Linus Lee is an independent researcher and engineer who builds — and argues for — software that expands what people can think. Writing and building in public since 2014 under the handle "thesephist," he has cataloged well over a hundred side projects, including two programming languages, a personal search engine, and a suite of self-made tools he uses to run his life. His current research asks how machine understanding of language can become new interfaces: ways of working with text and ideas that go beyond typing prompts into a box.
Origins: Korea, Indiana, Berkeley
Lee was born in South Korea and spent the first half of his childhood there before his family moved to Indiana, where he grew up near Purdue University in West Lafayette. He taught himself to code late in high school — JavaScript, Backbone, React — and as a junior took a job at Spensa Technologies, a precision-agriculture startup in the Purdue orbit, staying about two years through a gap period after graduation. In the same years he built the Apogee Citation Maker, a Chrome extension that generated MLA and APA citations and, by his project's account, passed 130,000 peak weekly users before he sold it in 2017. He also plugged into the student-startup scene through The Anvil, Cal Hacks, and the student-run venture fund Dorm Room Fund.
After roughly a year and a half away from school he enrolled at UC Berkeley to study computer science, completed three semesters, and dropped out in fall 2019 — a choice he recounts cheerfully in a rhyming autobiographical post. He briefly re-enrolled in 2020; no degree claim appears in the public record, and his own bio says only that he "went to school in Berkeley."
Between institutions and independence
His work history alternates between tool-building teams and self-directed stretches. A Replit engineering internship in summer 2019 produced lyrics.rip, the Markov-chain fake-lyrics generator he co-built with fellow intern Samarth Jajoo — covered by Genius, later past 800,000 hits, and used in 2020 by YouTuber Funk Turkey for the viral fake-Nirvana track "Smother." That fall he joined the nonprofit Hack Club to work on Hack Club Bank; in early 2021 he joined Ideaflow, a tools-for-thought startup, and moved to New York City, where he still lives.
He spent 2022 on independent research into tools for thought and generative-AI interfaces — including a Researcher in Residence position at Betaworks, where he demoed "Sentence Gradients," early controllable-text-generation work — then joined Notion's AI team as a research engineer. Announced in January 2023, the role ran "almost two years" through August 2024 and included prototyping AI product experiences such as a Q&A feature. In September 2024 he joined Thrive Capital as an entrepreneur-in-residence and advisor on AI interpretability, interfaces, and deployment, keeping his own research program running alongside the investing work.
The self-made stack
The through-line of his project catalog is unusual: he rebuilds the layers of his own computing life. He created Ink in 2019 — a minimal, dynamically typed functional language inspired by JavaScript, Go, and Lua — and later Oak, a pragmatic successor for his tools, with implementations including a self-hosting Ink-to-JavaScript compiler (September) and a Rust runtime (Schrift). On top sit Torus, his dependency-free UI framework, and the applications he actually uses: Monocle, a full-text search engine over tens of thousands of his own documents that he calls an "extended memory" and the closest thing he has experienced to Vannevar Bush's Memex; Polyx, a productivity suite covering contacts, tasks, files, and notes; Lucerne, a Twitter client; Merlot, the Markdown app where he writes; and Revery, a semantic-search layer over the same personal corpus.
He is explicit about the motive. Building your own tools, he told Metamuse, changes your relationship to software — you know where the bits go, and you are in control. Side projects, he told The Changelog, are not side hustles: they are small, personal learning tools, some maintained for years. He has even argued — against industry habit — that a piece of personal software can be finished. A few tools escaped the personal context anyway: Unim.press, a Reddit reader styled like a newspaper front page, has drawn more than 200,000 visitors.
The research program
His essays and talks converge on one program. Notation is the lever: "Notational intelligence" (2022) argues that better notations contribute more to effective intelligence than better automated tools — that notation design is as fundamental as tool building. Language models are the material: because they decouple how information is stored from how it is presented, they open interfaces beyond the chat box. And the interface is the point: he argues for direct manipulation — drag-and-drop and pinch-to-zoom over a model's latent space — so ideas become "things you can hold."
The Prism report (June 2024), produced during his Notion years, applied sparse autoencoders to text-embedding models, recovered tens of thousands of human-interpretable features, and demonstrated precise semantic edits — turning statements into questions, shifting style — by steering embeddings along feature directions and decoding through his Contra autoencoder models. He demoed related work at South Park Commons ("a notebook for seeing with language models") and the AI Engineer Summit ("The Hidden Life of Embeddings"), and discussed the program at length with The Gradient. Alongside the engineering sits a creative practice — two New Age piano albums, artwork, and short fiction at linus.coffee — that treats generative systems as aesthetic material too.
What the record does not settle
The record is thick on work and ideas, thin on vitals. No birthdate appears in the cited sources; his age at each milestone is inferential. Whether he completed a Berkeley degree is unsettled — a 2019 dropout and a 2020 re-enrollment are documented, a graduation is not. Role boundaries are approximate: Ideaflow's end and Notion's start are known only to the quarter. There is no Wikipedia article or Wikidata item for him, so identity rests on subject-controlled pages and the interview record. And the internal shape of the Thrive EIR — how it splits with independent research — is his to disclose, not ours to guess.
This index was compiled from public sources and does not imply the subject's endorsement. Citations live in the packet's source catalog.