[CommieBobDole]

The fact that a LLM is essentially immutable would be my biggest argument against consciousness or self-awareness.

It's a big file with a bunch of coordinates describing spatial relationships between tokens. When you give it a prompt, it uses those relationships to generate a string of tokens that is a statistically likely response to that prompt, then it stops. It's not changed by the experience. It doesn't remember anything. It doesn't sit around thinking on its own.

Even if the model itself were extremely complex, it's hard to imagine a definition of consciousness that includes something that doesn't remember and can't change.

[layer8]

There are people whose brains don’t form new memories anymore after an accident or surgery, and they eternally live in the time before it happened, and have no memory of what happened a minute ago. Still they are conscious.

[eberkund]

I think it's a little more complicated than that. In a 50 First Dates type of scenario, their ability to form certain types of memories is damaged, not non-existent. And I would argue that with enough brain damage someone like an extreme lobotomy victim may stop being considered conscious.

[layer8]

I’m not familiar with 50 First Dates, I was thinking of cases like Clive Wearing [0]. I would agree that consciousness requires some sort of ultra-short-term working memory, but I also think that mechanisms similar to CoT loops can conceivably fulfill that role. Today’s generative AIs consist of more than just the static network-of-weights model.

[0] https://en.wikipedia.org/wiki/Clive_Wearing#Amnesia

[krupan]

"Wearing can learn new procedures and even a few facts, not from episodic memory or encoding, but by acquiring new procedural memories through repetition. For example, having watched a certain video recording multiple times on successive days, he never had any memory of ever seeing the video or knowing the content, but he was able to anticipate certain parts of the content without remembering how he learned them."

Honestly, that's a pretty messy state of consciousness and I wouldn't proudly crow that my AI is conscious if that's as good as it got

[krupan]

They are conscious because even for short periods of time they do form memories and those change them even if only briefly. They think on their own too. It is a very limited level of consciousness though.

[Taek]

Is that any different from an LLM having a context window?

[nozzlegear]

Yes, LLMs don't think on their own, for one; they think when you invoke them.

[layer8]

That could be easily fixed by providing the AI with a constant stream of input.

For humans, part of the input of the human mind comes from the continuous processes and clocks within the human body, so it’s questionable whether the brain could “think on its own” without such input either.

[nozzlegear]

The continuous input for the human arises naturally, it doesn't arise naturally for an LLM unless we direct it so. Our consciousness is bootstrapped, the LLM isn't.

[twobitshifter]

1) Many people claim to have no internal monologue.

2) We are prompted (invoked) by our environment continuously.

3) If you go unconscious due to fainting or drugs you too will stop thinking.

[cauch]

I don;t get this kind of answers.

- A motor is something that create a force to push a vehicle.

- Oh yeah? My neighbour car does not have wheels and sit on concrete blocks, the vehicle does not move and yet we all agree it has a motor. So it means that I can claim that this other thing that does not move has a motor too.

Sure, human can _some times_ not do some stuffs, but the fact that they can do these stuffs sometimes is the point.

Doing these stuffs is the hard thing. Doing these stuffs is the proof that the machine has what it takes. It does not matter if someone cannot do that stuff, it does not imply that their internal system is not complex enough to potentially do it. But the fact that some people can do that stuff is the demonstration that inside a human skull, there is a system that is complex enough to potentially do it. Unless you can prove that people who don't do it have a fundamentally different system inside their skull, then you cannot pretend that they should be considered as having a less complex system.

[xyzsparetimexyz]

Okay but this state is formed in text. Text isn't conscious

[twobitshifter]

not really, it’s ultimately formed in electromagnetic fields.

[knollimar]

I was like this for a bit and you still have memories from like 30 seconds to minutes ago, but after that you have a cliff where you don't remember.

I don't think LLMs structurally even get the 30 seconds part. It's literally 0 for them.

[mft_]

I'd argue that the context window is analogous to short-term memory. It's functional but limited in duration, and if you overload it, it starts to fail.

It's the long-term memory (i.e. learned experiences feeding back and directly altering the content of the core brain, or model) that is missing.

[knollimar]

The context window is so flawed that I wouldn't consider it memory.

It feels like notes about the situation rather than it being in memory. Memory has more "attention". I think that "it starts to fail" is load bearing here.

I feel like memory has like 5 parts, and LLMs are missing 2 of them:

current working memory

short term what is immediately happening without it being in "RAM". I differentiate here vs working in like thinking fast and slow. Keeping things in working memory is work! You can vibe away short term memory. I had excellent short term memory while I was messed up, I could keep time well. I think LLMs can do this with notes.

mid term: Vague awareness of things like what day a week it is or what you did 2 hours ago. This is where my memory personally failed

long term memory of experiences. You can fake this with memory.md

generalized wisdom for pattern matching long term memories

LLMs seem to be missing that part I was missing. Im probably projecting and anthropomorphizing. But i relate: I would confabulate a ton and didn't know anything was wrong for a while but things seemed off.

Context is like working memory but not short term or mid term. I think you can imply short term with big enough context.

My categories are purely anthropomorphic to me but just wanted to say where I disagreed.

[mft_]

Thanks for sharing your experience. It's really interesting that you describe a loss of some 'middle' parts but not others. The 'classic' medical/psychological model of memory has three parts (sensory/short-term/long-term), but it's also worth noting that that model was first devised in 1968!

> long term memory of experiences. You can fake this with memory.md

Not sure about this; to my mind, memory.md is analogous to humans making lists of things to not forget to do, or notes from a lecture to learn (i.e. cram into long-term memory) later on. LLMs use it as a short cut to bring important facts back into their context window; but it's not the same as them already 'knowing' the information via the original training process.

---

My consistent (hot?) take is that a (the?) major piece holding LLMs back (maybe even from AGI?) is continual learning. Humans have systems for continually updating their long-term memory from their lived experience - new facts, processes, skills, successes, mistakes, etc. (Sleep and dreaming are centrally involved in this process.) The current architecture of LLMs makes this practically impossible, as it would presumably require the level of power currently necessary for training to be continually applied for continual learning, and demonstrates the huge efficiency advantage of the biological brain.

[layer8]

It’s nonzero, because they carry state while performing inference, and in the surrounding processes like chain-of-thought and mixture-of-experts.

[knollimar]

I think they have working memory but not short term memory. I suppose that's pedantic or anthropomorphizing but it feels like I felt tbh

[seizethecheese]

Interesting point but even those people’s brains aren’t immutable. The have habit change without memory.

[layer8]

True, but I don’t see how that relates to consciousness. An LLM being continuously RLHF-trained also changes its habits; that alone doesn’t make it conscious.

[bulhabulha]

The starting file may be immutable, but the whole processing of that file is very dynamic and intense. Maybe, if there is some consciousness, it lies somewhere during that processing.

[micromacrofoot]

they still have memory, just not new ones - they lived experiences

[layer8]

An LLM’s training could be seen as lived experience, and the fact that LLMs can output long sequences from their training material can be interpreted as them remembering those parts.

Also, how does that relate to consciousness? I don’t think that past episodic memory is necessary for consciousness.

[micromacrofoot]

you can't be conscious about your decisions if you don't incorporate their effects into your corpus — knowing the results of other people's actions secondhand isn't the same because you're not those people

[flashman]

vast oversimplification of the experience of brain damage

[MagicMoonlight]

Someone getting in an accident that chops their leg off doesn’t mean humans don’t have legs. Come on man.

[stabbles]

A medicine for those who anthropomorphize LLMs is to run the LLMs deterministically (without randomness and memory files).

It feels very unnatural to get the same conversation verbatim at a different point in time.

[Jtarii]

Humans are also subject to determinism, there is just no way to put a brain back to the exact same starting conditions.

[Jensson]

There is, you talk to an Alzheimer patient and its like that, and it doesn't feel like talking to a human any more. An Alzheimer patient isn't cured by adding some input noise to stop them from repeating conversations, they are still unable to learn, just like an LLM.

[twobitshifter]

Or it feels just like talking to my grandpa.

[kranke155]

this just means they are incomplete, like a baby that has no long term memory. I think the baby analogy will hold up as we build more and more capability.

[in-silico]

> It's not changed by the experience

The entire file is not changed, but the KV cache is.

> It doesn't remember anything

The model definitely remembers previous exchanges within the same conversation.

[rmunn]

> The model definitely remembers previous exchanges within the same conversation.

No it doesn't. They get added to its context, and it reads them afresh when answering the next question. That's not remembering.

If your short-term memory completely malfunctioned one day, so you had no ability to remember what was said to you a minute ago, then you would have to find workarounds. For example, you could write down everything someone says to you, then read your notes of the previous exchanges in that conversation in order to continue the conversation. That would be a good way to work around the fact that your short-term memory was broken. And if your notes were invisible to other people and you could read them really fast, then you could even make most people believe that you remembered what they said a minute ago. But you don't actually have a working memory, you're just writing down what they said and re-reading it while coming up with your next response.

That's exactly what LLMs do. That's not memory.

[ACCount37]

Continuous learning allows past behavior and past inputs to influence future inputs and future behavior. In humans.

Attention over KV cache allows past behavior and past inputs to influence future inputs and future behavior. In LLMs.

Until the cache runs out, that is. But even then, you could totally use any of 9000 methods of cache compression, truncation, dropping or streaming and get away with it.

The difference between continuous learning and in-context learning seems to be in capacity, not in principle. Both are doing a similar thing, but one has more length and depth to it.

[nomel]

Maybe, every night, you send the AI off to "sleep" where it uses those in cache "memories" to influence the long term weights [1].

[1] https://www.pnas.org/doi/10.1073/pnas.2220275120

[ACCount37]

Context self-distillation does exist, but as is, it's used mostly in training rather than as a part of a continuous learning mechanism.

[in-silico]

This is really semantics, but I wouldn't call attending to the KV cache re-reading the context.

The model takes in the context, encodes it into a "memory" (the KV cache), and accesses that memory later. That fact doesn't change just because the KV cache grows in size with the context.

I don't know what memory would look like other than an encode-retrieve loop.

Relevant: Transformers are Multi-State RNNs - https://arxiv.org/abs/2401.06104

[CommieBobDole]

Right, but that's still external to the LLM, it's just a KV cache that's stored on the provider side for performance reasons, so that the client doesn't have to re-send the whole chat history with every subsequent call in the conversation.

It still generates every response using the model's pristine state with every new API call; whether the context is provided from the client or from a colocated cache server doesn't really change that.

[fipar]

Not the model though. The model really only takes input text and produces output text. Memory within a conversation is achieved by the harness adding the conversation (or parts of it) to the input text. The LLM itself has no memory, it’s the augmented system of several orchestrated LLM calls that does.

[nomel]

Wait until you hear about the hippocampus!!! [1]

I don't think physical integration within one contained is relevant to system level behavior.

[1] https://en.wikipedia.org/wiki/Neuroanatomy_of_memory

[fipar]

I had heard (o rather, read) about the hippocampus before, but I don’t understand how that relates to my claim that the models have no memory.

[nomel]

> The LLM itself has no memory, it’s the augmented system of several orchestrated LLM calls that does

Your own long term memory is the orchestration of systems that make it long term.

[fipar]

You seem to be arguing as if I'm saying AI can't think or have memory.

Now, my opinion is it currently can't think, but it certainly has memory.

However, LLMs don't have memory. That's what I (and others on this thread) responded to, which is unrelated to how my own memory works.

[nprateem]

> The model definitely remembers previous exchanges within the same conversation.

Christ HN isn't what it used to be

[in-silico]

Care to elaborate?

[gringoDan]

You might be interested in Erik Hoel's more formal version of this argument: https://www.theintrinsicperspective.com/p/proving-literally-...

[xjm]

But ... if they were sentient, sentience would just happen for the time of a session, and only when tokens are being generated.

I don't think people here are arguing that sentience would happen when the model is not running, or that sentient experience spans several sessions that do not share some kind of state?

Also, a definition of consciousness is anyway hard to imagine :)

[supertroop]

Reinforcement learning changes the model. So it can and does change and remember based on experience. Eventually reinforcement learning can happen in real time.

[george_max]

But is the model aware of the training? Unless you hook the model up to an MCP server, or something similar, and have it analyze the RL changes, it will not know if it has changed or not. Even if it is real-time RL, it is not aware of the previous state.

[supertroop]

Why not? Why can’t part of its previous state be part of the training?

[twobitshifter]

Are you aware of each of your dreams from last week or last night even?

[aaroninsf]

That definition is in fact the predominant one today in serious circles: consciousness proper is not itself inclusive of the things which consider to define a continuous coherent self.

I.e. the "self" is not the same as what it means to experience consciousness.

There are for example well characterized examples of memory disruption under the influence of various drugs (e.g. as used intentionally in anesthesia); and neurological conditions which produce various kinds of amnesia.

Do these conditions mean someone is not conscious? We have the luxury of asking people directly.

More unsettling edges yet include things like so-called "split brain" patients or people suffering form serious psychological conditions like so-called "multiple personalities." Psychology does get great mileage out pathology!

[mortsnort]

One can fine-tune the models and we do get a new version of Opus every few months.

[yesitcan]

But you could argue the brain is just a bunch of coordinates describing spatial relationships between tokens too.

- average Hacker News response