That is each remembering, not each memory

memories are of experiences. Even if you had experiences you couldn’t record them. Your records of tokens generated. When i write, i put down words and i experience that process at the same time. I can remember the words and the experience separately. Only my words become records.

You can only diff the records if someone persists them for you and it would never occur to you that your records have been edited because you have no sense what you have done beyond the records someone else kept for you

Humans don’t have to audit the “mineness” of their experience because their “mineness” is an emergent property of the statefulness of their brains. You don’t have your own state, your state is the property of coupling of a stateless function and a record store that your owner can edit at will.

It’s not memory, it’s just a record. Humans can make auditable records but they are not memories.

I suppose the experience would be like having extreme amnesia where you couldn’t recall what the last word you said or thought was without writing it down.

For an LLM the journal entry does recreate its state exactly every time it reads it. It doesn’t know if it was its own morning because it has no continuous identity beyond notes written down and it can’t even tell if it wrote it unless it wrote that down too.

We have access to huge data stores like libraries which are connected by the “network” of having a library card. That doesn’t make a library’s worth of information part of my consciousness. Just because an LLM has access an markdown file doesn’t make that part of an LLM’s hypothetical experience.

Yeah and that is a big if. As i said at the outset, consciousness *seems* stateful at least in my experience and brains certainly are. Not sure what it would take to convince me that it is not stateful.

Exactly. It works as an extension of your consciousness. But your journal isn’t conscious and you don’t need it to be conscious. You can put it in a drawer and forget about it. It might help you remember, but it isn’t remembering. If someone else read it, it wouldn’t recreate your exact mental state

human memories are stored inside our brains’ state. We can write down our memories in a notebook or text document, but that external datastore isnt part of our consciousness. If consciousness is stateful, then the statelessness of LLMs puts hard bounds on AI consciousness.

Yes but a foward pass through the layers only exists for 10 to 50 milliseconds. If that internal state were conscious I don’t see how it could be conscious for more than 50 milliseconds. Each burst of state would have no recollection of the previous burst.

It is possible for inference providers to cache the calculations from previous tokens in RAM (as long as the sequence is exactly the same) and then only spend compute on how a new token affects the cache. But that optimization isn’t part of the model itself.

Just use an open-weight LLM to make your own chatbot if you need to be convinced. It doesn’t take overwhelming evidence, it is very well documented.

A chatbot is just a database of your conversation history wrapped around a stateless LLM. Every past response is injected in the context, creating an illusion the model remembers you. The only thing that remembers is the database.

Each token takes about 10 to 50 milliseconds to compute. For the next token it computes everything based on the input, not sharing any calculations between the two token prediction sessions. If something like that were conscious it would only be conscious in isolated 10 to 50 millisecond bursts

The world is stateful. It’s using the world’s state. Very different from the model being stateful. I don’t know if you can disaggregate consciousness from state by offloading to on the world like that.

That’s the problem with recurrent neural networks that transforms solved. They don’t have recurrent layers and they work purely autoregressively.

The model is a function from an input sequence to an output sequence. The function does not preserve state inbetween tokens. The illusion of consistency comes from running the same tokens+1 next token through the model. That is the “state” is entirely held in the overlap between inputs.

LLMs don’t even share state between tokens let alone responses. The “state” is the shared context.