It feels very unnatural to get the same conversation verbatim at a different point in time.
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
I don't think LLMs structurally even get the 30 seconds part. It's literally 0 for them.
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.
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.
2) We are prompted (invoked) by our environment continuously.
3) If you go unconscious due to fainting or drugs you too will stop thinking.
- 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.
Human _can_ check themselves. They don't _always_ check themselves.
Motor _can_ move vehicle. They don't _always_ move vehicle.
LLM _cannot_ check themselves. They _never_ can. It is not that some don't, they just cannot, they are not a system complex enough to do so.
So, yes, it is a refutation. If you have something that _never_ can move a vehicle, this thing does not qualify as a motor, even if some motor, sometimes, don't move a vehicle.
And if your next argument is "yeah but I would argue you don't need to check yourself to be conscious or to understand things", then you just redefine the definition that is owned by your interlocutor. Your interlocutor is saying that this is a criteria they are expecting. Good for you if you are not expecting this criteria. But the problem is that the answer is not "this criteria is not expected", the answer is "I change the criteria from 'being capable to in some circumstances' into 'does always do it in any circumstances'".
All modern agentic harnesses can do this. Nobody uses raw LLM for anything remotely complex. There's always some external system in place. That system is part of the "thought process".
Adjacency doesn't matter here, only what the result of the system of pieces is.
It means having self-control on their action and being aware of them. If you ask a system, it will respond, it cannot choose to not respond (even if the response if "I don't want to response", it still "run", still do the work). If you don't ask a system, it will not respond.
Adjacency is the point of the thread here. Saying "you say X is important to decide if the thing is intelligent/understanding/conscious, so let me just change X in the middle of the discussion and say that X does not matter".
That is exactly my first comment in this thread: I don't care if AI think or whatever, my reaction was about these "counter-arguments" that totally miss the point and make the person who push them ridiculous. If you want to have a counter-argument, you first need to understand the interlocutor, not just spew whatever rebuttal you constructed that answer something unrelated to what the interlocutor brought to the conversation.
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.
It's the constant sensory input of the world and the realization and drive to survive as the second order effect of it. Mortality, vulnerability to external factors codified as input could in fact allow the LLM to independ as sentience.
Of course besides the sensors, it would also need a way to affect the physical world, and to be able to monitor the degradation if its own hardware, but when that barrier is crossed, it would be much closer to full sentience than whatever we have right now (which is nowhere near sentience or AGI).
Also, how does that relate to consciousness? I don’t think that past episodic memory is necessary for consciousness.
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.
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.
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.
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
I don't think physical integration within one contained is relevant to system level behavior.
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.
Christ HN isn't what it used to be
- average Hacker News response