Visual summary of operating lessons from Kanjun Qiu.

Lessons from Kanjun Qiu

Kanjun Qiu co-founded Imbue, where she has led work on reasoning-capable AI agents intended to expand human agency. She previously founded the AI recruiting company Sourceress and served as Chief of Staff at Dropbox. This profile examines her views on AI, personal software, communities, research, and learning. — Kanjun Qiu — About.

Part 1: The Philosophy of Agency

  1. Use AI to expand what people can do: Qiu wants computers to take on tedious mental work so people can pursue larger goals and spend more time on creative, collaborative activities. — MAD Podcast Interview.
  2. Keep people in the driver's seat: For Qiu, a good AI future depends on people retaining decision-making authority while using AI to understand options and act more effectively. — Pioneers of AI Interview.
  3. Agency can be cultivated in practice: Qiu describes helping friends and teammates identify what blocks them, while building systems to make her own work more effective. — Latent Space Interview.
  4. Make consequential software controllable: Qiu argues that people affected by algorithmic decisions should be able to understand how those decisions are made, seek recourse, and change software that shapes their digital environment. — Decoder Interview.
  5. Rekindle personal computing: Qiu wants AI agents to revive the personal-computer ideal by giving individuals tools they can inspect, direct, and adapt to their own purposes. — Possible Interview.
  6. Guard against concentrated AI power: Qiu warns that AI could concentrate economic and political power unless people gain tools that preserve agency over their own decisions and lives. — Possible Interview.
  7. Expand the sense of what is possible: Qiu connects agency with authorship over one's life and with seeing examples that challenge limiting social narratives. — Possible Interview.
  8. Go beyond chat for agent interfaces: Qiu treats chat as an early interface for agents and argues that users need ways to inspect, modify, and fork an agent's reasoning or state. — Latent Space Interview.
  9. Examine the narratives you adopt: Qiu argues that narratives act like predictions: they shape what people notice and can blind individuals and cultures to possibilities. — We Should Be Very Careful What Narratives We Adopt.

Part 2: Building Reasoning-First AI

  1. Reasoning was a reliability bottleneck: In her 2023 account of Imbue's research, Qiu said pure reinforcement-learning agents could perform some mechanical tasks but struggled with reliable higher-level planning; that pushed the team toward language-based reasoning. — Latent Space Interview.
  2. Inspect an agent's reasoning: Qiu wants users to examine an agent's reasoning trace, change assumptions, and see how the conclusion changes before trusting consequential work. — Latent Space Interview.
  3. Train models for reasoning tasks: Imbue trained large foundation models optimized for reasoning because Qiu wanted agents that could help people delegate larger goals, not merely return one-off text completions. — MAD Podcast Interview.
  4. Verify AI output before relying on it: Qiu's NVIDIA conversation discusses the challenge of verifying AI outputs and Imbue's work to improve agents' verification capabilities; the episode does not establish how much research effort the company devoted to that work. — NVIDIA AI Podcast.
  5. Memory matters for longer tasks: Qiu identifies durable memory as one obstacle to agents handling difficult, long-running work, while noting that simpler agents can be useful sooner. — Latent Space Interview.
  6. Keep agent infrastructure inspectable: Qiu favors agent tools that people can inspect and adapt locally, so users retain control rather than depending entirely on centralized systems. — Possible Interview.
  7. Ask questions when the stakes require it: Using flight booking as an example, Qiu says a useful agent must recognize missing information and risky choices, asking for clarification without constantly interrupting the user. — MAD Podcast Interview.
  8. Pursue a theory of deep learning: Qiu says Imbue supports first-principles work on how neural networks learn because deeper understanding could help keep advanced AI open and less concentrated; she presents a scientific theory as an aim, not an established result. — Generally Intelligent: Learning Mechanics.
  9. Use critique and debugging to improve agents: Qiu describes model critique, self-critique, and debugging tools as ways to improve agent reliability alongside better reasoning and abstractions. — Latent Space Interview.
  10. Code provides explicit reasoning data: Qiu sees code as unusually explicit training material for reasoning: variables and functions expose structure, while coding agents reveal where models struggle and let them act on computers. — Latent Space Interview.

Part 3: Scaling Laws and Economic Shifts

  1. Watch the balance between labor and capital: Qiu worries that if AI makes wage work less valuable relative to capital, workers could lose economic and political power; she presents this as a risk to address, not a settled forecast. — Generally Intelligent: Liberty and AI.
  2. Choose an open AI ecosystem: Qiu sees AI at a crossroads between user-controlled software reminiscent of the early internet and centralized platform ecosystems that can lock people in. — Decoder Interview.
  3. Abundance is not the same as liberty: Qiu argues that meeting material needs alone would not ensure freedom; she wants people to have real agency and, potentially, ownership of productive tools. — Generally Intelligent: Liberty and AI.
  4. Build for worker-led automation: Qiu distinguishes tools that let workers automate parts of their own jobs from systems built chiefly to replace them. She says product builders can influence that choice, while acknowledging it will not solve every labor problem. — Generally Intelligent: Liberty and AI.
  5. Experiment at small scale before scaling: Qiu describes using smaller models to test training choices and transfer useful hyperparameter findings to larger runs; that is an efficiency strategy, not a claim that small models generally outperform large ones. — Latent Space Interview.
  6. Prepare for misuse as capabilities grow: Qiu includes the empowerment of bad actors among AI's major challenges and discusses both preventing harmful uses and making society more resilient to them. — Generally Intelligent: Liberty and AI.
  7. Software fundamentals still matter: Qiu expects algorithms, data structures, software engineering, and systems thinking to remain useful even as better tools make programming easier. — Pioneers of AI Interview.
  8. Lower software costs could enable niche tools: Qiu imagines that if writing software became much cheaper, people could make more specialized tools for their own needs instead of relying only on mass-market products. — NVIDIA GTC Fireside Chat.

Part 4: Human-Computer Interaction and Malleable Software

  1. Let people adapt the software they use: Qiu wants infrastructure that lets people modify personal software when they need to; she treats live end-user editing as a promising direction with unresolved versioning challenges. — Generally Intelligent: Malleable Software.
  2. Give agents shared visual context: Qiu argues that agents should be able to see and understand the interface a person is using, so both can refer to the same visible work rather than relying on chat alone. — Generally Intelligent: Malleable Software.
  3. Make modification possible without making it mandatory: Qiu wants to lower the barrier from simply using software to modifying it, while recognizing that most people will not want to edit most tools most of the time. — Generally Intelligent: Malleable Software.
  4. Make computing's power more accessible: Qiu credits her MIT Media Lab experience with showing her the power of computing and inspiring her later effort to make that power more accessible. — Pioneers of AI Interview.
  5. Use textiles to teach computing: At MIT's High-Low Tech group, Qiu explored conductive thread, fabric, and soft circuitry as ways to teach computer science and engineering concepts. — MIT High-Low Tech Biography.
  6. Help agents understand user intent: Qiu uses flight booking to show why an agent needs to recognize missing information and risky choices, then ask for clarification at the right moment. — MAD Podcast Interview.
  7. Use agents seriously to expose failures: Imbue tests agents in its own daily work because repeated real use reveals robustness and interface problems that a polished demo can hide. — Imbue: Introducing Imbue.
  8. Automate narrow engineering chores: Imbue built agents for tasks such as fixing type-checking and linting errors and used them alongside debugging tools to make research iteration faster. — Imbue: Introducing Imbue.

Part 5: Environment Design and Collective Intelligence

  1. Cultivate a productive scene: Qiu describes the Neighborhood as a place where friends learn from one another and ideas circulate; she uses the term scenius for that collective setting. — MAD Podcast Interview.
  2. Use shared living to explore human potential: Qiu co-ran The Archive, a 25-person house, for five years and recalls its culture of asking how residents might become better versions of themselves. — Ben Yeoh Interview.
  3. Build an adult-campus neighborhood: Qiu describes the Neighborhood as a more sustainable version of The Archive: friends living close enough for spontaneous encounters and a lively intellectual culture. She credits Jason Benn with driving the project. — Ben Yeoh Interview.
  4. Choose peers who expand your possibilities: Qiu says close peers shape what people believe they can do, so she seeks people she admires who broaden rather than narrow her sense of capability. — Ben Yeoh Interview.
  5. Build culture through beliefs and systems: Qiu describes an institution's culture as the beliefs its members share and the systems that reinforce them; this helps explain why an established organization may struggle to adopt an idea at odds with its core assumptions. — Ben Yeoh Interview.
  6. Treat institutions as design problems: Qiu asks whether the social processes of organizations and science could be deliberately redesigned and tested rather than treated as fixed. — MAD Podcast Interview.
  7. Let curiosity and play open research questions: Qiu describes playful exploration of an unanswered mathematics problem as part of what led her into research; curiosity and enjoyment, rather than an external test, kept her with the question. — Research as Understanding.

Part 6: Startups, Founders, and Capital

  1. Back kind and capable founders: Qiu describes Outset Capital's mission as supporting kind and capable early-stage founders in doing their life's work. — Kanjun Qiu — Current Work.
  2. Teach practical effectiveness as a company scales: Qiu recalls giving a How to Be Effective onboarding talk at Dropbox while serving as Chief of Staff; she presents building systems for effectiveness as part of her work with teams. — Latent Space Interview.
  3. Trust matters in AI recruiting: Qiu says Sourceress's candidate-finding system taught her how much hiring managers needed to trust that candidates had been vetted and selected for a reason; that experience informed Imbue's agent interfaces. — Latent Space Interview.
  4. Use a chief of staff as a flexible CEO proxy: Drawing on her experience as Dropbox's Chief of Staff and later as a CEO, Qiu says the role can act as a second set of hands for a leader across varied responsibilities, depending on the company's needs. — Outset: Hiring a Chief of Staff.
  5. Match funding to the research horizon: For Imbue, Qiu wanted a funder comfortable with a longer timeline and less immediate commercialization pressure while the team pursued an agent operating system. — MAD Podcast Interview.
  6. Treat AI research as engineering work: Qiu says Imbue's research depends heavily on engineering: internal tools and infrastructure make experiments reproducible and help the team build models. — MAD Podcast Interview.
  7. Look for overlooked latent potential: Qiu argues that funders should seek capable people or ideas whose potential has not yet been unlocked, rather than simply allocate more money where other funders already have. — Ben Yeoh Interview.
  8. Anticipate institutional resistance to change: Qiu calls attention to institutional antibodies: existing organizations may resist new approaches that threaten their position. She treats this as partly a competitive-system dynamic, not just a matter of internal culture. — Ben Yeoh Interview.

Part 7: Epistemology and the Nature of Discovery

  1. Treat research as deepening understanding: Qiu argues that research grows from trying to understand a question deeply enough to reach what nobody yet knows, rather than setting novelty itself as the goal. — Research as Understanding.
  2. Experiment with the social processes of science: Qiu and Michael Nielsen propose metascience as a way to design and test new funding, selection, and institutional processes that might unlock overlooked discovery potential. — A Vision of Metascience.
  3. Question beliefs learned in earlier environments: Qiu uses an overfitting metaphor for her own experience of fears learned earlier in life that no longer fit her present circumstances. She offers it as a personal framework, not a clinical finding. — Ben Yeoh Interview.
  4. Investigate what does not fit: In Qiu's account of research, observations that do not fit an explanation are useful prompts to ask why and explore beyond what is already known. — Research as Understanding.
  5. Build less-leaky abstractions for agents: Qiu compares current agent development with early computing and argues that better abstractions and interfaces could make agents more dependable and easier to build upon. — Latent Space Interview.
  6. Hire people who think from first principles: Qiu says Imbue looks for people who think from first principles, question the current state of things, and bring enough agency to shape projects around their distinctive strengths. — MAD Podcast Interview.
  7. Treat the structure of knowledge as an open question: Qiu questions whether knowledge only accumulates as dependencies on earlier discoveries; she suggests later thinkers may sometimes reconsolidate and simplify an idea enough to understand it more clearly. — Ben Yeoh Interview.
  8. Invite evidence against your conjectures: Qiu presents her Kanjectures as provisional explanations and explicitly invites readers to bring evidence for or against them. — Kanjun Qiu: Kanjectures.

Part 8: Cultivating Ambition and Learning

  1. Ask what generates human motivations: Qiu lists the drivers of human behavior and the development of motivations among the questions she is actively curious about; she does not present a settled formula for evaluating people. — Kanjun Qiu — Questions.
  2. Create entryways into computing: Qiu used computational-textile workshops to introduce middle- and high-school girls to computer science. She argued that new entryways can help challenge stereotypes about who belongs in the field. — MIT News: TIMtalks.
  3. Make room for creativity beyond fear: Qiu conjectures that people become more creative and contribution-oriented when basic needs are met and an internal sense of fear or scarcity also recedes; she distinguishes the two conditions rather than treating them as identical. — Humans Are Fundamentally Creative.
  4. Teach programming with sewable electronics: Qiu co-authored Sew Electric, a book that teaches children programming through projects with sewable electronics. — Kanjun Qiu — Past Work.