Key Takeaways

  • Elad Gil employs AI models to analyze founder images, claiming surprising accuracy in predicting personality traits relevant to founder quality.
  • This AI method simulates the micro-feature assessment humans naturally perform when meeting new people, but with enhanced scale and consistency.
  • Gil also leverages AI for complex topic deep dives, aggregating clinical trial data and primary sources to efficiently challenge conventional wisdom.
  • His process involves structured prompts to extract data, generate summaries, and cross-reference information across multiple AI models.

The Method

Elad Gil employs a two-pronged AI research method.

First, for founder evaluation, Gil uploads founder images to AI models. He then prompts the AI to predict personality traits and founder quality based on micro-features. Gil states this "works pretty well," noting, “We quickly try to create an assessment of that person and their personality and what they're like. And there's all these micro features like do you have crows feet by your eyes which suggest that your smiles are genuine and what does that imply about the sense of humor you have or fured your brow over time and what does that you know so there's all these like micro features.” This AI-assisted visual analysis speeds up an assessment normally done intuitively.

Second, for complex topic analysis, Gil conducts “deep dives with models into like questions that I just find interesting.” His approach involves:

1. Prompting for Aggregation: Asking multiple AI models to gather diverse data, such as clinical trial information or different types of research.

2. Source Extraction: Requiring the AI to “give me the primary sources” for all aggregated data, allowing for direct verification.

3. Summary Generation: Requesting concise summaries and charts of the findings.

4. Data Validation: Applying “a whole series of prompts around that to kind of also clean data and check it,” ensuring accuracy and enabling Gil to “go back and double check the data and then reread through the literature.” This structured prompt engineering helps identify non-obvious correlations, like maternal age having a greater impact on autism than paternal age in some studies.

Where This Breaks Down

This method relies heavily on the AI's interpretive ability and access to training data, which introduces several failure points. First, personality predictions from images are a black box. The AI's internal model for "good founder" or "genuine smile" is opaque, making it impossible to audit for bias or false correlations. Gil's "practice smiling people" comment suggests he finds it surprising, not necessarily robust. It risks perpetuating existing biases present in the training data, for instance, favoring certain demographics or facial features that have no causal link to actual founder success.

Second, the quality of AI deep dives depends on the accuracy and completeness of the public data the models were trained on. While Gil emphasizes asking for primary sources, the AI might hallucinate sources or misinterpret complex scientific literature, particularly in evolving or highly specialized fields. It's a powerful discovery tool, but not a replacement for expert human synthesis and critical review. Its strength is uncovering overlooked connections, not providing definitive truths without independent validation.

What to Do With This

This week, pick one complex problem you're researching for your business – perhaps a new market segment, a competitor's strategy, or a technical challenge. Design a series of 5-7 targeted prompts asking an advanced AI model (e.g., GPT-4, Claude 3) to: 1) aggregate diverse data points, 2) extract specific primary sources or evidence, and 3) summarize findings into a bulleted list or table. Compare the AI's output and cited sources against your existing understanding or traditional research to identify overlooked insights or contradictory evidence, focusing on challenging your assumptions rather than just confirming them.