Linus Lee

Linus Lee researches how people can understand and directly control AI-assisted knowledge tools. He joined Notion’s AI team in January 2023, worked on collaborative features including AI Autofill, and developed Prism, an experimental system for interpreting and editing text embeddings. In September 2024, he joined Thrive Capital as an entrepreneur in residence and advisor, supporting founders’ use of AI while pursuing research on interpretability, knowledge representation, and interfaces.[1][5][4]

Lee chose Notion partly to learn how exploratory ideas become widely used products. In his January 2023 essay, he described enjoying the earliest stages of an idea but wanting to improve at execution and distribution. He also saw an opportunity to influence AI’s emerging interface conventions: Notion could prototype quickly while reaching enough users to learn from practical use.[3][1]

His design agenda extended beyond generating prose. In a June 2023 interview, Lee distinguished content generation, knowledge synthesis, and automation. He argued that much knowledge work consists of understanding information, combining ideas, and deciding what to do before writing the resulting artifact. The interface challenge, in his view, was to make language models useful in those earlier stages.[1]

Lee advocated constraints that make useful actions apparent. His example was the same input box presented as either chat or search: a messaging interface invites greetings and open-ended conversation, while a search bar encourages keywords and questions. He argued that this framing helps align user expectations with model strengths. At Notion, the team paired curated prompts with a custom-prompt escape hatch, preserving expert control and allowing users to discover new applications.[1]

Lee’s work on AI Autofill brought these questions into a collaborative workspace. Autofill could extract information from pages into database properties—for example, identifying pain points or competitor products across user interviews—and update values when the underlying page changed. Lee recalled testing summaries on roughly 6,000 pages, consuming approximately two or three million tokens in about 20 minutes. This was a prototyping test, rather than a production benchmark. He joined after Notion’s original AI prototype, which he credited to Ivan Zhao and Simon Last.[1]

Collaboration led Lee to emphasize “zones of influence” for AI. An autofilled cell or AI block gives generated material a visible boundary, helping users distinguish it from their own writing. He argued that users should also understand what an AI can access and what it has changed. More precise control, he suggested, can reduce the burden of trusting an unconstrained agent; constant approval notifications create their own problem when automation scales.[1]

For prompt engineering, Lee described team practices centered on examples and regression testing. Multilingual few-shot examples helped prevent English instructions from producing English answers to documents in other languages. Handwritten cases and voluntarily shared user feedback tested whether prompt changes broke previously supported behavior. The team also used language models to judge outputs, including whether a summary was reasonable and in the correct language; Lee did not recall whether generation and evaluation used different models.[1]

Lee conducted most of his Prism research at Notion between December 2023 and February 2024, publishing a revised report in June 2024. He applied sparse autoencoders to text embeddings, building on dictionary-learning research, particularly Anthropic’s work. The autoencoder reconstructs an embedding through a larger, sparse dictionary of features, providing candidate explanations for what the embedding represents. Lee reported around 8,000 features for the lg-v6 configuration’s 1,024-dimensional embedding space.[5]

Prism also explored semantic editing: modify an embedding along a feature direction, then decode it into text with Lee’s Contra text autoencoder models. Lee introduced feature gradients to address unwanted changes from overlapping feature directions. The method starts with feature-vector addition, then uses gradient descent at inference time to bring the embedding’s sparse-autoencoder activations closer to the desired values. He reported that this could preserve more wording and structure in precise edits, while simple addition sometimes worked better for broad changes in topic or style.[5]

Building on OpenAI’s automated interpretability work, Lee used GPT-4 to label features and introduced normalized aggregate scoring. It subtracts the fraction of non-activating examples that fit a label from the fraction of highly activating examples that fit it, helping filter out descriptions so general that they explain almost any sentence. He acknowledged ideas, advice, resources, and encouragement from Gytis Daujotas, Matthew Siu, Dan Shipper, Theo Bleier, Maya Bakir, Betaworks, and others.[5]

Lee characterized Prism as a proof of concept. Its editing was not yet competitive with prompting larger models for general-purpose tasks, and he had not rigorously compared edit methods. Automated labeling traded rigor for cost efficiency; relying on top-activating examples could obscure other meanings represented by a feature. His feature gallery was hand-curated rather than a systematic effectiveness study.[5]

Lee left Notion’s AI engineering team in August 2024 and joined Thrive earlier that September after a brief break. In his September 11 announcement, he said he wanted to understand how the broader world was adapting to AI and spend more time on interpretability and interfaces. His role combined helping founders understand and deploy AI with continuing his own research.[4]

These interface questions predated Lee’s Notion work. His earlier career included a researcher-in-residence role at Betaworks and software engineering at Replit, Hack Club, and Spensa. Looking back in June 2023, he described taking all of 2022 off after working mostly as a web engineer at seed-stage startups. Savings and confidence that he could find work enabled the experiment. His starting question was how to improve text interfaces; AI became a means of exploring it rather than the predetermined purpose of the year.[2][1]

During that year, Lee prototyped continuous sliders for text length, philosophical character, and sentiment. He distinguished these controls from Amelia’s preset writing brushes discussed in the same interview. In his 2022 interface essay, he argued that complex tasks benefit from shared context made visible in the interface, rather than buried in conversation. A language-model-powered UNIX shell he built let him issue natural-language commands while working with the agent in a shared environment.[1][6]

Lee’s personal tools also included Monocle, a search engine for which he wrote a custom full-text index. He described sharing such projects so others could understand their workings, rather than primarily seeking mass adoption. His practical advice was to use familiar tools, scope a project to a short period, and follow the rule of “only adding one new unknown per project”—a way to finish something usable before expanding it.[1]

His earlier interest in programming languages, compilers, and interpreters informed another public position: for suitable decision-making tasks, a model could write an inspectable program instead of making each decision opaquely. Lee’s example was generating a regular expression to extract dates. The program becomes a “boundary object” that humans and computers can share, test, debug, and maintain. He advocated combining model-generated code with established program-verification techniques.[1]