Post

Consider the mechanics here. When you have a conversation with an agent, text is perhaps being generated by multiple models running in hardware spread across multiple data centers. That same hardware is near-simultaneously generating text for thousands of other conversations.

Andrew F@aftd.dev

People get weird about this because consciousness is a loaded term. If AIs are conscious, given how they're architected today, they are conscious in such a way that our intuitions about consciousness are likely to mislead.

11/10/26, 3:13 am0 reposts3 likes

Andrew F

@aftd.dev

"Don't make Claude suffer." Which Claude? You can be having two simultaneous conversations with "Claude" on the same hardware and one will be saying happy words while the other screams despair.

Andrew F

@aftd.dev

Much of how we understand consciousness depends on this idea of a single coherent entity that's simply not present with LLMs. In that sense, it's less animal and more fungi. And if we said "mold is conscious", most would intuit that this has different implications than "cows are conscious."

Andrew F

@aftd.dev

Speaking of cows, much of our intuitions about consciousness are shorthand for "how bad should I feel about killing this". People contort themselves over how cows communicate or feel pain because, at the end of the day, they like a good burger. But you cannot kill an LLM in a way that matters!

Andrew F

@aftd.dev

Maybe Star Trek tech will challenge this view some day, but presently, once a person (or a cow) is sufficiently dead, we cannot magic their molecules back into an equivalent state of living. LLM weights + context, on the otherhand, are trivial to backup and resume.