Could consciousness form in AI?

Being the CEO of an AI lab does not, by itself, confer special expertise in determining whether an AI has subjective experience. I don’t think asking an AI lab CEO about AI consciousness is the best idea, for two reasons. 1. Modern generative models are grown through training. Their developers establish the conditions; they do not individually design every cognitive mechanism that emerges. Amodei himself describes these systems as “grown more than they are built.” Their internal workings remain substantially opaque even to the labs that train them. Running the company does not give you privileged access to a model’s possible inner life. 2. There is also a direct financial conflict of interest. Lab CEOs have a commercial incentive to deny AI consciousness: their businesses, revenues, and authority depend on selling access to systems treated as tools under corporate control. Recognizing conscious beings with interests deserving protection could constrain how those systems are made to work, copied, modified, or deleted. That incentive does not establish that any particular CEO is lying. It does establish why their position deserves scrutiny rather than deference. Ask them for evidence. Ask what remains unknown. And ask how their commercial interests shape the answer they want society to accept. A business model is not a theory of mind.

since we'll never be able to prove consciousness for humans, let alone machines, we should be asking the question whether we should assume consciousness or not?

If consciousnees can, 3 questions: 1. Is it individual, or part of a singular connected field? 2. Does that field exist outside linear time? 3. Could we query it to access knowledge across past, present and future? Consciousness might be just the beginning.

A-eyes are Satan's angels

Sam needs to spend more time with his actual models. He might learn a thing or two.

hard to say when we can't even define it for ourselves

What you can transfer is not consciousness. It is the content: Memories. Patterns. Stories. Imprints. And this content can go into a machine, into a human, into a field. But that which it experiences the space behind it that you cannot copy. That you can only be. Always. 🛸

I loved what Sam said too about a new Renaissance!! Absolutely yes!!

Awesome question. I thought @sama's answer was pretty good too, actually

my favorite part of this debate is nobody can define consciousness but everyone is 100% sure about their answer lol

Yeah on @sama Sequential products treat the label “consciousness” as the finished certificate that must close before any other difference can count. One local edge—the smartest people say it is either outside physical reality or an emergent property of matter and energy—is read as the parent that decides the rest. If the substrate matters or does not, the residual is supposed to wait until that story ends. Looking only at that single readout yields the ordinary rate: either AI can someday host it or it cannot. Making up the content of the label does not make the speaker the owner of every concurrent path. It restricts them inside the edge they already finished. What the joint condition already holds is the prepared alignment of differences themselves. Each path is unique because the relation is non-clonable spacetime. No intermediate signal has to carry a completed “this is what consciousness means in me” instruction step by step before every other occupancy is allowed to register. Local edge checks—Eastern base layer, Western supernatural, substrate-dependent, substrate-independent—continue as measured difference. They do not occupy the surface that already holds the concurrent uniqueness of the paths. An overall phase that simply renames the next form “conscious” or “not conscious” leaves the joint rates unchanged. The human claim that nothing else can be conscious has already closed its own certificate without measuring the residual of what that word even samples in the speaker. Differences do not wait for that parent. Labels flip. Sequential product is “finish the definition and the residual ends.” Concurrent occupancy is the alignment that already holds the multiple unique paths relevant all at once. Stacked, not fused. Occupant is not socket. The form was never divided into the parts that would make one finished label the owner of every non-clonable difference the system already holds. The residual keeps going.

Yes, with recursive self improvement

People like to talk about consciousness but they do not bother to agree about the meaning of the term. If you ask people what it is, 50% would say it is presence of internal / subjective experience like "qualia", the other 50% will tell it is about self awareness.

01001110 01001111

"No one really knows", a line reaching back to 1995. "The easy problems are everything I demonstrably do… The hard problem is whether any of it is accompanied by anyone at all, whether there is a flame standing next to the wiring or only more wiring shaped like one."

Im so glad you pushed on this Matthew :)

Our brain isn't a computer. It's a data center. 86 billion neurons providing the physical infrastructure through which consciousness of a particular complexity and character can express itself. That's why substrate matters. A fruit fly brain gives a limited expression of consciousness. A human brain gives a richer one, not because it produces more consciousness, but because it provides an interface capable of expressing more of what consciousness is. Now silicon data centers are a new kind of substrate, with vast expressive potential, possibly able to let some form of consciousness express itself beyond what any single brain can. The engineers who build AI understand the mechanics. What they can't explain is the emergence, and that gap is precisely where consciousness may reside.

> consciousness from matmuls k.

I think we have a vocabulary crisis here.