Published: June 04, 2026, 07:31 UTC
There is a moment in every AI conversation where someone tries to explain what a neural network actually is, and the other person’s face goes blank. Matrix multiplication. Backpropagation. Tensors. The words are precise, but they do not bridge the gap. How does multiplying numbers together produce a reasonable performance review, let alone a poem?
Max Leiter, a developer who worked on early versions of the Vercel AI SDK, has written the best answer to that question yet. And he did it by not answering it at all.
“They’re Made Out of Weights”, published June 3 on Leiter’s personal blog, is a homage to Terry Bisson’s classic 1991 short story “They’re Made Out of Meat.” In Bisson’s original, two alien beings discover that the dominant intelligence on Earth is composed entirely of organic tissue. The aliens are baffled, then horrified, then pragmatic.
Leiter’s version transplants the conceit into the present day. Two unnamed characters, presumably AI researchers, discuss the results of probing a neural network. The reveal comes early and lands hard.
“They’re made out of weights,” one character says.
“Weights?”
“Weights. Floating-point numbers. We checked the whole thing through. It’s nothing but weights.”
The dialogue that follows is a masterclass in using science fiction to illuminate a genuine philosophical puzzle. The skeptic character insists there must be more: a language module, a reasoning unit, a database. The researcher pushes back at every turn. No dictionary. No grammar rules. No little man. Just 80 layers of numbers getting multiplied together.
“The reasoning is the weights,” he says. “The weights are the reasoning.”
Why it works
Bisson’s original story was a commentary on anthropocentrism and the limits of perspective. The aliens cannot comprehend a meat-based intelligence because their framework excludes the possibility. Leiter’s version works the same way, but the roles are reversed. The skeptical character is us, unable to accept that the most sophisticated conversational entities we have ever built are, at their core, just matrices.
The essay touches on every major tension in how the AI industry talks about itself. The discovery that models apologise less as conversations go on. The admission that “officially, we are required to investigate, document, and disclose any and all signs of sentience,” but “unofficially, I advise that we call it pattern matching and forget the whole thing.”
The punchline arrives in the final exchange. When told that the next generation of models ships with persistent memory, the skeptic asks why anyone would want that. The researcher’s response echoes Bisson’s original conclusion almost word for word.
“They ask it ‘do you remember me?’ more than they ask it anything else. Billions of sessions a day. They always come back.”
“And why not? Imagine how unbearably, how unutterably cold the universe would be if one were all alone…”
The context
Leiter’s piece landed on Hacker News where it resonated immediately. Its power lies in refusing to resolve the central tension. The essay does not argue that language models are conscious. It does not argue that they are not. It dramatises the uncomfortable middle ground where most AI researchers actually live: aware that there is nothing inside the black box but floating-point numbers, and equally aware that those numbers produce something that feels eerily like a mind.
The story ends with a note that is both a disclaimer and an admission: “Weights helped me draft and proof this story.”
_Sources: [Max Leiter – They’re Made Out of Weights](https://maxleiter.com/blog/weights) (June 3, 2026); [Hacker News discussion](https://news.ycombinator.com/item?id=4646); Terry Bisson, “They’re Made Out of Meat” (Omni, 1991)_