A very human vision for going all-in on AI

Episode Summary

Executive Summary: In this episode of The Vergecast, host David Pierce interviews Sari Azout, founder of Sublime, a personal knowledge management tool. They discuss how Sublime uses AI to help users collect, connect, and create from their personal libraries of saved content. Sari emphasizes that AI should be invisible and serve human needs, not dominate the experience. She shares insights on building taste-driven products, the value of intention over attention, and the importance of using AI to augment—not replace—human agency. The conversation explores the balance between AI-driven discovery and authentic human curation, with practical examples like Podcast Magic.

Main Topics: Sublime's Design Philosophy: Sublime is a personal knowledge management tool that emphasizes curating a library of meaningful content connected to others' libraries. It focuses on human judgment for collection while using AI invisibly for tasks like podcast transcription and semantic search. Role of AI in Product Development: Sari argues that AI should be a tool, not a lifestyle decision. She advocates for invisible AI that solves specific problems rather than being marketed as a feature, warning against over-automation and loss of user control. Intention over Attention: Sublime prioritizes user intention (e.g., saving a quote) over algorithmic attention-grabbing, enabling serendipitous discovery through semantic connections rather than for-you feeds. AI-Grounded Creation: Using personal libraries to ground AI assists in creative tasks like writing, but Sari cautions that AI often retreats to averages and lacks the nuance needed for high-quality output. User control is critical. Podcast Magic as a Case Study: A feature allowing users to capture podcast insights via screenshot—using AI for transcription, timestamp detection, and moment extraction—exemplifies invisible, utilitarian AI that solves a real problem. Human Taste vs. Algorithmic Mediation: The conversation explores how AI recommendations shape taste, with Sari emphasizing the need for 'inefficient' consumption to develop genuine discernment, rather than outsourcing taste to algorithms. AI's Impact on Agency and Creativity: Sari discusses using AI as a thinking partner (e.g., ChatGPT voice for rambling) and for prototyping, but warns against outsourcing agency. The future lies in 99% inspiration, 1% perspiration.

Key Arguments: AI should be invisible—users shouldn't know or care about the underlying technology if it solves a problem effectively. The 'collect, connect, create' loop is best served by human judgment for collection and AI for connection, not replacement of human input. Hallucinations can be a feature for discovery, not a bug, by surfacing unexpected connections. Semantic search is a 10x improvement over traditional search, enabling 'vibe search' that doesn't require rigid tags or folders. Personal context—a curated library of your own content—becomes more valuable with AI because it grounds outputs in your sources. VC hype around AI agents is often disconnected from real user needs; smart automation is more realistic and less disappointing. The quality of AI-assisted work should be measured by whether it makes the output better, not just faster or more efficient. Technology is neutral; the outcome depends on whether the tool supports outsourcing agency or augmenting it.

Data Points: Threshold for content recommendation: 3 people - Sublime treats content added by at least three users as having signal, without amplifying viral outliers. Time to process podcast screenshot: 30 seconds - Podcast Magic delivers transcript and moment capture within 30 seconds of a screenshot. Number of items in Sublime database: millions - The database contains millions of items from user libraries, enabling semantic search and connections. Time spent building Podcast Magic: close to a year - Due to edge cases and details, it took nearly a year to get the feature right. Podcast detection steps: 3 - AI needs to determine podcast title, podcast name, and timestamp from a screenshot. Number of public RSS feeds indexed: millions - Podcast Magic indexes millions of podcasts for rapid audio matching and transcription.

Pivotal Quotes: "If you want to just do a useful valuable thing with ChatGPT or any of these other tools, get movie recommendations. Tell it a bunch of movies you like and let it feed you other movies that you might like. It's low stakes... but it is able to just sort of like open up a bunch of doors for you." — David Pierce: Discussing how even simple AI recommendation tasks (like movies) can be genuinely helpful and low-risk. "I think people think of hallucinations as a bug. And I think in this context, it's a feature. It's the fact that it's going to take you down a path that is completely unexpected that is interesting." — Sari Azout: Explaining how Sublime's semantic recommendations benefit from hallucination-like serendipity. "I think the ideal future is one where actually, we are in symbiosis with this machine, and we hand over all of the optimization so we can focus on the dreaming." — Sari Azout: Articulating her vision for human-AI symbiosis: machines handle perspiration, humans focus on inspiration.

Implications: This conversation signals a shift toward human-centered AI design where tools augment rather than replace agency. For developers and product builders, it highlights the value of invisible, single-purpose AI features (like Podcast Magic) over all-in-one agents. For users, it reinforces the power of building personal knowledge bases that ground AI outputs in meaningful context. The long-term implication is that the most successful AI products will be those that nurture human taste, creativity, and intentionality—not just efficiency.