Key Moments
Beating Google at Search with Neural PageRank and $5M of H200s — with Will Bryk of Exa.ai
Latent Space PodcastKey Moments
Exa.ai is revolutionizing search with neural PageRank, aiming for perfect information retrieval and challenging Google's established methods with advanced AI and significant compute resources.
Key Insights
Exa.ai is building a new search engine based on 'neural PageRank' and advanced AI, aiming to understand queries with greater nuance than traditional search engines like Google.
The company evolved from Metaphor Systems, initially focusing on link prediction as a core training objective, analogous to predicting the next token in language models.
Exa.ai's approach involves significant compute power, with a recent $5 million investment in H200s, enabling more complex and comprehensive search queries.
Variable compute products and the concept of 'search latency tolerance' are central to Exa.ai's vision, suggesting search can take minutes or even days for comprehensive results, rather than milliseconds.
Exa.ai positions itself as the 'OpenAI of search,' acting as a research-driven infrastructure company that also offers accessible products, differentiating from services that merely layer LLMs on top of existing search APIs.
The platform aims to combat 'SEO slop' prevalent in current search results by focusing on genuine relevance and quality, influencing how content is created and discovered on the internet.
THE ORIGINS AND VISION OF EXA.AI
Exa.ai, formerly Metaphor Systems, was founded with an audacious goal: to build a search engine demonstrably better than Google. The company's journey began in the summer of 2021, a time marked by the emergence of powerful language models like GPT-3, which showcased an unprecedented ability to understand nuanced human language. This contrasted sharply with the perceived stagnation of traditional search engines. Exa.ai's core vision has always been to create a search engine that truly understands user intent, leveraging the insights and capabilities of LLMs to achieve 'perfect search'.
NEURAL PAGERANK AND LINK PREDICTION TECHNOLOGY
At its core, Exa.ai's innovation lies in its 'neural PageRank' technology, which evolved from an initial focus on link prediction. This process involves training models to predict links that follow a given text, akin to how language models predict the next token. By hiding URLs and asking the model to predict them based on surrounding context, Exa.ai builds a search engine that can infer relevant links from a query. This method is seen as more powerful than traditional PageRank, as it learns diverse ways users might refer to content while also assessing its importance.
INVESTING IN COMPUTE AND VARIABLE COMPUTING
Achieving comprehensive and nuanced search requires substantial computational resources. Exa.ai has made significant investments, including a recent $5 million allocation for H200 GPUs, to power its advanced models. The company is pioneering the concept of variable compute products, acknowledging that different queries require vastly different amounts of processing power. This allows users to allocate compute budgets, enabling searches that might take minutes or even hours but yield far more comprehensive and accurate results than the traditional sub-second responses.
THE EXA.AI PRODUCT ECOSYSTEM
Exa.ai offers a suite of tools beyond its core search engine. This includes capabilities for web scraping, providing users with clean, parsed content from the web, which is crucial for building AI applications. The company also offers a 'list builder' function, designed to generate comprehensive lists based on complex queries—a feature directly addressing the perceived shortcomings of current search engines in providing exhaustive and relevant results. These offerings position Exa.ai as both an infrastructure provider and an accessible product.
CHALLENGING THE SEARCH PARADIGM
Exa.ai differentiates itself from rivals like Perplexity and those using traditional search APIs by building its own search infrastructure from the ground up. The company actively combats 'SEO slop' and low-quality content that often plagues search results. By focusing on neural networks and deep understanding, Exa.ai aims to incentivize the creation of high-quality, discoverable content, thereby influencing the future evolution of the internet and how information is accessed and valued.
THE FUTURE OF SEARCH AND ARTIFICIAL INTELLIGENCE
Looking ahead, Exa.ai envisions a future where search is seamlessly integrated with advanced AI agents. As LLMs become more sophisticated interfaces, they will be able to pose complex, paragraph-based queries to search engines like Exa.ai, which are designed to handle them. The company's philosophy extends to internal culture, fostering a fun, collaborative, and counter-consensus environment, exemplified by practical innovations like nap pods. This blend of cutting-edge technology and unique company culture aims to redefine the search landscape.
Mentioned in This Episode
●Products
●Software & Apps
●Companies
●Concepts
●People Referenced
Common Questions
Exa.ai aims to be a 'perfect search engine' by using neural networks and advanced AI to understand queries deeply, unlike traditional keyword-based search engines like Google. They focus on high-quality, relevant results rather than sheer volume.
Topics
Mentioned in this video
A language model mentioned as a catalyst for Exa's development, highlighting the potential of LLMs in search.
A neural network architecture that inspired Exa's link prediction model, involving predicting the next link in a similar way to predicting the next token.
A competitor in the enterprise search space, mentioned in the context of B2B applications of search technology.
A language model that Exa uses for certain tasks like labeling queries, showcasing its ability to understand complex meaning beyond keywords.
A platform where a user experienced Exa hallucinating YouTube links instead of support guides.
Amazon Web Services, used by Exa.ai for inference alongside their own compute cluster.
A conversational AI model from OpenAI, mentioned as a parallel to Exa's early release strategy and its impact on user understanding of search.
A major search engine that Exa.ai aims to surpass. The video contrasts Exa's approach with Google's traditional keyword-based methods.
A cloud storage company whose executive appeared on the podcast to discuss enterprise search solutions.
A company focused on building a perfect search engine, formerly known as Metaphor Systems. They are developing advanced search algorithms and infrastructure.
A space exploration company where Will Bryk interned, highlighting an anecdote about a rule against interns touching Elon Musk.
A prominent AI research lab, frequently compared to Exa.ai for its research-driven approach and product releases like ChatGPT.
More from Latent Space
View all 185 summaries
86 minNVIDIA's AI Engineers: Brev, Dynamo and Agent Inference at Planetary Scale and "Speed of Light"
72 minCursor's Third Era: Cloud Agents — ft. Sam Whitmore, Jonas Nelle, Cursor
77 minWhy Every Agent Needs a Box — Aaron Levie, Box
42 min⚡️ Polsia: Solo Founder Tiny Team from 0 to 1m ARR in 1 month & the future of Self-Running Companies
Found this useful? Build your knowledge library
Get AI-powered summaries of any YouTube video, podcast, or article in seconds. Save them to your personal pods and access them anytime.
Try Summify free