the potential implications of such an experiment.
- Speaker 1: Humans' natural curiosity is key to AI breakthroughs.
- Speaker 1: Mimicking Fineman's curiosity, not just research skills, is crucial.
AI's potential to revolutionize drug discovery and the ethical implications of its power.
- Speaker 1: AGI could be compute-limited, not just data-limited.
- Speaker 2: AGI could be like a “nerdy, rapid, iterative thinking” that needs compute.
- Speaker 1: Value of transformer-like AI breakthrough could be trillions of dollars.
- Speaker 1: Breakthroughs in AI may be limited by clever use of iterative compute.
AI's potential to provide new knowledge and challenge human understanding.
- Speaker 1 discusses the potential for AI to arrive at new truths and challenge current understanding in physics and other fields.
- Speakers discuss the potential of AI to provide groundbreaking insights, but also the importance of considering link structure in addition to text patterns.
- Speaker 1: AI insights needed to uncover COVID origins, avoid politicization.
AI's potential to revolutionize truth seeking and problem-solving.
- Elon Musk and others may benefit from AI-assisted innovation, potentially leading to groundbreaking discoveries.
- AI Ilya and Andrei could create a million copies of each, leading to interesting echo chambers or new truth seeking.
Using AI to improve search experiences, including generating code and querying databases.
- Speaker 1: AI complete products like GitHub Copilot and Google Search benefit from advances in AI, creating a flywheel effect.
- Speaker 1: Self-driving cars and other products that use more data to improve AI models are examples of AI completeness.
- Speaker 1 discusses disrupting search experiences by querying databases directly through natural language questions.
- Speaker 2 clarifies that previous attempts at querying databases required complicated SQL, but new models can translate questions into SQL and run them against the database.
Using AI to improve search functionality on Twitter.
- Speaker 1 and co-founders used GitHub Copilot and SQL to build a question answering bot for Twitter, focusing on AI-related tweets.
- They demoed the bot to influential people in the AI field, including Yan, Lacan, Jeff, Dean, and Ray, who liked it.
- Speaker 1: People search about themselves on social media, even though they can easily access their own profiles.
- Speaker 1 and 2: Perplexity's initial growth came from people sharing screenshots of their search results, which showed interesting and sometimes hilarious insights about themselves.
Creating a knowledge-centric company with a mission to help people discover new things.
- Speaker 1 and co-founders aimed to create a conversational search product with suggested questions, growing usage exponentially.
- Speaker 1: Obsessing over knowledge and curiosity, not just search or answering questions.
AI-powered text generation and its limitations.
- Speakers discuss rag, a framework for retrieval-augmented generation, which retrieves relevant documents and paragraphs to write answers for queries.
- Speaker 1 explains how a model can hallucinate in a conversational AI system, citing three ways it can happen: poor snippets, insufficient information, and too much detail.
- Speaker 2 agrees and adds that the model can also go wrong by referencing a transcript of this podcast, potentially leading to irrelevant answers.
- Speaker 1 explains how Google's crawler works, including decision-making processes for crawling web pages.
Search algorithms and ranking signals for web pages.
- Speaker 1 explains how Google's ranking system works, including the crawling, building, and indexing processes.
- Speaker 2 highlights the challenges of converting raw web content into a vector database for ranking, including disentangling different dimensions and capturing meaningful semantics.
- Speaker 1 emphasizes the importance of combining embeddings with traditional retrieval methods for effective search.
- Speaker 1: LM helps improve search accuracy despite poor queries.
AI models for language processing, focusing on latency and throughput.
- Speaker 1: “Precision and recall are important, but so is flexibility in model training.”
- Speaker 2: “Sonar is a pre-trained LM that can be swapped out for different models.”
- Speaker 1 discusses the importance of tracking tail latency in a system, citing a paper by Jeff Dean.
- Speaker 1 highlights the collaboration with Nvidia on optimizing kernel-level throughput for llama-based models.
Scaling compute for AI models, with focus on AWS and Google Cloud.
- Netflix uses AWS for nearly all computing and storage needs, with over 1000 server instances and a virtual studio in the cloud.
- Speaker 1 argues that using AWS infrastructure can help startups scale quickly and recruit engineers.
Starting a company, founder advice, and coping with the sacrifices of entrepreneurship.
- Speakers discuss the importance of compute power for AI research and development.
- Start with an idea you're passionate about, as it will give you the strength to persevere.
- Understand your own dopamine system to find the right product fit.
- Founder's support system crucial for success in startups.
Entrepreneurship, hard work, and passion.
- Speaker 1: New insights & fresh perspectives are key to being a good founder.
- Speaker 2: Passion & sacrifice are necessary for starting a successful startup.
- Speaker 1 advises young people to focus on their passions and explore their interests early on in life, as it sets the foundation for future success.
- Speaker 1 and Speaker 2 discuss the importance of work-life balance, with Speaker 1 emphasizing the value of hard work and Speaker 2 highlighting the need for empathy and supportive relationships.
The future of search, internet, and knowledge discovery.
- Speaker 1: Messi is goat, but Ronaldo's journey inspires (12 words)
- Speaker 2: Understanding human journey is key to success (14 words)
- Speaker 1 predicts future of search and internet will prioritize knowledge discovery over search and answer engines.
- Speaker 1: Empowering people to seek truth, fact-check, and learn from each other's experiences.
- Speaker 1: Creating a platform for people to share knowledge, like a search engine for understanding.
AI, knowledge sharing, and the importance of catering to human curiosity without drama.
- Speaker 2 discusses Perplexity Pages, a tool for organizing and sharing knowledge, and how it can be used to create a Wikipedia-style page.
- Speaker 1 emphasizes the importance of understanding the mission of Perplexity Pages, which is not about changing the search, but about making people smarter and delivering knowledge.
- Speaker 2 finds Wikipedia fascinating, mentions Drake Equation and bias in Wikipedia.
- Speaker 1 agrees, highlights importance of catering to human curiosity without drama in social apps.
AI-powered language model and its capabilities.
- Speaker 1 and 2 discuss the challenge of maximizing engagement while avoiding the “delicious temptation” of prioritizing views and clicks over quality content.
- Speaker 2 seeks to maximize curiosity and provide wildcard, on-topic, and beginner-friendly content, while Speaker 1 aims to start with a simple solution and gradually increase the complexity of the LLM.
- Speaker 1 seeks help understanding investing terms, wants simple explanations.
- Speaker 1 and 2 discuss the potential of increasing context window size in AI models.
AI's potential to enhance human life, including curiosity, memory, and emotional connections.
- Speaker discusses potential benefits of AI systems with memory capabilities.
- Speaker 1: Curiosity makes humans special, and AI can enhance it.
- Speaker 2: AI can be like deep friends, empowering relationships with humans.
- Speaker 1 hopes that with AI, people can pursue their interests and build true connections, leading to a more fulfilling life.
- Speaker 2 mentions Maslow's hierarchy of needs, suggesting that AI can help address basic needs before focusing on personal growth.
AI, human connection, and the future of consciousness.
- Speaker 1: AI should prioritize long-term human flourishing, not just short-term dopamine hits.
- Speaker 2: AI coaches could help individuals become better human beings, not just provide temporary emotional support.
- Speakers discuss the importance of curiosity and knowledge in preserving human consciousness in a technologically advanced future.
- AI can help bridge understanding gaps between people with different perspectives, reducing biases and fostering empathy.
