Introduction

> When it comes to AI, the potential for deep reasoning breakthroughs lies in the ability to enhance inference compute over time. The idea of AI disappearing and returning with mind-blowing answers after thorough research could mark a significant advancement in real reasoning capabilities.

> By combining search and large language models to provide answers with citations to human-created sources on the web, we can reduce hallucinations in AI responses and make information more reliable for both research purposes and casual exploration. This approach not only enhances user experience but also paves the way for innovation in machine learning and information retrieval techniques.

How Perplexity works

> - Perplexity is designed as an answer engine that leverages a combination of search engine capabilities and a large language model to provide well-formatted answers backed by sources, similar to how an academic paper is structured.

> - The concept behind Perplexity is not just about getting an answer, but about initiating a journey of knowledge discovery where users can explore related questions and expand their understanding, making it more than just a search engine but a knowledge discovery engine.

How Google works

> - Pro Search's structured approach with steps like comparing Perplexity and Google reflects a methodical procedure.

> - Perplexity excels in direct answers and summaries but falls short in accuracy and speed compared to Google, especially for simple navigational queries.

> - The challenge in search lies not just in advancing models but in custom UI presentation tailored to user intent, as seen in offering real-time information and enhancing user experience.

> - Perplexity's focus on building a unique UI rather than directly competing with Google's existing model reflects a strategic approach to differentiate and innovate in the search space.

Larry Page and Sergey Brin

> The key insight from Google's founders, Larry Page and Sergey Brin, was their unique approach to search engines by focusing on link structure for ranking signals rather than traditional text-based similarity.

> Larry Page's emphasis on hiring PhDs for core infrastructure and research at Google, along with a relentless focus on user experience and latency, inspired a similar discipline in building products like Perplexity.

> Admiring Google's philosophy that "the user is never wrong", the guest emphasizes the importance of understanding user intent, reducing user effort, and predicting user needs to create a seamless and magical product experience, balancing simplicity with power-user features and growth challenges.

Jeff Bezos

> One key point I shared was how I draw inspiration from various entrepreneurs. For instance, from Bezos, I learned the importance of clarity of thought and the value of occasionally documenting strategies for gaining clarity, even if not for sharing extensively.

> Another key insight I discussed was the significance of relentless focus on operational excellence and customer obsession seen in successful founders like Bezos. This approach to clarifying decisions, focusing on problem-solving over minor details, and prioritizing user experience is crucial for long-term success in any venture.

Elon Musk

> - Learning from Elon Musk's approach, I find that focusing on distribution in business and maintaining a direct relationship with users is crucial. Avoiding reliance on intermediaries for distribution can lead to success, even if it may seem challenging to achieve critical mass.

> - Elon Musk's emphasis on first principles thinking and willingness to delve into details, such as personally doing data annotation for Tesla's autopilot system, highlights the importance of understanding every aspect of the business to identify bottlenecks and simplify processes. Constantly questioning why things are done in a certain way can lead to breakthrough innovations and more efficient systems.

Jensen Huang

> One key insight I shared was about Jensen's approach to leadership, questioning conventional wisdom, and his long-term strategic planning spanning decades, which sets him apart in the industry.

> Another important point discussed was the necessity for meticulous planning and attention to detail in the hardware industry, highlighted through the example of Nvidia GPUs and the impact of potential mistakes on future competitiveness and success.

Mark Zuckerberg

> I think Zuckerberg's work in open source is amazing. It's great that Meta is leading in AI and sharing high-quality open-source models like Llama 3 70B. It's important for more players to have access to advanced models, not just a few big companies.

> Zuckerberg's success in open-sourcing models like Llama 3 70B can help level the playing field in AI. It's crucial for the success of many others in the industry, bringing about a world with more diverse and capable players.

Yann LeCun

> Yann LeCun has been instrumental in shaping the field of AI, not just through his own contributions like ConvNet and self-supervised learning, but also by educating and mentoring a generation of top scientists who have gone on to do great work, like Koray, Wojciech Zaremba, and the inventors of DALL-E and ChatGPT. His early insight that most of the intelligence lies in unsupervised learning, with supervised learning as the icing on the cake and RL as the cherry, has had a profound impact on the direction of AI research, setting the stage for models like ChatGPT.

> Yann LeCun's controversial ideas about the limitations of autoregressive models, advocating for latent space reasoning and emphasizing the importance of open source for AI safety, have sparked important conversations in the AI community. His push for reasoning in an abstract representation and the belief that open source is crucial for transparency and security in AI development highlight the innovative and forward-thinking perspectives he brings to the field.

Breakthroughs in AI

> The evolution of models like Transformers and LLMs has been driven by key insights like attention and masking, enabling more efficient parallel computation during training, leading to significant advancements in natural language processing.

> Post-training phases like RLHF (Replay, Learn, and Human Feedback) play a crucial role in fine-tuning models for specific tasks, ensuring controllability and behavior, essential for building effective and usable AI products.

> Innovations like the STaR paper, focusing on bootstrapping reasoning with a chain of thought, demonstrate the importance of forcing models to engage in reasoning pathways to avoid overfitting, improve performance on NLP tasks, and potentially lead to the development of more intelligent agents in the future.

Curiosity

> - Achieving an intelligence explosion through self-supervised post-training is a possibility, with the idea of interacting with humans periodically to gain intelligence and improve being a key concept.

> - While AI advancements show potential for reasoning breakthroughs, the innate human quality of curiosity, especially the natural curiosity to explore and seek answers without external rewards, remains a unique aspect yet to be fully captured in AI development.

$1 trillion dollar question

> The future of AI involves unlocking its potential through iterative compute, leading to groundbreaking discoveries and solutions. It's less about pre-training or post-training and more about the rapid, iterative thinking that AI systems can engage in with sufficient inference compute.

> The ultimate value of AGI lies in its ability to create new knowledge and truths, surpassing human limitations in understanding complex and controversial topics. The goal is to move beyond debates and ideologies towards a deeper sense of truth, potentially revolutionizing fields like medicine and engineering.

> The ideal scenario involves AI systems assisting top thinkers like Elon Musk or experts in various fields, offering fresh perspectives and insights to drive innovation and groundbreaking discoveries. The challenge lies in balancing multiple copies of such AI systems to avoid echo chambers and promote diverse perspectives without hard-coding curiosity.

Perplexity origin story

> - Our journey with Perplexity started with a desire to harness the power of LLMs to create user-facing applications, inspired by the success of GitHub Copilot. It marked a shift where AI itself became the focal point of companies, not just a tool for data collection.

> - We aimed to create a search experience where users could query relational databases in natural language using SQL generated by our models. The initial focus on Twitter search opened doors to investors and prominent figures, showcasing the magic of offering something previously impossible and practical.

> - Transitioning from Twitter search to web search, we stumbled into the vision of becoming the world's most knowledge-centric company, driven by curiosity and the mission to guide users towards discovering new things, inspired by Amazon's customer-centric approach and aiming for a purpose beyond simply competing with giants like Google.

RAG

> In Perplexity, we ensure factual grounding by only using information retrieved from documents, which helps to reduce hallucinations in the generated answers.

>

> Hallucinations can seep in due to model skill issues, poor snippets in retrieved documents, excessive detail overload in the model, or retrieval of irrelevant documents, highlighting areas for improvement in the system.

>

> The indexing process involves crawling the web, rendering modern web pages, respecting politeness policies, processing content for ranking systems, and utilizing BM25 algorithm for effective information retrieval.

>

> While aiming for lower latency, maintaining tail latency through optimizing components like search and LLM layers is crucial, requiring decisions on scaling compute resources, model providers, and the trade-offs between in-house and cloud solutions, which impact system performance and user experience.

1 million H100 GPUs

> - The future of AI development and competition may rely heavily on powerful hardware, with a focus on winning through computational capabilities. However, achieving breakthroughs in AI that decouple reasoning and facts could lead to more efficient models that do not solely depend on massive clusters of GPUs.

>

> - Exploring ways to represent knowledge in a more efficient and abstract manner, while making reasoning iterative and parameter decoupled, could revolutionize the field of AI and reduce the necessity for extremely large GPU clusters in the future.

Advice for startups

> - Starting a company requires traits like relentless determination and grit, but it's crucial to work on something you are truly passionate about, not just what you think the market wants. It's about understanding where your dopamine comes from and finding founder-market fit based on genuine interest and dedication.

> - Sacrifices are part of the founder journey, and having a strong support system is essential to cope with the challenges. The commitment and dedication to a meaningful idea can drive you through the tough times and remind you of the good fortune of serving millions through your product.

> - In your youth, focus on channeling your passion into meaningful work, even if it requires sacrifices and hard work. Use your time wisely, explore, surround yourself with people who drive you to be better, and embrace obsession and dedication towards your goals, planting seeds early on for future success.

Future of search

> - The future of search and the internet is evolving towards knowledge discovery, moving beyond traditional search and answer engines towards guiding people to discover new knowledge and cater to human curiosity.

> - By empowering individuals to seek truth and engage in knowledge discovery, the hope is to create a better world with increased knowledge and a shift towards fact-checking rather than relying on biased information.

> - Perplexity aims to revolutionize knowledge sharing through tools like Perplexity Pages and personalized experiences, focusing on maximizing human curiosity and learning, ultimately leading towards a future where human curiosity thrives alongside cutting-edge AI capabilities.

Future of AI

> Building deep connections with AI could lead to a potential future where people spend more time interacting with AI than with other humans, especially in work settings, aiming to streamline tasks and productivity.

> The ultimate goal is to create personal AIs that act as empowering coaches, helping individuals achieve their life goals and grow as human beings, fostering meaningful and enduring relationships that aid mutual flourishing.

> By fostering curiosity, understanding, and reducing biases through AI, we have the opportunity to break out of echo chambers, bridge divides, and pave the way for a more peaceful and loving coexistence, ultimately creating a brighter future where AI assists us in understanding the world and each other better.