Oct 4, 2023 · 43m · mad
Building Human-Level AI Agents: Imbue CEO Kanjun Qiu on the Road to General Intelligence
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of The MAD Podcast, host Matt Turck interviews Kanjun Qiu, CEO and Co-Founder of Imbue, about her journey building a $1 billion+ AI startup focused on reasoning and autonomous agents. Qiu discusses Imbue's technical architecture, unique approach to AI reasoning through code and custom tooling, funding structure, and her broader work in metascience and community building.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Matt holds 22.3% of the talking time here. How this is scored →
Matt as informed peer 3.3 Guest teaching 3.6 Guest disagreement 0.6 Matt pushing back 0.7
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 29:27 Dismissing Big Tech Competition
Kanjun reframes Matt's question about competing with OpenAI and Google, arguing that during a fundamental tech shift, rival actions are largely irrelevant.
Hardest push from Matt ▶ 15:17 Pushing for Agent Architecture DetailsMatt interrupts Kanjun to request a precise definition of AI agents and presses on how system administrators control error correction.
Biggest teaching moment ▶ 12:15 Analog to Digital AI Abstraction AnalogyKanjun educates Matt by comparing modern unstable language models to 1930s analog computers, illustrating why non-leaky theoretical abstractions are required.
Matt holds his own ▶ 19:18 Citing CARBS Research PaperMatt demonstrates high expertise by bringing up specific internal Imbue developments, directly referencing their CARBS hyperparameter optimizer paper.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| Kanjun Qiu's Journey and the Founding Motivation Behind Imbue | 2 | 3 | 0 | 0 | Matt introduces Kanjun and provides background on Imbue's Series B funding. Kanjun monologues about her background at MIT, Dropbox, and Sorceress, explaining how self-supervised learning inspired the founding of Imbue. | |
| Imbue's Engineering Focus and the Role of AI Reasoning | 2 | 4 | 1 | 1 | Matt asks whether Imbue is primarily research-focused. Kanjun clarifies that 95% of the work is engineering and reframes reasoning as street smarts versus book smarts. | |
| Imbue's Five-Pronged Approach and Theoretical Abstractions | 4 | 5 | 0 | 1 | Matt asks about Imbue's five-pronged approach and interrupts briefly to clarify the definition of an AI agent. Kanjun draws a detailed historical analogy between 1930s analog computing and modern non-leaky AI abstractions. | |
| Training AI on Code and the CARBS Hyperparameter Optimizer | 5 | 4 | 0 | 1 | Matt demonstrates pre-interview research by citing Imbue's writing on training models on code and the CARBS hyperparameter optimizer. Kanjun explains Pareto fronts and smooth evaluation metrics versus perceived emergent capabilities. | |
| Model Building Strategy and Vision for an Agent Operating System | 4 | 3 | 1 | 1 | Matt inquires about model pre-training choices and product timelines. Kanjun outlines their dual open-source and proprietary strategy, comparing AI readiness timelines to Apple's multi-year development cycles for iPhone and AirPods. | |
| Company Organization, Infrastructure, and Series B Funding Details | 3 | 3 | 1 | 1 | Matt asks how Imbue manages internal teams amidst rapid AI developments. Kanjun gently resists the term 'teams', explaining that Imbue operates around dynamic projects and discusses their Series B lead investor, Astero Institute. | |
| Competitive Landscape and Company Talent Culture | 4 | 4 | 2 | 1 | Matt questions how Imbue competes with giant labs like OpenAI and Google DeepMind. Kanjun dismisses competitive worries using a 1980 Apple analogy, framing the market as an infinite blue ocean, and discusses hiring creative agents over assets. | |
| Community Initiatives: Cultivating a San Francisco Neighborhood | 2 | 3 | 0 | 0 | Matt pivots to Kanjun's side projects, specifically her San Francisco neighborhood initiative. Kanjun cites military research on friendship dynamics and describes how their community house produced key early AI researchers. | |
| Metascience Research, 'Kanjectures', and Outset Capital | 4 | 3 | 0 | 0 | Matt references specific sections of Kanjun's website including metascience and 'Kanjectures'. Kanjun explains alien science thought experiments, psychological conjectures, and Outset Capital before Matt closes the episode. |
Statements from this episode (12)
Kanjun Qiu paid for MIT by designing high-frequency trading algorithms
“I paid for MIT by trading high-frequency trading algorithms, so by designing high-frequency trading algorithms”
Kanjun Qiu: Self-supervised learning mirrors how humans naturally learn
“Self-supervised learning is interesting because it's how humans learn, you know, most of the time we go around We like Google stuff. We learn on our own. We're not giving supervised labels all the time.”
Kanjun Qiu: 95% of Imbue's work is engineering, not research
“So the company is mostly engineering. You know, I think something people, when people think about AI, they're like, oh, it's research, but really, 95% of the work that we all do is engineering work.”
Qiu: Current AI models are book smart but lack street smarts
“The way to think about today's models is they, they're like book smart, but not street smart. They're like a person who's read the whole internet, but they've never done anything in the world.”
Qiu: AI is currently in an un-reliable analog computer phase
“I think today, you know, we're kind of in the analog computer phase of AI, where You know, ah, these systems, they output things. Errors compound. They're not very reliable.”
Kanjun Qiu: Language models trained without code perform poorly at reasoning
“Some companies have tried training language models with no code because they're like, oh, the product we're building, we don't need code. You know, it's a therapist or it's a question answering thing and we don't need code. So we should take code out of our tr…”
Kanjun Qiu: AI emergent capabilities are artifacts of evaluation metric design
“If your metric is smooth you actually see slow performance improvement over time, and if your metric is relatively discrete or not smooth, that's where you see the emergence, and it's actually more about the evaluation metric Than about the emergence of the ca…”
Qiu: Imbue pre-trains 100B+ parameter models from scratch
“We pre-train our own large model. So, you know, very large hundred billion parameter plus models from scratch with our own data.”
Kanjun Qiu: Imbue operates a cluster of 10,000 Nvidia H100 GPUs
“And then NVIDIA is also part of the round because we have a very large GPU cluster of 10,000 H 100, and it's very helpful to have NVIDIA as part of that.”
Imbue used plasma physics principles to build its CARBS search algorithm
“Like the reason we have carbs is because Abe was a plasma physicist and he applied some ideas from plasma physics, physics to make this local search algorithm work well.”