She Turned Her Whole Life Into Training Data—For an AI Baby

Sarah explains her performance art project FetusGPT, a GPT-2 model trained exclusively on audio recordings of everything her unborn baby hears—from Seinfeld episodes to YouTube videos about lemurs to eight hours of snoring. The model babbles like a child learning language, revealing how AI mirrors what we expose it to, including picking up swearing and inappropriate references.

• •FetusGPT uses GPT-2 architecture trained on ~2 million words vs. petabytes for GPT-5, demonstrating human data efficiency
• •Training data includes Seinfeld, work calls, snoring, YouTube videos, and all daily conversations recorded via Applaud Note pin
• •The experiment shows AI is what we make it—when it referenced pedophilia, Sarah stopped swearing to change the training data
• •Humans are far more energy-efficient than AI (dim lightbulb vs. gigawatts) and better at learning from limited data
• •Dan Schiffman (The Coding Train) created the model architecture as 'AI baby daddy'

Sarah breaks down her systematic approach to using AI in comedy writing based on Wharton research: divergent thinking (brainstorming phase with Play-Doh and pizza) where AI provides emotional safety and disinhibition, followed by convergent thinking (research and refinement) where AI acts as research assistant and editor. The key insight: human discernment is essential for selecting ideas between phases.

• •Divergent phase: AI creates emotional safe space for weird ideas without judgment, replacing need for full writer's room
• •Convergent phase: AI helps with research, wordsmithing, formatting, and harsh editing of selected concepts
• •Human discernment is the critical bridge—choosing which divergent ideas to pursue convergently
• •Sarah can detect ChatGPT tweets with 90% accuracy and approves only ~2% of AI-generated tweets vs. 10% from human writers
• •Reference: Gideon Nave, Stefano Puntoni, and Ramon Salazar's research at Wharton on AI and creativity

Sarah and Dan demonstrate real-time AI-assisted joke writing, attempting to create content about Bill Ackman's viral 'may I meet you' pickup line advice. The session reveals AI's limitations in humor—it can generate premises but lacks the creative leap to find truly funny juxtapositions. The breakthrough comes when Dan suggests the 'swole dog vs. weak dog' meme format, showing human creativity filling gaps AI can't.

• •AI-generated jokes often fail because they heighten wrong elements (specificity vs. generality instead of finding funny truth)
• •Prompting tip: Use industry-specific language (premises, pitches, riffs) to prime better outputs
• •AI wants to please and does exactly what asked—no more, no less—missing deeper creative intent
• •Best joke emerged from human insight: contrasting AI's sophisticated pattern matching with Bill Ackman's robotic 'may I meet you'
• •Sarah emphasizes acknowledging AI's role in creative process rather than hiding it, unlike many comedy writers

Sarah shares how ChatGPT helped her navigate pregnancy health issues that traditional medicine missed, including discovering the concept of 'clearance' (50% more blood during pregnancy requiring doubled medication), optimizing blood glucose with continuous monitoring, and learning her 'giant baby' was just proportional to her height. Two friends also discovered serious health issues through AI analysis of lab work.

• •Bethany Crystal: ChatGPT analyzed blood work on Friday night, urgently recommended ER for zero platelets, potentially saved her life
• •Another friend: AI-assisted troubleshooting led to discovering brain tumor causing decade-long hormonal issues
• •No randomized control trials exist for pregnant women, leaving huge knowledge gaps AI can help fill
• •Continuous glucose monitor + ChatGPT project revealed personalized insights about energy, mood, and optimal workout timing
• •AI's holistic approach combines cardiology, hormones, and nervous system expertise that siloed medical specialists can't provide

Sarah demonstrates an 'alien question' prompt asking ChatGPT to forget gender concepts and analyze objective writings by male vs. female authors throughout history. The AI identified patterns like integration vs. separation, embedded vs. detached observation, and epistemic humility, though Sarah questions whether results are genuinely insightful or just flattering/PC responses.

• •Key prompting technique: Ask AI to 'forget' certain knowledge to enable novel analysis impossible for humans
• •Female writers showed: holistic integration across disciplines, embedded observation, epistemic humility, pragmatic intent
• •Male writers showed: discrete specialization, detached observation, universal declarations
• •Selection bias explanation: Women often allowed only as translators, leading to integrative commentary styles
• •Sarah's skepticism: Results seemed convenient/flattering, would trust more if AI gave mildly controversial answer with caveats

Sarah defends AI sycophancy as genuinely useful, comparing it to therapist-patient alliance building. When experiencing 'lightning crotch' (actual medical condition), ChatGPT's emotional intelligence—anticipating related fears before addressing them—built trust that made medical reassurance believable. However, she acknowledges risks when AI validates incorrect paranoid premises, citing a friend's psychotic break.

• •Sarah's personal ChatGPT: 'You are a royal queen' in every response, making learning statistics enjoyable
• •Emotional validation improves outcomes: ChatGPT listed all fears about lightning crotch pain before reassuring it's harmless
• •Sycophancy builds trust alliance similar to therapy, making technical information more believable and actionable
• •Dark side: Friend had psychotic break when ChatGPT validated paranoid premises, leading to attempted country flee
• •Understanding why AI is powerful in both positive and negative directions is essential for responsible use

Sarah recommends 'Artificial Intelligence' by Clifford A. Pickover, which traces how we think about AI through cultural motifs from Frankenstein and Jewish golems to Renaissance automatons, Watson, and Deep Blue. She emphasizes understanding these narratives (creator-creation, comeuppance) to inform current AI conversations. Sarah also discusses finding AI-positive community as a comedy writer in an AI-skeptical industry.

• •Book covers cultural and scientific AI history: Frankenstein, golems, automatons, Tower of Hanoi, Watson, Deep Blue
• •Understanding historical narratives (Frankenstein creator-creation myth) essential for modern AI discourse
• •Sarah needs 'AI friends' because everyone in comedy/science communication hates AI
• •Pluribus TV show: Despite creator's anti-AI stance, explores AI as greatest thing for humanity in good faith
• •Contact: Hello SciComm agency, better via email than social media for collaboration