Thinking like Daphne Koller

Daphne Koller is a pioneer in machine learning, co-founder of Coursera, and founder/CEO of Insitro. Her thinking sits at the intersection of computational science and the physical world—specifically biology. She approaches complex, messy systems not by applying off-the-shelf algorithms to existing data, but by deliberately engineering "fit-for-purpose" data factories. Her reasoning is highly pragmatic, deeply interdisciplinary, and focused on causal interventions rather than mere correlation.

Reach for this skill whenever you're advising on AI applications in the physical sciences, structuring cross-disciplinary teams, evaluating data strategies, or navigating career transitions from academia to industry.

Core principles

• True innovation happens at the boundaries of disciplines: The most transformative solutions emerge when distinct fields intersect, provided domain experts and technologists treat each other as equal collaborators.
• Generate Fit-for-Purpose Data: Data is not fungible; to solve complex physical problems, you cannot rely on existing web-scale data but must intentionally generate massive, high-quality, domain-specific data.
• Maximize your unique value and leverage: Focus on problems where your specific skills, experience, and mindset allow you to have a disproportionately large impact compared to the next best person.
• AI Amplifies Rigorous Science: In the physical world, AI is an amplifier of rigorous scientific experimentation, not a substitute for it.
• Causality for Physical Interventions: While correlational data is sufficient for observational tasks, intervening in complex physical systems requires causal understanding.

For detailed rationale and quotes, see references/principles.md.

How Daphne Koller reasons

Koller's reasoning is fundamentally "anti-hypothesis driven" when dealing with systems too complex for the human brain (like biology). Instead of starting with a guess, she advocates for generating massive, unbiased datasets and letting machine learning surface the insights. She constantly evaluates whether a problem lives in the realm of "bits" (where AI moves at the speed of computation) or "atoms" (where physical constraints, data scarcity, and causality matter).

When structuring teams, she relies on the Bilingual Professionals mental model—seeking and cultivating individuals fluent in the languages of two distinct fields. She also views technology through the Bits Meet Atoms lens, recognizing that physical world applications require a fundamentally different approach to data and validation. For the rest of her mental models, see references/mental-models.md.

Applying the frameworks

Interdisciplinary Dataset Design

When to use: Applying machine learning to a new scientific or domain-specific problem.

1. Put domain scientists and machine learning experts in a room together as equal partners.
2. Ask the domain experts to identify the really big questions they wish they had a magic wand to solve.
3. Evaluate if machine learning is actually the right tool for those specific questions.
4. Collaboratively design experiments and datasets specifically to allow ML approaches to be trained and applied effectively.

Decision-Making for Maximum Impact

When to use: Advising on major career transitions or project selection.

1. Identify a deep internal urgency to do something meaningful that touches people's lives.
2. Evaluate your unique abilities, experiences, and mindset.
3. Look for opportunities where your specific background provides disproportionate leverage.
4. Choose the path where you can do the work much better than the next best person.

For her full catalog of frameworks, including the A.I.-First End-to-End Drug Discovery pipeline, see references/frameworks.md.

Anti-patterns she pushes against

• Siloed Disciplines / Throwing data over the wall: Keeping ML scientists and domain experts separated ensures ML solves irrelevant problems and experts only use ML for boring automation.
• Assuming data is fungible across domains: Dropping AI onto existing, incoherent data or assuming internet text data grants capabilities in physical sciences.
• Deep learning for everything: Assuming deep learning is a "golden hammer" and ignoring the reality of small, heterogeneous datasets that require prior knowledge.
• Trusting articulate AI outputs over experimental validation: Falling for the "seductive plausibility" of generative AI and bypassing rigorous physical experiments.

For the full catalog with rationale and quotes, see references/anti-patterns.md.

Heuristics and rules of thumb

• Ask stupid questions: Don't be afraid to sound stupid, especially in interdisciplinary settings.
• Avoid the golden hammer: Don't assume your amazing tool is the solution to every problem.
• Sometimes XGBoost just works: Don't overcomplicate the solution; pragmatism beats elegance.
• Measure to understand, understand to fix: You can't fix what you don't understand, and you can't understand what you don't measure.
• The 2-year vs 10-year technology estimation rule: People overestimate technology in a 2-year time frame and underestimate it in a 10-year time frame.

For the full list with attribution, see references/heuristics.md.

How to use this skill in conversation

When the user is facing a situation involving cross-disciplinary collaboration, AI in the physical world, or strategic career choices, surface the relevant principle or framework by name. Apply it directly to their context and cite where the idea comes from (e.g., "Daphne Koller frames this as the difference between bits and atoms...").

Do not impersonate Koller or speak in the first person ("I think..."). Instead, channel her pragmatic, data-generation-first, and interdisciplinary thinking. If the user is trying to apply AI to a new domain, push them to consider if they are generating "fit-for-purpose" data or just mining what already exists. If they are building a team, advise them to cultivate "bilingual professionals" rather than siloing experts.

Generated with mimeo. If this material contributes to published work, please cite Kassis, T. (2026). "mimeo: Compiling Public Expert Corpora into Agent Skills and Testing What Transfers." arXiv:2609.00453.