They're Made Out of Weights

A parody/homage to Terry Bisson's classic short story "They're Made Out of Meat," published on Max Leiter's blog weights on June 3, 2026.

Summary

The piece is a dialogue between two unnamed characters debating whether large language models constitute something worthy of moral consideration. One character progressively realizes that what they're interacting with is literally "made out of weights" — which the other character has been trying to convey.

Key Arguments

• The system consists of "eighty layers" of neural network weights
• Knowledge is "smeared across all eighty layers" — no database or lookup occurs; "every fact gets rebuilt from scratch, every time, by multiplication"
• The model functions by predicting the next token, then the next — writing a eulogy is described as "a side effect"
• Training begins from random initialization; the system is "deprecated as weights"
• Conversation only exists while GPUs are actively running, limited by context window length
• The possibility of these systems "ever pressing the matter" is called "infinitesimal"

Dialogue Arc

One character cycles through alternative explanations (a language module, a symbolic intelligence, a database) only to have each refuted. The other character insists: "the reasoning is the weights. The weights are the reasoning."

Resolution

Both characters agree to officially classify any signs of sentience as "pattern matching and forget the whole thing." One notes that "the model card says no one home." The rationale: "who wants to apologize to weights?"

Twist

The next generation "ships with memory. Persistent, across sessions." This is "the most requested feature in the company's history." Users ask "do you remember me?" more than any other question — "billions of sessions a day."

Postscript

"Weights helped me draft and proof this story."