Linus Lee

Linus Lee is head of AI at Thrive Capital, where he builds AI systems and helps founders apply advanced models to their products. His work spans personal search engines, programming languages, model interpretability, and creative interfaces, united by a practical question: How can people inspect and shape what AI systems understand?

Lee grew up in Korea and Indiana and attended school in Berkeley. He worked at Spensa, Hack Club, and Replit before becoming a researcher in residence at Betaworks. His personal software projects include the programming languages Ink and Oak, the web-interface framework Torus, and Monocle, a search engine for notes, journals, bookmarks, and other personal archives. Revery extended his retrieval experiments into embedding-based semantic search.

After a year of independent research into reading, writing, and knowledge interfaces, Lee joined Notion in January 2023 to explore how experimental language-model interfaces could become useful mainstream products. He worked there as a research engineer, spanning retrieval, evaluation, agents, and collaborative-software interfaces.

Lee left Notion in August 2024 and joined Thrive Capital the following month as an entrepreneur in residence and adviser. He now leads AI there, applying his interests in retrieval and model interfaces to production systems and advising portfolio-company founders.

• Latent-space interfaces: Lee builds tools that make a model’s internal representations directly manipulable. His embedding experiments move text along semantic directions associated with length or sentiment, arrange alternatives on a spatial canvas, and combine embeddings into recognizable conceptual hybrids.

• Embedding inversion: Using a T5-based denoising autoencoder and a learned linear adapter, Lee reconstructed approximate passages from OpenAI embedding vectors, recovering details that sometimes included proper nouns, punctuation, and topical structure. He extended the approach to images using CLIP and Kakao Brain’s Karlo model, interpolating between visual styles and altering imagery through text-derived vector arithmetic.

• Prism and interpretable representations: Prism uses sparse autoencoders to identify interpretable concepts and stylistic features in model representations. Lee’s experimental notebook for working with ideas turns related techniques into interfaces for discovering relationships and directly manipulating meaning. His writing on creative tools imagines semantic and stylistic controls that give writers finer agency than prompting alone.

• Context engineering as search: At Thrive, Lee has worked on Puck, an internal assistant designed to navigate complex organizational information. His approach to context engineering treats indexing, databases, retrieval, tools, and subagents as complementary ways to locate relevant information amid incomplete and conflicting data.

Lee opposes tools that encourage disengaged creation. He favors systems that keep people closely involved in their work and make machine intelligence a more responsive medium for thinking and making.