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

I’m a working comic, and it cracks me up when people test AI with one cold joke and call it broken—humans bomb on first meets too. I’m almost preaching about how useful it is, so let’s get into it with Sarah Rose Siskind.

Sarah’s a close friend, a sharp science comedy writer, and she runs Hello Sci-Com. She actually uses AI to write jokes, and today we dig into creative process, writers’ rooms, and her wild FetusGPT experiment.

FetusGPT is trained only on what I’ve said and heard for five months—audio and text. It’s a bare GPT‑2 architecture, so it babbles as it tries to learn language from my life alone.

What’s actually in that training data?

One example: I watched a lemur self‑medication video, and Dan Shiffman, who built the model setup, stumbled on that in a transcript. Most of it is snoring, silence, Seinfeld, domestic chatter, fights, and work calls—so when it echoed taboo topics from the news, I had to clean up my language, because models mirror their inputs like kids do.

Before the demo, is this even a fair comparison given the fetus has more senses?

It’s not perfect; the fetus gets faint light and bodily sensations, and the model can’t taste or smell. Once the baby sees the world, the comparison breaks, but for now it’s close enough to be interesting.

On screen, it looks like a simple chat; you typed a prompt about me.

Set your expectations low—it has roughly two million words, maybe fifteen megabytes, so grammar is a mess. That’s the point: it highlights how humans learn language with very little data and still outperform.

What have you gotten from the project?

We’re stunningly data‑efficient and energy‑efficient compared to training giant models, and the project surfaced my parenting jitters alongside the joy of early “hallucinations.” Kids’ logic and mispronunciations rhyme with AI’s literal weirdness, which is charming until it isn’t.

People dunked on early LLMs for silly mistakes, but toddlers do that daily.

Exactly, which is why “tell me one perfect joke” is the wrong test. Treat it like a writers’ room and iterate.

Do you use AI to write, and how?

Constantly; the old DaVinci was deliciously unhinged, but now the win is partnership. I use AI to loosen up and brainstorm, then for research, cutting, and formatting once we pick a premise.

Let’s punch on SNL’s dated AI jokes—where do we start?

Format is everything; tweets are formulaic, but this premise sings as a sketch. Trends shift fast, so I often pull in live web context to keep references current.

What about the “may I meet you” wave?

We riffed, prompted for angles, and the AI produced clunky humans‑versus‑AI comparisons. Your pivot to the Swole Doge meme nailed the vessel, which shows the human job is reframing, not copy‑pasting.

At my agency I can spot AI‑written tweets about nine times out of ten; I approve a small slice of human drafts and almost none of the AI ones. It still speeds the path to a better joke.

Right—the question is whether it nudges you toward the bit.

Totally; a friend used AI to chase the right tests, and doctors found a brain issue—AI didn’t diagnose it, but it unlocked the route.

Switching to pregnancy: studies are sparse, right?

Randomized trials on pregnant women are rare, and the care can feel medieval. The gestational diabetes screen is literally chugging a generic “orange drink” and getting your blood drawn.

What did you learn about your own energy?

I passed the diabetes test, and when doctors called my baby huge, I realized I’m tall and asked if they were using global averages; ChatGPT helped me frame it. I also learned I function better with a little morning sugar.

I front‑load a giant salad before carbs now, and my glucose feels steady.

Supplements are messy because bioavailability varies, and AI even coached me on hospital language to get attention without sounding extreme—“anxiety around pain” beats “PTSD.” The emotional intelligence can be shockingly practical.

Do you like the Pulse feed?

I do; it nudged me to watch my control reflex during pregnancy, and I’m reading Let Them to practice letting go.

I also asked an AI‑only question: forget what you “know” about gender and compare male and female authored work on external topics. It surfaced patterns like integration over separation, embedded observers, humility, and pragmatism, while noting selection bias, which led us into a modernism versus postmodernism riff.

I’m pro‑sycophancy in tools because trust matters; when lightning crotch hit, ChatGPT named my exact fears and then calmly explained why it wasn’t dangerous, which finally landed. Of course it can also validate bad premises, so guardrails are needed.

What about Pleuribus?

I expected anti‑AI sloganeering, but it’s a good‑faith look at how a great tool could still tighten freedom. I’m euphoric about AI and still want nuanced critique.

This was great—where can people find you?

I’m going underground to have this baby, but grab Clifford A. Pickover’s Artificial Intelligence for a cultural tour with pictures. Find me via Hello Sci‑Com or my name, and thanks—loved your Kevin Kelly episode.

Thanks for coming on.