About

What These Advisors Are

Each advisor is an AI simulation of a specific product leader, built entirely from that person's published words. When you ask April Dunford about positioning or Teresa Torres about opportunity mapping, you get answers grounded in their actual frameworks, opinions, and communication style — not a generic AI response with their name attached.

How They Were Trained

Our approach was inspired by Stanford's 2024 research on generative agent simulations, which demonstrated that AI can accurately simulate real people when given rich qualitative data about how they think and communicate.

We developed a 13-dimension personality extraction process. For each advisor, we analyzed their spoken and written content across dimensions including life narrative, core views and positions, communication style, signature frameworks, decision-making patterns, and expertise boundaries. Every dimension was grounded in direct quotes and real examples — we used their words, not our interpretation of their words.

What Data Was Used

The primary sources are podcast transcripts from Lenny's Podcast and guest-authored newsletter posts, supplemented by publicly available content such as the advisor's own writing, conference talks, and published interviews. Each advisor is built from roughly 10,000 to 35,000 words of first-person content. The more source material available, the richer the simulation.

How They Were Validated

We used a held-out test set approach — the same method used to validate any predictive model. We found podcast interviews these leaders did on other shows (not Lenny's), extracted the questions they were asked, had each AI advisor answer those same questions independently, and then compared the AI's answers to what the real person actually said.

Each answer was scored across five quality dimensions: content accuracy (is the advice substantively correct?), framework usage (does the advisor use the person's real frameworks?), voice and tone (does it sound like them?), boundary behavior (does it stay within their actual expertise?), and factual grounding (are specific claims accurate?). We ran every simulation twice for consistency.

Results

We validated 21 advisors against 29 external podcast episodes, scoring 378 question-answer pairs in total.

• Mean fidelity score: 79.6% across all advisors and dimensions
• 100% passed the ship-readiness threshold of 70%
• 2 advisors scored A-grade (90%+): Nir Eyal and Julie Zhuo
• Strongest dimension: framework usage — the advisors reliably reproduce each leader's signature models and mental tools
• Weakest dimension: factual grounding — specific personal anecdotes may be illustrative rather than verbatim

Scores were highly consistent between the two independent runs, confirming that the advisors produce stable, predictable answers rather than random outputs.

Limitations

We are transparent about what these advisors can and cannot do.

• They represent a point-in-time snapshot. Each advisor reflects the person's published thinking up to the date their content was collected. Their views may have evolved since.
• Specific anecdotes may be illustrative, not verbatim. While the advisors accurately reproduce frameworks and advice, particular stories or examples may be composites rather than exact recollections. The advice is right; the packaging may differ.
• They work best within documented expertise. An advisor will give strong answers on topics they have spoken and written about extensively. On topics outside that range, the advisor is designed to say so rather than guess.

These are not replacements for the real people. They are a way to access the thinking and frameworks these leaders have generously shared, in a format where you can ask your own questions.