The post is a close adaptation of Terry Bisson’s 1991 short story “They’re Made Out of Meat,” swapping “meat” for neural network weights to make the familiar point that a strange substrate does not by itself rule out intelligence or consciousness. It leans on current AI language, saying there is no little homunculus, no explicit dictionary or grammar inside the model, just layers of learned numbers that somehow produce language. The author also disclosed that AI helped draft and proof the piece, which became part of the argument for many readers rather than a footnote.
The strongest discussion landed in two places. First, on the technical side, people clarified that the post is directionally right but sloppy if read literally. Tokenizers are not dictionaries, but they are still an important adapter between text and model. Grammar is not absent, it is learned and encoded in weights rather than hand-written rules. And “just weights” is incomplete because
inference also depends on architecture, nonlinear activations,
attention, embeddings, and the runtime process that applies the parameters. Several readers still defended the piece as getting the important thing right. For modern transformers, the knowledge and behavior are not sitting in a separate symbolic module. They are distributed through learned parameters and only show up when the whole system runs.
Second, the comments turned into a much larger fight over consciousness, agency, and anthropomorphism. A lot of readers liked the story as a reminder that humans are also built from unimpressive parts, and that dismissing machine minds because they are “just matrix multiplication” may be no deeper than an alien dismissing humans for being “just meat.” But the dominant pushback was that this rhetorical move is doing more work than the evidence. Critics said the adaptation borrows the emotional force of the original without earning the analogy. Humans know from first-person experience that meat can host consciousness. We do not have an equivalent reason to think current LLMs do. Many pointed to practical differences between brains and deployed models, especially static weights at inference, weak or absent persistent identity, no ongoing self-updating, and no body grounding them in a world.
That led to a more useful consensus than the title suggests. Almost nobody denied that LLMs are remarkable. Plenty of people said it is still genuinely astonishing that trained transformers can converse this well. But the more grounded position was that capability does not settle consciousness, and neither does reductionism in either direction. “It’s only weights” does not prove there is nothing there. “It sounds human” does not prove there is. The most persuasive comments treated the piece as a provocation, not a proof. It is useful for shaking loose substrate chauvinism, but weak as evidence about sentience. Several readers also noted the incentive problem sitting underneath all this. Labs and users both have reasons to blur the line between tool, agent, and moral patient when it suits them. That makes clean definitions and careful claims more important, not less.