Kanjun Qiu

Kanjun Qiu is the CEO and co-founder of Imbue. Her work combines reasoning-oriented AI research with interfaces that let people inspect and direct agents. She has articulated two related challenges: making agents dependable enough for ongoing work, and ensuring that increasingly powerful software remains under its users’ control.[1][3][11]

Before Imbue, Qiu worked on educational electronics, company operations and AI recruiting. She co-authored Sew Electric, a book teaching programming through sewable electronics, and experimented with hardware at the MIT Media Lab. After college, she served as Dropbox’s chief of staff. She later co-founded and led Sourceress, an AI recruiting company.[2]

In her October 2023 Latent Space interview, Qiu described Imbue as her third company with Josh Albrecht. She recalled that Albrecht recruited her from Dropbox to Ember, where they built a VR headset using a laser raster projector. After deciding that VR was not their strongest interest, they considered AI research but chose recruiting as a practical way to learn what deploying AI in the real world required.[3]

Sourceress found potential candidates and helped employers contact them; before language models could handle outreach, human writers customized emails with assistance from automation. Qiu recalled a business difficulty: successful hiring could reduce a customer’s need for the service. She also identified a lesson that informed Imbue’s interface work: employers needed to understand why particular candidates had been selected, rather than simply receive a list. She said Sourceress never fully reached the level of trust and dependability she wanted.[3]

Qiu dates serious exploration of the future research lab to late 2019 and early 2020, when progress in self-supervised learning across images and language encouraged her and Albrecht to investigate more capable computers. They founded Imbue in 2021. The company initially operated as Generally Intelligent; its September 2023 announcement disclosed the rename to Imbue and a $200 million Series B at a valuation above $1 billion.[3][11][4]

An early experiment was Avalon, introduced in 2022: a reinforcement-learning benchmark built with the Godot game engine, using procedurally generated worlds and twenty tasks with a shared sparse reward structure. Qiu explained that varying task difficulty created a curriculum. Easier environments gave agents more opportunities to succeed and receive reward, which she interpreted as providing more learning signal.[5][3]

Qiu recalled that Avalon agents could learn mechanical skills such as climbing and opening doors with multiple switches, yet struggled with higher-level planning and consistent execution. The team’s later account likewise reports better results from tuned PPO on basic tasks than from most reinforcement-learning algorithms on complex, compositional tasks. Qiu said these experiences pushed the team toward reasoning in natural language, with inspiration from Chelsea Finn’s work on SayCan.[3][5]

In October 2023, Qiu’s blunt assessment—“The agents don't work.”—was immediately qualified by successes on small, specific tasks such as changing a button’s color. Her concern was dependable delegation over larger goals. She described reasoning as deciding what information is missing, gathering it, weighing uncertainty, revising a plan and knowing when to ask the user a question.[3]

Qiu’s proposed approach combined reasoning-focused pretraining data, internal use of coding agents and debugging interfaces. Rather than assuming that reasoning would emerge from undirected data collection, she advocated generating examples of explicit reasoning strategies, such as checking the evidence behind a claim. Imbue’s announced strategy also included spending more computation during inference to reach more robust conclusions.[3][4]

She described code as “almost like a curriculum for reasoning”: variables and functions make relationships explicit, while generated code gives an agent a way to act on a computer. Using coding agents internally also exposed failures. When a test-generation agent went off course, she wanted to inspect why it had produced a particular test. Her interface experiments let users inspect execution, fork a run, change prompts or plans, and examine how those changes affected the result.[3]

In “Empowering humans in the age of AI,” published in May 2025 and updated in April 2026, Qiu explained a strategic reconsideration. Imbue had initially expected task-automating agents to distribute power. She came to argue that centrally controlled agents could instead leave their builders governing users’ decisions, with incentives that might diverge from those users’ interests. Her answer was to equip people to create, customize and control their own AI tools.[11]

Qiu connected that ambition to software engineering practices. She argued that generating an initial app was insufficient if subsequent bug fixes and features made it fragile or difficult to maintain. Her proposed coding environment would incorporate testing, safe execution and change management into AI-assisted development. Desired applications included a feed that protected her attention and an app protecting her grandmother from Chinese-language scam calls—examples of personally shaped tools, rather than claims of products already built.[11]

Sculptor, shipped in October 2025, gave the coding-interface work a concrete form. The open-source desktop app runs coding agents in parallel in isolated Git workspaces, with branches, terminals and views of proposed changes. Developers can inspect edits before merging them into their codebase. Its engineering lead, Bryden Fogelman, describes it as originating in Imbue’s own need for multiple agents, isolated branches and full review before merge.[6][7][8]

The Sculptor repository documents support for Claude Code, Pi and terminal-based agents. Although the product page describes Sculptor as stable, the repository calls it an experimental research preview, warning of mistakes, bugs and potentially substantial changes.[6][7]

Qiu’s 2023 account of leadership emphasized helping colleagues develop work around their strengths. “We have projects, not teams.” was her description of Imbue’s structure at the time. She credited the collaborator she called Abe with creating CARBS, the hyperparameter optimizer, while trying to automate his own research process and drawing on ideas from plasma physics. She also favored releasing tools because they would help others, rather than merely demonstrate activity.[3]

That approach faced pressure during product development. In his January 2026 retrospective, “A more radical Imbue,” Albrecht recalled that he and Qiu tightened decisions and processes while sprinting toward Sculptor’s launch. He said they independently concluded over the holiday break that this way of working had lost something central to the company. His resulting proposal was smaller open software projects, with project owners empowered to make or delegate decisions.[8]

Beyond Imbue, Qiu invests in early-stage founders with Ali Rohde and Josh Albrecht at Outset Capital, and hosts conversations with deep-learning researchers alongside Albrecht. She also investigated the processes and progress of science with Michael Nielsen. She co-founded the communal home The Archive and describes building a San Francisco Neighborhood with Jason Benn and former housemates, extending her interest in environments that support learning and agency beyond the workplace.[2]