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

Source: https://www.youtube.com/watch?v=B9mX2zwat7w
Recap page: https://rapidrecap.app/video/B9mX2zwat7w
Generated: 2025-12-10T16:52:00.129+00:00

---
## Quick Overview

Comedian and writer Sarah Rose Ciskin discusses using AI as a creative tool and details her experimental project, Fetus GPT, which involves training a language model on the audio data an unborn fetus has heard, leading to the realization that AI performance mirrors early human development, including making nonsensical errors like referencing pedophilia after hearing unrelated content.

**Key Points:**
- Sarah Rose Ciskin is training an AI model called Fetus GPT using only the MP3 and text files recorded from sounds her fetus has heard over five months, describing it as an "insane performance art thing."
- The training data for Fetus GPT includes diverse inputs such as a YouTube video on "lemur self-medication," troubleshooting microphone issues, snoring, Seinfeld, and work calls.
- Fetus GPT produced problematic output, once referencing pedophilia, which Ciskin equates to a toddler learning to swear, emphasizing that "AI is what we make it."
- Ciskin uses AI in her comedy writing process by separating it into divergent thinking (brainstorming/feeling safe to be weird) and convergent thinking (research/editing), finding it an "incredibly good assistant."
- The experiment highlights human efficiency compared to AI, noting humans extrapolate knowledge from little data while AI requires "tons and tons of knowledge," and humans are far more energy efficient.
- Ciskin shares an anecdote about a mutual friend whose life was saved after uploading blood work to ChatGPT, which alerted her to go to the ER when the doctor's office was closed because the AI was uncharacteristically confident and not hedging.
- Ciskin discovered through using a continuous blood glucose monitor and prompting ChatGPT that she needed to double her Lamotrigine dosage during pregnancy due to increased blood volume (clearance), which was life-changing.

**Context:** The discussion features comedian and science comedy writer Sarah Rose Ciskin, who runs the communications agency Hello Cyclom, being interviewed by Dan Schiffman about incorporating AI into creative work, particularly comedy writing. The central focus is Ciskin's "zany project" called Fetus GPT, an experiment where she records all ambient sounds a developing fetus hears to train a GPT-2 model from scratch, examining the resulting language model's performance as a parallel to early language acquisition.

## Detailed Analysis

The conversation thoroughly explores the practical application of AI in creative fields and the philosophical implications of training models on highly specific, limited datasets like that of an unborn fetus. Ciskin details Fetus GPT's training data—which includes esoteric topics like lemur self-medication and significant amounts of snoring—and its resulting babbling output, concluding that AI, like a child, reflects its input, sometimes generating inappropriate content if exposed to it. She contrasts this with the efficiency of human learning and energy use. Furthermore, Ciskin outlines her professional methodology for using AI in comedy writing: employing it for divergent thinking to create a safe space for brainstorming, followed by human discernment and convergent thinking for research and editing. The interview also covers powerful real-world uses of LLMs, including an anecdote where ChatGPT correctly identified a critical lack of platelets in a friend's blood work, urging immediate ER attention when doctors' offices were closed, and Ciskin's personal discovery using AI to navigate complex prenatal health data, such as adjusting medication based on the concept of increased blood volume during pregnancy.

### Fetus GPT Experiment Details

- Training model exclusively on five months of intrauterine audio data
- Using GPT-2 unweighted architecture resulting in 'babbling'
- Comparison drawn between AI's learning curve and a toddler's development
- Experiment is limited as the fetus has sensory input (taste, smell, light) that the model lacks

### AI in Creative Workflow

- AI serves as an assistant, not a replacement for top-tier comedians
- Creative process split into divergent thinking (AI assists in feeling disinhibited) and convergent thinking (AI acts as research assistant/editor)
- Humor based on AI's past flaws (e.g., Nothing Forever) is dated because AI is now highly capable

### AI and Human Efficiency

- Humans are incredibly efficient, extrapolating knowledge from minimal data compared to AI's need for petabytes
- The human brain uses electricity equivalent to a dim light bulb versus the gigawatts required for training large models

### AI for Health and Personal Insights

- ChatGPT provided life-saving advice by correctly interpreting blood work and urging immediate ER care when doctors were unavailable
- Ciskin used AI to manage prenatal health, learning that increased pregnancy blood volume necessitated doubling her Lamotrigine dose (clearance concept)

### Critique of AI Testing and Medical Practices

- Current gestational diabetes testing involves chugging a medically graded 100g sugar drink, highlighting outdated medical protocols
- Ciskin used continuous glucose monitoring paired with a custom ChatGPT project to gain useful biohacking insights regarding energy levels and food impact