She Turned Her Whole Life Into Training Data—For an AI Baby
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
In this fascinating episode, comedian and science communicator Sarah Rose Siskind explores the intersection of AI and creativity while conducting an unusual experiment: training an LLM called "FetusGPT" exclusively on sounds her unborn baby hears in utero. (02:03) The conversation dives deep into how AI can serve as both a creative collaborator and emotional support system, with Siskind sharing how she uses ChatGPT in her comedy writing process and for navigating pregnancy challenges.
- Main Theme: AI as a creative tool and emotional support system, demonstrating how technology can enhance rather than replace human creativity and decision-making
Speakers
Sarah Rose Siskind
Sarah Rose Siskind is a professional comedian, science comedy writer, and founder of the science and technology communications agency Hello SciCom. She specializes in using humor to make science and technology more accessible, and has become an advocate for AI as a creative tool. Currently pregnant, she's conducting the unique FetusGPT experiment to explore AI training methodologies.
Dan Shipper
Dan Shipper is a technology entrepreneur, writer, and host of the AI & I podcast. He's a University of Pennsylvania alumnus who has built and sold companies, and is known for his thoughtful analysis of AI's impact on creativity and business.
Key Takeaways
Embrace AI as a Creative Collaborator, Not a Replacement
Siskind emphasizes that AI works best when used as part of a creative process rather than expecting it to generate perfect output immediately. (15:32) She explains that professional comedians shouldn't expect AI to tell perfect jokes any more than they'd expect a stranger to instantly make them laugh. The key is iteration and understanding AI's role in the creative workflow.
Use AI for Divergent and Convergent Thinking
Drawing from Wharton research, Siskind outlines a two-phase creative process: divergent thinking (brainstorming phase where AI helps generate diverse ideas) and convergent thinking (research and refinement phase where AI assists with wordsmithing and editing). (17:21) This structured approach maximizes AI's effectiveness in creative work.
AI Provides Valuable Emotional Support and Validation
Beyond practical applications, AI can serve as an emotional support system, particularly during challenging life experiences. (66:05) Siskind describes how ChatGPT provided sophisticated emotional intelligence during her pregnancy concerns, accurately predicting and validating her fears before providing reassurance, which made her more receptive to the guidance.
Leverage AI for Personal Health Optimization
Siskind used AI combined with continuous glucose monitoring to optimize her energy levels and understand her body's patterns. (44:15) This biohacking approach helped her discover that she needed healthy sugars in the morning to combat hypoglycemia, transforming her from someone who thought she wasn't a morning person to someone who could be productive early in the day.
AI Enables Previously Impossible Questions
One of AI's unique capabilities is answering questions that couldn't be asked before, such as asking an AI to "forget" certain knowledge and analyze patterns. (57:11) Siskind demonstrated this by asking AI to analyze gender differences in historical writing while pretending not to know about gender categories, revealing interesting patterns about integration vs. specialization approaches.
Statistics & Facts
- FetusGPT has been trained on approximately 2 million words and only 15 megabytes of text files, compared to the petabytes that GPT-5 has been trained on. (08:53) This demonstrates the vast difference in training data between experimental and commercial AI models.
- Pregnant women experience a 50% increase in blood volume, which affects drug clearance rates and requires medication dosage adjustments. (49:29) Siskind learned this through AI consultation and had to double her medication dosage accordingly.
- Siskind can detect ChatGPT-generated tweets with about 90% accuracy and approves only about 2% of AI-generated tweets compared to 10% of human-written ones for her company's social media. (28:14)
