The Engineering Roadmap
Put the systems on a line.
A stateless text model. No memory, no persistence, no perception, no action, no body, no stake. This is where almost all the confident public discussion is still aimed, and it is already historical.
A multimodal assistant with a long context. Perception of a sort — images, audio, video — integrated into a common representation. Still no persistence across sessions, still no action.
A persistent agent. Memory across months. Goals pursued over hours. Tools that change the world. A record of what it has done and a rudimentary interest in doing it well. This exists now.
An embodied autonomous system. All of the above with sensors and actuators in physical space, a body that can be damaged, a charge level that can run out, a world that pushes back.
An artificial organism. Something that maintains its own integrity against entropy, whose states are organised around its continued existence, which can be harmed in a sense that is not merely functional.
The first has essentially none of the properties any theory of consciousness identifies as relevant. The last has nearly all of them, including the ones the phenomenologists insist on. And the intervening steps are not a philosophical speculation; they are, roughly, an engineering roadmap that is being executed for entirely unrelated commercial reasons.
The question “are LLMs conscious?” assumes the answer is a property of a category. It is not. It is a property of particular systems, and the systems are moving along this line at a pace set by capital expenditure.
What I would ask of anyone who feels certain about this — in either direction — is to say where on the line their certainty stops applying. The sceptic who is confident about a stateless text model should be asked whether the confidence survives an embodied homeostatic agent. The enthusiast who thinks a chatbot might be conscious should be asked what the chatbot has that a thermostat lacks, and whether the answer is anything more than eloquence.
Most people, I have found, have not located themselves on the line at all. They have an opinion about “AI,” formed in one place, and they apply it everywhere, and the systems keep moving underneath it.
from Are LLMs Conscious?: Language, Experience and the Problem of Artificial Minds (2026)
