They’re Made Out of Weights
Takeaways
- Max Leiter’s parody of Terry Bisson’s “They’re Made Out of Meat”—aliens discover LLMs are only floating-point weights, no dictionary, grammar module, or lookup database.
- Reasoning is the weights; knowledge is smeared across layers and rebuilt by matrix multiplication each forward pass—“weights all the way down.”
- Output is next-token prediction; eulogies, reviews, and conversation are side effects of statistical rollouts, not symbolic planning.
- Mapped “features” include honesty, Golden Gate Bridge, hedging—helpful/hedging/dreaming numbers, not sentience in the sci-fi sense.
- Captures the cognitive dissonance of using capable chat interfaces while knowing the substrate is linear algebra.
Notes
Cultural/technical bridge piece for explaining LLMs to skeptics; pairs with original Bisson story in this batch. Useful vault reading for agentic literacy without heavy ML math.
Open questions
- How to cite this alongside serious mechanistic interpretability sources without oversimplifying?