Episode Summary
Executive Summary: Sridhar Ramaswamy argues that AI and large language models make a fundamentally better search experience possible: one that answers questions directly, summarizes pages with citations, and combines retrieval with generation. He explains Neva’s journey from privacy-focused search to cited AI summaries, discusses costs, distribution, and monetization, and predicts broad disruption to publishing, ads, and tool-based AI workflows.
Main Topics: Why Neva was founded (Priority: 5/5): Ramaswamy describes Neva as a back-to-basics attempt to rethink search after leaving Google, motivated by a belief that search could be better if freed from ad-driven constraints and redesigned around user value. Privacy, ads-free search, and consumer behavior (Priority: 4/5): He explains that Neva initially focused on private, ads-free search, but learned that changing default search behavior is hard and that privacy mattered differently by region, with stronger traction in Europe than the U.S. AI summaries as the breakthrough product (Priority: 5/5): The conversation centers on cited summaries enabled by LLMs, which let Neva move from links to direct answers. Ramaswamy frames this as the first scalable way to deliver the kind of answer-oriented search users already prefer. Retrieval-augmented generation and search’s future (Priority: 5/5): He argues that the future is not chatbot versus search engine, but search engines that use LLMs plus retrieval and tools. He believes many queries can be answered better with RAG, while others remain hard. Cost, model size, and economics (Priority: 4/5): Ramaswamy addresses the expense of LLM-powered search, noting that very large models can be costly per query, but that smaller fine-tuned models and task-specific architectures can make the approach viable. Distribution, partnerships, and publisher impact (Priority: 4/5): He says distribution remains Neva’s biggest challenge and explores partnerships, alternative search experiences, and publisher-facing conversational search. He also warns that AI answers may consolidate content creation and weaken smaller publishers. Broader AI disruption beyond search (Priority: 3/5): Ramaswamy predicts major disruption in advertising, content generation, and agentic tools that combine LLMs with APIs, calculators, and programs—potentially enabling AI SREs, code reviewers, and other action-oriented systems.
Key Arguments: Search is still early in its AI evolution, and users will increasingly expect direct answers rather than opaque lists of links. Neva’s original privacy-first, ads-free positioning was valuable but did not by itself overcome consumer inertia in the U.S. LLMs finally make scalable cited summaries possible, solving a problem that earlier featured snippets and one-boxes could not handle at scale. The winning model is retrieval-augmented generation: a search engine should be a tool used by an LLM, not replaced by a pure chatbot. LLM costs are high for large models, but task-specific smaller models and fine-tuning can make many search use cases economically feasible. The real challenge is not only technical accuracy but query types: some questions are still hard, while many tail queries are now answerable in natural language. Distribution will require a mix of superior product, partnerships, and new publisher integrations because habits are sticky and search defaults are powerful. AI-generated summaries may reduce traffic to smaller publishers and encourage consolidation around larger content platforms with their own conversational interfaces. Advertising will remain economically powerful but is structurally awkward for answer-first products; search quality must remain high to avoid ad overload. The most promising next wave is tool-using AI systems that can invoke search, code, APIs, and other functions to complete actions, not just generate text.
Data Points: Neva launch year: 2019 - Ramaswamy says Neva was co-founded in 2019 after leaving Google. Target query coverage for cited summaries: 50–70% - He says Neva can write a single authoritative answer for roughly 50, 60, 70% of queries. Earlier featured-snippet coverage: 5–7% - He notes Google’s featured snippets never scaled beyond about five to seven percent of coverage. OpenAI API price reduction: 10x - He cites OpenAI reducing API cost by a factor of 10 as a dramatic market change. Estimated large-model call cost: ~$0.05 per query - He says an average very large model call can cost about five cents. Implied serving cost: ~$50 CPM - He translates the five-cent per query cost into roughly $50 per thousand queries. Average U.S. search RPM: $40–$50 - He compares LLM serving costs with U.S. search revenue per thousand queries. Average global RPM: ~$20 - He says global average RPM is substantially lower than U.S. RPM. Model size for summarization: 5–10 billion parameters - He says Neva is comfortable using models in the five to ten billion parameter range for summarization. Flash-based index iteration: 2 days - He says Neva can replace the entire index over a span of two days using its flash-based system.
Pivotal Quotes: "The difference between a chatbot and a search engine that combines a chatbot and retrieval is going to just look more and more bloody going forward." — Sridhar Ramaswamy: On the long-term convergence of search and chat interfaces. "If you can provide a believable answer to a question, people are always going to prefer that over any number of links." — Sridhar Ramaswamy: On why answer-first search experiences are compelling to users. "We are very much at the beginning." — Sridhar Ramaswamy: On how early AI-enabled search still is and how much more can be built.
Implications: Search is shifting from link navigation to answer and action systems. That raises pressure on publishers, favors larger platforms with proprietary data, and opens opportunities for AI tools that combine retrieval, summaries, and execution.